FactsExposureLeaderboardAI StocksCapexMembership

ALLFACTSAI · The Intelligence Club

The AI boom
decoded
for investors

Daily verified AI facts. Investing intelligence. Model leaderboards. And by email, free: The Daily Receipts every weekday, The Weekly Brief every Saturday.

Free forever. Six emails a week — The Daily Receipts on weekdays, The Weekly Brief on Saturday — drop either in one click.

▶ Watch the trailer
While Amazon, Microsoft, Meta, Alphabet and Oracle have all sharply increased AI capital spending, Apple's cash paid for property, plant and equipment in the nine months to 27 June 2026 actually fell, to Apple's nine-month capex of $6.8 billion, down 28% from $9.47 billion a year earlierApple Inc. (SEC 10-Q) · Nine months ended 27 Jun 2026Epoch AI tracks how much AI compute a dollar buys, measured in constant 2025 dollars across the chips companies actually purchased each quarter. Since 2023, AI chip price-performance has grown 49% a year, doubling every 1.7 yearsEpoch AI · August 2026AI data centres worldwide reached a power capacity comparable to New York State at peak demand in 2025, which Stanford's AI Index puts at 29.6 gigawattsStanford HAI, AI Index 2026 · 2025AI hiring intensity doesn't track the size of a country's AI industry. Stanford's 2026 AI Index, drawing on Lightcast's job-postings database across dozens of countries, found the 2025 leader for AI-skill job-posting share by a wide margin was Singapore, at 4.69% of all job postings — nearly double the 2.6% share in the United StatesStanford AI Index · 2025 data (2026 AI Index)AI models can now win gold at the International Mathematical Olympiad, yet Stanford's 2026 AI Index found they still fumble a task most children master early. Across ClockBench's 180 clock designs and 720 questions, the top model, GPT-5.4 High, read analog clocks correctly only 50.6% of the time, versus 90.1% for humansStanford AI Index · 2026 AI Index ReportStanford's 2026 AI Index reports that organizations' use of generative AI specifically, not just AI overall, grew sharply in 2025: the share of organizations that regularly use generative AI in at least one business function reached 79%, up from 71% in 2024Stanford HAI AI Index (McKinsey Survey) · 2025Stanford's 2026 AI Index compared generative AI's adoption curve after its mass-market debut to that of earlier transformative technologies, and found generative AI reached 53% population adoption within three years, faster than the PC or the internetStanford HAI AI Index · 2026OpenAI does not disclose GPT-4o's resource footprint, so Stanford's 2026 AI Index Report turned to independent researchers' estimates of the water needed to cool the servers behind it — as much as 1.58 million kilolitres a year — more than 1.2 million people drink annuallyStanford HAI, AI Index · 2026 AI Index ReportAI training's carbon footprint has scaled by orders of magnitude. Stanford's 2026 AI Index estimates that training AlexNet in 2012 produced about 0.01 tons of CO2 equivalent, while training xAI's Grok 4 in 2025 produced roughly 72,816 tons of CO2 equivalent — over 1,150 times the lifetime emissions of an average carStanford HAI, AI Index · 2026 ReportFrontier AI models are catching up fast to benchmarks meant to resist them. A year ago they scored under 10% on Humanity's Last Exam, a 2,700-question test designed to be hard for AI and favorable to human experts. Now accuracy has reached 38.3% on Humanity's Last Exam, up from under 10% a year earlierStanford AI Index 2026, citing CAIS · 2026 AI Index ReportThe information sector — telecom, media, tech services — led US corporate AI hiring in 2025. Lightcast data cited in Stanford's 2026 AI Index found AI skills required in 13.2% of the sector's job postings, up from 7.8% in 2024Lightcast, via Stanford AI Index · 2025 data, pub. 2026McKinsey's State of AI survey, cited in Stanford's 2026 AI Index, asked organizations already using AI where they had seen the clearest gains. The single most commonly cited outcome in 2025 was improved innovation, named by 64% of respondents, more than any other organizational outcome measuredMcKinsey, via Stanford HAI · 2025AI agents attempting OSWorld's real-world computer tasks across Ubuntu, Windows and macOS — file operations, multi-app workflows — historically topped out at just 1% to 12% success. Stanford's 2026 AI Index reports the best model, Claude Opus 4.5, now reaches 66.3% accuracy, within 6 percentage points of the 72.35% human baseline 66.3% accuracy on OSWorld's real-computer-task benchmarkStanford HAI 2026 AI Index (Epoch AI data) · early 2026Global private investment in AI reached $344.7 billion in 2025, up 127.5% year-over-year, with generative AI companies alone capturing $170.9 billion of that total $344.7 billion in global private AI investment in 2025Stanford HAI AI Index 2026 (citing Quid) · 2025 full yearStanford's 2026 AI Index reports that four out of five US high school and college students now use generative AI for schoolwork, even though only half of middle and high schools have any AI policy and just 6% of teachers describe those policies as clear four out of five US students use AI for schoolworkStanford HAI AI Index 2026, Ch. 7 Education · 2026 reportSWE-bench Verified hands an AI model a real GitHub issue and a codebase and checks whether the patch it writes actually fixes the bug. Stanford's 2026 AI Index reports the leading model, Claude 4.5 Opus in high-reasoning mode, had solved about 76.8% of the benchmark's issues as of February 2026Stanford AI Index · Feb 2026TrendForce says a shortage of critical optoelectronic components — EML and CW lasers — is reshaping the AI optical transceiver supply chain, and projects the global market to grow from $16.5 billion in 2025 to $26 billion in 2026, an AI optical transceiver market up more than 57% year-over-yearTrendForce · April 2026Applied Materials sells the deposition, etch and inspection machines chipmakers must install before a single AI chip exists, so its results run a step ahead of the chip cycle. Its third fiscal quarter, ended 26 July 2026, was the biggest in the company's history: Applied Materials booked a record $9.12bn of revenue, up 25% year over yearApplied Materials 8-K (SEC) · Quarter ended 26 July 2026Applied Materials' Semiconductor Systems segment — the deposition, etch and inspection tools chipmakers install before fabricating a single AI chip — reported fiscal third-quarter 2026 revenue up 26.5% year-over-year to $7.04 billion in Applied Materials' Semiconductor Systems segment revenueApplied Materials, Inc. · Q3 FY2026 (ended 26 Jul 2026)AMD's Data Center segment, the one selling Instinct GPUs and EPYC server chips, took $16.6 billion of the company's record $34.6 billion of revenue in 2025 — 48% of all AMD revenueAMD FY2025 results · year ended 27 Dec 2025Amkor Technology, the outsourced chip packaging and test house scaling up advanced packaging for AI processors, reported record second-quarter 2026 net sales of $1.90 billion, up 26% year-over-year $1.90 billion in record Q2 net sales, +26% YoYAmkor Technology · Q2 2026Amazon's cash capital expenditure across the whole of 2025 came to $131.81 billion in one yearAmazon 10-K, SEC EDGAR · FY2025Amazon's most recently reported quarterly cash capital expenditure reached $54.21 billionAmazon, SEC EDGAR (XBRL filings) · Q2 2026Amazon's cash capital expenditure in the first quarter of 2026 was $44.2 billion in a single quarterAmazon 10-Q, SEC EDGAR · Q1 2026Arista Networks, whose switches carry AI training and inference traffic inside hyperscaler data centres, reported total revenue for the quarter to 30 June 2026 of $3.036 billion, its first quarter above $3 billion and up 37.7% year-over-yearArista Networks · Q2 2026Amphenol's Communications Solutions segment, whose interconnects and cabling carry AI-related IT datacom traffic, reported net sales of $5.38 billion in the quarter ended 30 June 2026, up 85% year-over-year $5.38 billionAmphenol Corporation 10-Q · Q2 2026Applied Digital had signed long-term hyperscaler leases for 1,410 megawatts of AI data-centre capacity across five campuses as of May 31, 2026, worth approximately $36 billion in contracted revenue over the initial 15-year lease terms 1.41 gigawatts of AI data-centre capacity leased for $36 billionApplied Digital Corp. · as of May 31, 2026AppLovin's revenue grew 53% year-over-year to $1.924 billion in the second quarter of 2026, powered by its AI-driven Axon ad-matching engine, with net income up 55% to $1.267 billion AppLovin revenue: $1.92 billion in Q2 2026, up 53% year-over-yearAppLovin Corporation · Q2 2026Arm's total royalty revenue grew 11% year-over-year to $671 million in its fiscal fourth quarter of 2026, and the company said its data centre royalty revenue within that more than doubled year-over-year on rising Arm-based server chip deployments by hyperscalers Arm data-centre royalties more than doubled YoYArm Holdings · Q4 FY2026Arm Holdings' royalty revenue grew 22% year-over-year to $715 million in its fiscal first quarter of 2027, as data-center royalties more than doubled $715 million in Arm royalty revenue, up 22% year-over-yearArm Holdings · Fiscal Q1 2027 (Apr–Jun 2026)Every advanced AI chip needs packaging and testing before it ships, and ASE Technology, the world's largest outsourced semiconductor packaging and test house, is racing to add capacity for it. Equipment spending in its packaging and test segment jumped in the quarter to 30 June 2026, from $942 million a year earlier to $1.646 billion in ASE's advanced-packaging equipment spending, up 74.7% year-over-yearASE Technology Holding · Q2 2026Every advanced AI chip is printed on an ASML machine, and the company says AI demand has its customers accelerating their expansion plans. On 15 July 2026 ASML said it would lift its 2026 low-NA EUV capacity of around 65 systems, for 2027, by 30% more EUV machines a yearASML Q2 2026 results · 15 July 2026ASML's CEO says ongoing AI-related investment is driving demand for advanced logic and memory chips, and it shows up in the numbers -- the lithography maker's second-quarter 2026 total net sales climbed roughly 21% year-over-year, from 7.7 billion euros to 9.3 billion eurosASML Holding · Q2 2026Astera Labs makes the connectivity chips that link AI accelerators inside data-centre racks, and demand pushed its revenue for the quarter ended 30 June 2026 up 104% year-on-year to $392.4 millionAstera Labs 10-Q (SEC) · Q2 2026Astera Labs, whose PCIe and CXL connectivity chips link AI accelerators inside hyperscaler racks, posted record second-quarter 2026 revenue of $392.4 million, up 104% year-over-year, at a GAAP gross margin of 73.3% GAAP gross margin for Astera Labs on record $392.4 million quarterly revenueAstera Labs, Inc. · Q2 2026 (ended 28 Jun 2026)Broadcom reported $20 billion of AI revenue in its 2025 financial year, out of $63.89 billion in total — 31.3% of everything it soldBroadcom FY2025 Form 10-K · FY2025Broadcom's entire 2025 financial year produced $20 billion of AI revenue. In the single quarter to 3 May 2026, its AI semiconductor revenue — custom accelerators and AI networking — reached $10.8 billion in one quarter, up 143% year on yearBroadcom · Q2 FY2026 (to 3 May 2026)Broadcom's custom AI accelerators and networking chips for hyperscale AI clusters just delivered their fastest growth quarter yet — in the fiscal quarter ended 2 August 2026, revenue reached $16.7 billion in AI semiconductor revenue, up 221% year-over-yearBroadcom Inc. · Q3 FY2026Alibaba is racing to keep pace with AI demand on its cloud platform. Capital expenditure in the quarter ended 30 June 2026 reached RMB67.68 billion ($9.98 billion), up 75% year-on-year from RMB38.68 billionAlibaba Group · quarter ended 30 Jun 2026Baidu's Apollo Go robotaxi service has expanded to 28 cities worldwide and its fleet has now driven over 240 million kilometres with nobody behind the wheel, out of more than 350 million autonomous kilometres in total Apollo Go's fully driverless total: 240 million+ kilometres across 28 citiesBaidu Q2 2026 results · Q2 2026Baidu's AI Cloud Infra revenue grew 50% year-over-year in the quarter to 30 June 2026, powered by GPU Cloud demand up 283%, and the company poured capital into keeping pace. Group-wide capital expenditure nearly tripled year-over-year to RMB11.39 billion ($1.68 billion), Baidu's Q2 2026 capital expenditure, up 199.7% year-over-yearBaidu · Q2 2026The Bureau of Labor Statistics built a new four-tier AI exposure classification for every US occupation, combining AI-capability measures with real-world Claude and Copilot usage data mapped onto job tasks. Summing 2025 base-year employment across its 831 rated occupations, BLS's own data shows 53.2 million US workers, 31% of total employment, hold jobs rated 'Very high' AI exposureUS Bureau of Labor Statistics · 2025 employment projectionsBLS's 2024-34 employment projections show AI adoption fueling outsized growth in the occupations that build and secure it: data scientist jobs are projected to grow 33.5%, more than ten times the 3.1% projected growth rate for all US occupations combined 33.5% projected growth in US data scientist jobs through 2034U.S. Bureau of Labor Statistics · 2024-34 projectionsConstellation Energy's newly acquired Calpine subsidiary signed a new agreement in February 2026 to connect and power a CyrusOne data centre next to its Freestone Energy Center in Texas — plus an exclusive option on a second phase of equal size — for an initial 380 megawatts of power committed to a single AI-era data centre in Texas (Calpine-CyrusOne deal)Constellation Energy Corporation · 24 Feb 2026AI benchmarks typically score just one model on a single run, but a new study testing 21 LLMs across 16 benchmarks spanning coding, reasoning, medicine and agentic tasks found that correcting for both single-model and single-run bias reveals real achievable performance is 82% better than standard single-model benchmark scoresarXiv (Fowler et al.) · Jun 2026Cerebras Systems, the wafer-scale AI chip maker whose cloud inference revenue nearly quadrupled year-over-year, disclosed in its second-quarter 2026 earnings release that it now has 600 MW of data center capacity under contractCerebras Systems · Q2 2026Cadence Design Systems, whose software chipmakers use to design AI accelerators, reported $1.584 billion of revenue in its second quarter of 2026, up 24.2% from $1.275 billion a year earlier, with quarter-end backlog reaching a record $8.1 billion as CEO Anirudh Devgan cited accelerating demand for AI-driven design tools Cadence's Q2 2026 revenue: $1.584 billion, up 24.2% year-over-yearCadence Design Systems · Q2 2026Long-dated nuclear supply is being locked up now for demand that only arrives at the end of the decade. In the quarter to 30 June 2026 Constellation Energy signed power purchase agreements running 15 to 20 years with investment-grade customers, beginning between 2029 and 2032: Constellation signed another 920MW of long-term nuclear supply, none of it starting before 2029Constellation Energy 8-K (SEC) · Q2 2026A Census Bureau working paper using its 2026 BTOS AI supplement found that among US firms that adopted AI between November 2025 and January 2026, the scope of use stayed narrow: 57% of AI-adopting firms use it in three or fewer business functionsUS Census Bureau · April 202644% of US firms using AI report it augmenting workers' existing tasks rather than replacing them, versus just 5% reporting AI substituting tasks outright, according to Census Bureau research on business AI adoption 44%U.S. Census Bureau · 2026 BTOS AI supplementAmong U.S. firms where AI is substituting worker tasks, the share reporting a 'large number' of tasks replaced rose from 2.5% in the Dec 2023-Feb 2024 Census BTOS AI supplement to 7% in the Nov 2025-Jan 2026 supplement Share of AI-task-substituting US firms reporting large-scale task replacement rose from 2.5% to 7%US Census Bureau BTOS AI Supplement · April 2026The Census Bureau's Business Trends and Outlook Survey found AI adoption climbing fastest among big employers: as of May 2026, 37% of firms with 250+ staff used AI, 32% at 100-249, under 20% at the smallestUS Census Bureau · May 2026The Census Bureau's Business Trends and Outlook Survey also breaks AI adoption down by industry — in the wave ending 3 May 2026, the Information sector (telecom, media, tech services) led every US industry, with 39.7% of firms using AI, more than double the 19.8% national averageU.S. Census Bureau · May 2026 (BTOS)A Census Bureau working paper covering AI use at over 117,000 US firms found AI-related employment decreases occurred at only 2% of firms, with job cuts linked to firm-wide AI deployment rather than individual worker AI use Only 2% of 117,000+ US firms cut jobs after adopting AI, Census Bureau findsUS Census Bureau (CES-WP-26-25) · Apr 202655% of U.S. workers say they have used AI on the job for at least one of 11 listed tasks, and among those who used it in the past week, 31% said it saved them one to two hours of work, per the Census Bureau's March 2026 Household Trends and Outlook Pulse Survey 55% of U.S. workers have used AI on the job, Census Bureau survey findsU.S. Census Bureau · March 2026 HTOPSA Census Bureau working paper using Quarterly Workforce Indicators administrative payroll data found that early-career employment — ages 22 to 24 — in the most AI-exposed industries fell, over the ten quarters since ChatGPT's release, by a regression-adjusted 12% decline in early-career (22-24) US employment in the most AI-exposed industries since ChatGPT's launchU.S. Census Bureau · April 2026 (CES-WP-26-27)A Census Bureau working paper on the 2026 AI supplement to the Business Trends and Outlook Survey found that, among firms using AI in any business function, Sales and Marketing was the single most common use case, ahead of Strategy and Business Development at 45% and IT at 41% 52% of AI-adopting firms use it in Sales and MarketingU.S. Census Bureau · CES WP, Apr 2026Nearly 750 executives in the Duke/Federal Reserve CFO Survey expect AI to barely touch total headcount — aggregate employment down less than 0.4% in 2026. What shifts is the mix, over the next three years: routine clerical work down more than 2 percentage points of employment shareNBER w34984, The CFO Survey · 2026Q1 survey waveChatGPT was estimated to have reached 100 million monthly users roughly two months after launchUBS, reported by The Register · February 2023The total computing power of the stock of AI chips is growing about 3.4× per yearEpoch AI · 2026Ciena, the pure-play optical networking supplier wiring hyperscale AI data centers together, just posted a record fiscal third quarter — revenue for the quarter ended 1 August 2026 reached $1.67 billion in quarterly revenue, up 37% year-over-yearCiena Corporation · Q3 FY2026CL-Bench, a new benchmark testing whether AI agents actually improve with experience across sequential real-world tasks, found that even the best system -- Claude Sonnet 4.6 using full-context in-context learning -- achieved only a 25.4% normalized gain over its own stateless baseline Best AI agent's continual-learning gain on CL-Bench: 25.4%Asawa et al. (UC Berkeley/Snorkel AI), arXiv · June 2026A UC Berkeley, Sao Paulo and Stanford study of 141 countries found that national Claude usage rises 12x for every 0.10-point increase in a country's AI exposure score, based on Anthropic's Economic Index data for February 2026 Claude usage rises 12x per +0.10 rise in national AI exposureUC Berkeley/Stanford study · arXiv, Aug 2026Cloudflare's capital expenditure fell to 7% of revenue in the quarter to 30 June 2026, down from 12% a year earlier, even as revenue grew 36% year-over-year to $696 million 7% of revenue, down from 12% a year earlierCloudflare Q2 2026 earnings release · Q2 2026Epoch AI asked LLM judges how long an experienced engineer would need, without AI, to reproduce each merged pull request from the 41 collaborators on OpenAI's public Codex repository. In the second quarter of 2026, work beyond what anyone could finish in a single unassisted day accounted for 8% of contributor-days, up from 2% a year earlierEpoch AI · Q2 2026Coherent's Datacenter & Communications segment, the optical transceivers that link GPUs inside AI data centers, brought in $1.615 billion in the quarter to 30 June 2026, up 58.6% from $1.018 billion a year earlier as demand for 800G and 1.6T transceivers accelerated Coherent's AI-datacenter segment revenue: $1.615 billion, up 58.6% year-over-yearCoherent Corp · Q4 FY2026 (June 2026)Coherent's Datacenter & Communications segment, which supplies optical transceivers for AI data centers, generated $1.62 billion in revenue in the quarter ended June 30, 2026, up 58.6% from a year earlier $1.62 billionCoherent Corp. · Q4 FY2026CoreWeave's capital expenditures for Q2 2026 (April-June) totaled approximately $6.42 billion, derived from its SEC 10-Q filings as the $14.12 billion six-month total less the $7.70 billion spent in Q1 2026 CoreWeave's Q2 2026 capital expenditures: approximately $6.42 billionCoreWeave, Inc. (SEC 10-Q) · Q2 2026Core Scientific has contracted about 1.1 gigawatts of leased data-center power capacity to AI and HPC customers, including CoreWeave and AMD, representing more than $24 billion of potential revenue, the company said in its second-quarter 2026 results Core Scientific's leased AI data-center capacity: 1.1 gigawatts, over $24 billion in contractsCore Scientific · Q2 2026Credo Technology, whose active electrical cables and optical DSPs connect AI accelerators inside hyperscaler racks, reported $479.0 million in quarterly revenue, up 114.7% year-over-yearCredo Technology · Q1 FY2027Credo Technology makes the active electrical cables and optical DSPs that connect AI accelerators inside hyperscaler racks, and demand for that interconnect gear pushed its revenue for the quarter ended 26 April 2026 up 157.0% year-over-year to $437.0 million in quarterly revenueCredo Technology · Q4 FY2026Salesforce brands its whole product line around AI agents and reports separately what customers actually pay for them. In the quarter to 30 April 2026, on total revenue of $11.13 billion, that came to $1.2 billion of Agentforce annual recurring revenue, up 205% year on yearSalesforce Q1 FY2027 results · Quarter to 30 April 2026Salesforce brands its whole product line around AI agents and reports separately what customers actually pay for them. In the quarter to 31 July 2026, Agentforce annual recurring revenue exceeded $1.5 billion in Agentforce ARR, up over 240% year-over-yearSalesforce Q2 FY2027 results · Q2 FY2027Salesforce's quarter to 31 July 2026 brought total revenue of $11.3 billion, up 11% year-over-year, with subscription and support revenue of $10.8 billion providing the bulk of the growth $11.3 billion in Q2 FY2027 revenueSalesforce · Q2 FY2027CoreWeave rents out AI computing capacity, and building that capacity is capital-intensive. Its first-quarter 2026 cash flow statement shows capital expenditure on property and equipment, including capitalised internal-use software, jumped to $7.7 billion in Q1 2026 capital spending, a 5.5x jump from $1.4 billion a year earlierCoreWeave · Q1 2026Datadog's own 10-Q filing shows AI is now a material swing factor in its growth: for the quarter ended 30 June 2026, its AI-native customer cohort — which includes its single largest customer — contributed high-single-digit percentage points of Datadog's total year-over-year revenue growthDatadog Q2 2026 10-Q · Q2 2026Dell's AI server business booked a record $60.9 billion in orders and exited its second fiscal quarter of 2027 with a record $95 billion backlog, on revenue of $16.4 billion $95 billion backlogDell Technologies · Q2 FY2027 (ended Aug 1, 2026)Dell's quarter to 28 May 2026 was the one where AI servers took over the business: total revenue hit a record $43.8 billion, up 88% year on year, and of that, $16.1 billion came from AI servers aloneDell Technologies 8-K · Q1 FY2027, ended 28 May 2026Data-center landlords are racing to keep pace with AI and cloud demand. Digital Realty's second-quarter 2026 financial supplement to investors disclosed a global construction pipeline of Digital Realty's 1,402 megawatts of IT capacity under construction, 54% already pre-leasedDigital Realty Q2 2026 supplement · Q2 2026Digital Realty ended Q2 2026 with a record $1.9 billion backlog of signed-but-not-commenced data center leases, driven by surging AI and hyperscale demand $1.9 billion data center lease backlogDigital Realty Trust · Q2 2026 (reported July 23, 2026)As cloud providers shifted from AI training to inference, conventional DRAM contract prices jumped 93% to 98% in a single quarter — and on TrendForce's count of the first quarter of 2026, global DRAM revenue hit $97 billion, up 81%TrendForce · Q1 2026Epoch AI's Capabilities Index scores frontier models on one aggregate scale, and fits a break in the trend at April 2024, when reasoning models and reinforcement learning took over frontier training. Before the break the frontier gained 8.3 index points a year; after it, 15.5 points a year, 1.85 times as fastEpoch AI · December 2025NBER asked 69 economists what the US economy would look like in 2050 if AI capabilities advance rapidly. Their median answer was 3.5% annual GDP growth, alongside a labour force participation rate falling from today's 62% to 55.0% by 2050, the level before women entered the workforce en masseNBER working paper 35046 · April 2026AI data centres are pushing US electricity generation to record highs. Total output from the country's power sector in the first half of 2026 rose 1.8% versus the first half of 2025, adding 37 billion kilowatt-hoursEIA, Short-Term Energy Outlook · Aug 2026 STEOEIA modelled what would happen if data-centre power demand grew faster than forecast. Its February 2026 baseline had US coal generation falling 9.3% between 2025 and 2027; in the higher-demand scenario, coal plants run on their existing spare capacity and the decline shrinks to a 5.0% fall in US coal generation instead of 9.3%EIA Short-Term Energy Outlook analysis · March 2026EIA's September 2026 Short-Term Energy Outlook forecasts commercial-sector electricity sales climbing 3.3% in 2026, accounting for 63% of that year's total US electricity-sales growth, with data-centre development cited as a key driver 63% of all US electricity-sales growth in 2026EIA Short-Term Energy Outlook · September 2026EIA's Annual Energy Outlook 2026 projects that in its High Electricity Demand case, US data-centre servers alone — before any cooling or other building overhead — could draw as much as 818 billion kWh of server electricity a year by 2050US Energy Information Administration · AEO2026Data-centre development is now the main driver of new US electricity demand, and the EIA's September 2026 Short-Term Energy Outlook forecasts US electricity generation climbing 2.2% in 2026 to a record, before rising a further 1.7% in 2027, to 4,368 billion kilowatt-hours of US electricity generation in 2026U.S. EIA · Sep 2026 STEOEIA's Annual Energy Outlook 2026 estimates that data-centre servers alone, before any cooling or other building load, already accounted for 7% of all US commercial-sector electricity in 2025US Energy Information Administration · AEO2026, May 2026EIA's September 2026 Short-Term Energy Outlook, which attributes record US electricity demand to data-center development, forecasts the average residential electricity price rising from 17.30 cents per kilowatt-hour in 2025 to 18.20 cents in 2026 and 18.59 cents in 2027 18.20 cents per kilowatt-hourEIA · September 2026 STEOEIA's September 2026 Short-Term Energy Outlook projects the West South Central region (Texas, Oklahoma, Louisiana, Arkansas) will drive nearly 20% of total U.S. electricity sales growth in 2026 and nearly 40% in 2027, propelled largely by data center development despite Texas's grid-connection pause for new data center projects West South Central region accounts for nearly 40% of 2027 U.S. electricity sales growth (EIA)EIA Short-Term Energy Outlook · September 2026On 3 August 2026 the governor of Texas paused new data-centre development while projects already under review are assessed. In the same month's Short-Term Energy Outlook the EIA cut its forecast for Texas electricity load growth in 2027 to 6% growth in Texas load, down from 14% forecast a month earlierEIA Short-Term Energy Outlook · August 2026Texas paused new data-centre development on 3 August 2026 while it reviews the projects already in the queue, and the EIA cut its forecast for the state in response: Texas electricity load growing 6% in 2027, not the 14% forecast a month earlierEIA Short-Term Energy Outlook · August 2026Data centers are reshaping state-level power demand — the EIA reports commercial electricity sales in Virginia, home to the world's largest concentration of data centers in PJM's Dominion zone, rose by nearly 30.0 million megawatthours between 2019 and 2025, the fastest growth of any US state except TexasU.S. Energy Information Administration · 2019-2025NBER's Firm Data on AI paper found a striking gap in expectations: while executives surveyed expect AI to cut their firms' employment by 0.7% over the next three years, US employees surveyed separately in December 2025 predicted AI would raise US firms' employment about 0.5% over the next three yearsNBER (Yotzov et al.) · March 2026AI inference is consuming NAND flash faster than suppliers can allocate it. Enterprise SSD contract prices rose roughly 80% in the first quarter of 2026, and on TrendForce's count the top five brands' combined revenue hit a record $18.46 billion — enterprise SSD revenue rose 86.1% in a single quarterTrendForce · Q1 2026Epoch AI modeled the total cost of ownership of a typical 1-gigawatt AI data center built by a US hyperscaler around Nvidia's GB200 NVL72 systems, and found the up-front capital expenditure alone comes to $38 billion for one hyperscale AI data centerEpoch AI · May 2026Epoch AI's AI Data Centers tracker now covers an estimated 46% of the world's deployed AI computing capacity, up from 26% as of March 2026, spanning 87 sites with a combined 13.2 gigawatts of IT power and about 13.7 million H100-equivalent GPUs, as of 21 September 2026 46% of the world's deployed AI computing capacityEpoch AI · 21 September 2026Investment in AI-related data center construction, compute hardware and networking equipment reached about 0.8% of US GDP in the first quarter of 2026, pushing total computing-infrastructure investment to roughly 1.5% of GDP — more than double the 2015-2022 average of about 0.7%, per Epoch AI AI infrastructure investment: ~0.8% of US GDP in Q1 2026Epoch AI · Q1 2026The 84 AI data centers tracked by Epoch AI have a combined real-world power footprint of 16.9 gigawatts as of August 2026, more than New York City's 11 GW peak demand 16.9 GWEpoch AI · August 2026Epoch AI tested three commercial AI detectors — Pangram, GPTZero, Originality.ai — pinned to their June 2026 versions. Against plain machine-generated text they catch nearly everything, but shown five samples of a specific author's writing and told to imitate the style, AI detectors missed 13% of style-imitated AI passages, versus under 1% of plain AI textEpoch AI · Jul 2026Epoch AI rebuilt the cost structure of three AI companies — Anthropic, MiniMax and Z.ai — from IPO filings and reporting. Staff, despite the industry's headline pay packages, take under a quarter of spending. Research and inference compute together take 54% to 62% of an AI company's costsEpoch AI · 2025 accountsEpoch AI finds a $2,500 consumer gaming GPU (RTX 5090) can now run open models matching frontier AI performance from 6 to 12 months earlier, with the lag ranging from 6.3 months on the Artificial Analysis Intelligence Index to 12.4 months on LM Arena Elo 6 to 12 monthsEpoch AI · 2026Epoch AI fits a trend through the models whose context windows ranked among the ten longest on their release date, and finds that since mid-2023 the frontier context window has been getting about 30× longer every yearEpoch AI · since mid-2023Frontier models have started finding software flaws at scale. Counting the high- and critical-severity CVEs disclosed by 21 major vendors — Microsoft, Google, Apple and Cisco among them — Epoch AI puts July 2026 at about 2,500 serious CVEs in July 2026, five times the monthly record set before April 2026Epoch AI · July 2026Epoch AI's tracked data-center dataset captures only 26% of global AI computing capacity as of March 2026, covering 5.2 million H100-equivalent chips versus an estimated 20.2 million H100-equivalent chips sold through 2025 Epoch AI's data-center dataset covers just 26% of global AI computing capacity as of March 2026Epoch AI · March 2026Epoch AI tracks the largest AI data centre in the world by IT power capacity, not raw compute -- and on that measure the record has roughly doubled every 10 months since mid-2024, climbing from Google's 237-megawatt Papillion facility to xAI's Colossus 2, now at 946 megawatts of IT power capacityEpoch AI · Sept 2026Epoch AI had frontier models repeatedly play the board game Earthborne Rangers to test whether they improve with practice, and found GPT-5.5 and Claude Opus 4.8 both mismanaged its core fatigue mechanic at 2.1 fatigue per round, versus 0.6 for expert humans and 3.5 for purely random playEpoch AI · July 2026Epoch AI finds the frontier of its Capabilities Index has advanced 14 points a year for reasoning models since September 2024, more than double the 6 points a year non-reasoning models achieved from GPT-4 to GPT-4.5 14 points a year, vs 6 for non-reasoning modelsEpoch AI · September 2026Epoch AI tracked 333 compute estimates for notable AI models released between 2010 and May 2024, and found that training compute for the frontier — the top 10 models by compute at release — has been growing each year to reach 5.3x the prior year's levelEpoch AI · 2010–May 2024 dataOpenAI, Anthropic and xAI are the three labs most associated with sparking the AI boom, yet Epoch AI estimates the trio together controlled well under 4 million H100-equivalent GPUs at the end of 2025 — only 20% to 30% of the world's total operational AI computing powerEpoch AI · End of 2025Epoch AI's FrontierMath Erdős benchmark tasks AI models with proving or disproving 68 significant unsolved Erdős problems in Lean, each within a fixed $300, 72-hour budget. Running five frontier models in August 2026, only one produced any verified solution: GPT-6 Astra solved 2 of 68 problems (3%) — every other model tested scored 0%Epoch AI · Aug 2026Epoch AI audited its FrontierMath benchmark after OpenAI flagged suspiciously high scores, and its 12 June 2026 v2 release corrected errors in 42% of the original problems, leaving a verified set of 338 problems across four tiers 42% of FrontierMath's original problems contained errorsEpoch AI · June 2026 (v2 release)Epoch AI modeled the power budget of a frontier AI data centre built around Nvidia's GB200 architecture, and found the GPUs themselves -- the part everyone pictures when they think 'AI compute' -- account for only about 40% of total facility power at peakEpoch AI · Dec 2025Epoch AI estimates Huawei will produce less than 4% as much AI compute, measured in H100-equivalents, as Nvidia in 2026 -- and without access to foreign high-bandwidth memory, that share could fall to around 1% by 2028 under 4% of Nvidia's AI compute in 2026Epoch AI · Sep 2026AI capability is getting cheaper to reach, not just bigger. Epoch AI tracked what it cost in tokens to hit the same score — about 27% accuracy on its FrontierMath benchmark — first with o4-mini in April 2025, then with GPT-5.2 in December 2025, and found the same capability got ~3x cheaper in eight months, a ~5-10x-a-year declineEpoch AI · Feb 2026Epoch AI and Ipsos asked 469 US workers who had used AI for work in the previous seven days whose subscription they were on. In computer, engineering and science jobs 67% use a tool their employer provides — everywhere else the free tier wins: 46% of US workers using AI at work are on a free plan, against 37% given a tool by their employerEpoch AI / Ipsos · Fieldwork 10-19 July 2026Epoch AI and Ipsos surveyed 1,106 employed US adults between 10 and 19 July 2026, on the Ipsos KnowledgePanel, and found 20% say AI now handles a task they used to hand off to a human coworker or contractorEpoch AI / Ipsos · Jul 2026Epoch AI compared how OpenAI's and Anthropic's flagship models handle very long prompts, and found GPT-5.6 Terra and Sol slow down quadratically as context grows while Claude Sonnet 5 and Opus 5 stay almost linear -- the marginal latency to process the next 10,000 tokens for GPT-5.6 Terra rose from 0.071 seconds at a 100,000-token prompt to 1.66 seconds at a 10-million-token promptEpoch AI · Sept 2026Epoch AI's new AI Chip Users explorer estimates Meta Superintelligence Labs held about 996,000 H100-equivalent GPUs of AI computing capacity as of December 2025, putting it behind Google DeepMind and Anthropic among the frontier labs it tracks Meta's ~996,000 H100-equivalent GPUsEpoch AI · December 2025Epoch AI and METR's new MirrorCode benchmark found the top AI model correctly rebuilt real software from behavior alone in 56% of cases, up from about 30% for leading models roughly eight months earlier Top AI model now solves 56% of Epoch AI's toughest long-horizon coding tests, up from about 30% eight months agoEpoch AI / METR (MirrorCode benchmark) · Jul 2026The share of US adults using AI at least six days a week more than doubled in six months, rising from 8.2% in March 2026 to 18.7% in August 2026 Near-daily US AI use rose from 8.2% to 18.7% in six monthsEpoch AI / Ipsos polling · Mar–Aug 2026Equinix, the world's largest data-center operator, ties its buildout of hyperscale capacity directly to demand it links explicitly to the adoption of AI. Its capital expenditure in the quarter ended 30 June 2026 reached Equinix's Q2 2026 capital expenditures: $1.578 billion, up 59.6% year-over-yearEquinix · Q2 2026Eaton's Electrical Global segment, which sells the power distribution and switching gear AI data centres depend on, ended the second quarter of 2026 with total backlog up 103% year-over-year Eaton Electrical Global backlog up 103% year-over-yearEaton · Q2 2026Asked what three years of AI had actually changed inside their own firms, nearly 6,000 senior executives across the US, UK, Germany and Australia mostly reported nothing: nine in ten saw no impact on employment or productivityNBER working paper 34836 · February 2026Fabrinet, the Thai contract manufacturer that assembles the optical transceivers linking Nvidia's and other hyperscalers' AI clusters, just posted a record quarter — revenue for the fiscal fourth quarter, ended 26 June 2026, hit $1.32 billion in quarterly revenue, up 45% year-over-yearFabrinet · Q4 FY2026Of the notable AI models released in 2025, Stanford's AI Index counts over 90% as coming from industryStanford HAI, AI Index 2026 · 2025Across 42 notable language models, Epoch AI estimates frontier training costs have risen from roughly $2 million for GPT-3 to up to nearly $390 million per runEpoch AI · 2024Epoch AI's Open Problems benchmark collects research mathematics problems that have resisted serious attempts by professional mathematicians. Expanded to 50 problems on 31 July 2026, the count two weeks later — with the newest AI solution still marked provisional — stood at 4 solved by AI, 1 by a human, 45 still openEpoch AI FrontierMath · 17 August 2026Alphabet told investors in its second-quarter 2026 results that Gemini models now process 22 billion API tokens a minute, and that the Gemini app has reached 950 million monthly active usersAlphabet 8-K · Q2 2026Brynjolfsson et al. asked US adults what they'd need to be paid to give up generative AI for a month, and Stanford's 2026 AI Index reports their estimate that total US consumer surplus grew from $112 billion in 2025 to $172 billion in annual consumer surplus by early 2026Stanford AI Index (Brynjolfsson et al.) · early 2026In a randomised field experiment across 66 firms and 7,137 knowledge workers, a generative AI tool built into everyday work apps saved the people who used it two hours a week on email, with no detected change in their tasksNBER working paper 33795 · revised Nov 2025Mentions of generative AI skills in US AI job postings grew 111% from 2024 to 2025, even as their share of total AI postings slipped, as employer demand pivots toward agentic-AI skills, per Lightcast data cited in Stanford's AI Index 111%Stanford HAI · 2025Across 5,179 customer-support agents, a generative AI assistant raised issues resolved per hour by 14% on average — and by 34% for the least experienced staffBrynjolfsson, Li & Raymond, NBER · April 2023GE Vernova builds the gas turbines and grid equipment that electricity supply runs on, and AI data centres are now one of its largest sources of new demand. Its gas equipment under contract rose from 100 GW to 116 GW during the second quarter of 2026, and data-centre orders through the first half of the year reached over $5 billion of data-centre orders, more than double the whole of 2025GE Vernova Q2 2026 results · First half of 2026GE Vernova's Electrification segment, which sells the transformers and switchgear data centres need to connect to the grid, booked over $5 billion in data-centre equipment orders in the first half of 2026 alone -- more than its entire 2025 total GE Vernova's H1 2026 data-centre orders: over $5 billion, beating all of 2025GE Vernova · Q2 2026A large-language-model analysis of 18,797 work tasks across 124 economies found automation-exposed task shares ranging from 3.3% in the least-exposed economy to 61.6% in the most, rising with income but varying widely within income groups 3.3% to 61.6% across 124 economiesGlobal Automation Atlas (Garg, Crosta & Baier) · 2026Alphabet's most recently reported quarterly cash spending on property and equipment reached $44.92 billionAlphabet, SEC EDGAR (XBRL filings) · Q2 2026Alphabet's cash capital expenditure in the first quarter of 2026 reached $35.67 billion in a single quarterAlphabet 10-Q, SEC EDGAR · Q1 2026Alphabet's quarterly cash capital expenditure has roughly tripled in two years, from $12.01 billion in the first quarter of 2024 to $35.67 billion in the first quarter of 2026Alphabet 10-Q, SEC EDGAR · Q1 2024 - Q1 2026Epoch AI estimates the single largest owner of AI computing power worldwide is not a chipmaker or a cloud reseller but Google, which held, measured in H100-equivalents as of the fourth quarter of 2025, about a quarter of all global AI compute, mostly its own TPUs rather than Nvidia GPUsEpoch AI · Q4 2025Epoch AI's Capabilities Index scores language models on one aggregate scale, and nothing since has held first place as long as GPT-4 did after its March 2023 release — roughly a year. The second-longest lead, by OpenAI's o1, lasted a little over three months, less than a third of GPT-4's run at the topEpoch AI Capabilities Index · July 2026OpenAI's GPT-6 Astra scores 97.6% on Epoch AI's FrontierMath Tier 4 (v2) math benchmark, versus 87.8% for Anthropic's Claude Fable 5.1 GPT-6 Astra: 97.6% on FrontierMath Tier 4 (v2), vs Claude Fable 5.1's 87.8%Epoch AI · September 2026Traditional computing powers like Germany, Japan and France now play marginal roles in AI infrastructure. Epoch AI's tracking of global GPU cluster performance in May 2025 put China second at 15%, behind a United States hosting about three-quarters of the world's GPU cluster performanceEpoch AI · May 2025US private AI investment reached $285.9 billion in 2025 — 23.1 times China's $12.4 billion and 48.5 times the UK's $5.9 billion — per Stanford's 2026 AI Index 23.1 timesStanford AI Index (Quid data) · 2025High-bandwidth memory is taking over the memory industry's capacity: TrendForce expects HBM to account, among the top three DRAM suppliers by the end of 2026, for 22% of DRAM wafer input, up from 18% a year earlierTrendForce · June 2026 projectionAmazon, Google, Meta, Microsoft and Oracle together held, measured in H100-equivalents as of Q4 2025, an estimated 71% of the world's AI computeEpoch AI · Q4 2025A January 2026 IMF staff discussion note tracking the spread of AI-related job skills across US local labor markets found that in occupations highly exposed to AI but with little scope to complement it, five years after AI-skill demand appeared, employment was 3.6% lower than in other regionsIMF · Jan 2026About 20% of Finland's workforce holds jobs with high AI exposure but low complementarity — including software developers, finance professionals and administrative staff — putting them at greater risk of job displacement, per IMF analysis of 2023 EU labor data 20%IMF · 2026 Selected Issues PaperAn IMF staff analysis of Israel's labour market found that 23.1% of national employment sits in occupations with high AI exposure but low task complementarity -- the group most vulnerable to AI-driven displacement rather than augmentation 23.1% of Israeli jobs: high AI exposure, low complementarityIMF Selected Issues Paper · July 2026Epoch AI models the world's AI inference capacity, at fixed model size and context length, as growing 3.4× a year — but it puts token demand growth at around 10× a year, three times faster than supplyEpoch AI · May 2026Innodata's data engineers annotate, label and curate the training data behind frontier AI models, and demand for that work keeps climbing. Its second-quarter 2026 revenue hit a record $92.1 million, up 58% year-over-yearInnodata Inc. · Q2 2026Intel's Data Center and AI segment, the unit selling Xeon server chips and Gaudi/Crescent Island AI accelerators, reported revenue for the quarter ended 27 June 2026 up 59% year-over-year to $6.3 billion in quarterly segment revenueIntel Corporation · Q2 2026IREN, the bitcoin miner turned AI cloud operator, said its AI Cloud Services revenue grew roughly 8x year-over-year to $128.8 million in the fiscal year ended June 30, 2026, and the company now has $4 billion of ARR contracted for 2026 GPU capacity, with $1 billion already operatingIREN FY26 results release · FY2026IREN, a bitcoin miner turned AI cloud operator, told investors in its quarter ended 31 March 2026 that it now holds executed grid connection agreements across seven data-centre sites in Texas, Oklahoma and British Columbia for a combined 4,510 megawatts of AI data-centre power capacityIREN Ltd · Q3 FY2026 10-QIron Mountain is building AI-ready colocation data centres alongside its legacy records-storage business, but its own quarterly filing shows growth capital expenditure on that data-centre build-out fell in the six months to 30 June 2026, from $953.0 million a year earlier, to $880.1 million in data-centre growth capex, down 8%Iron Mountain (SEC 10-Q) · H1 2026KLA Corporation, whose process-control and inspection tools qualify AI chip production before wafers ship, posted fiscal fourth-quarter 2026 results of $3.66 billion in revenue, up 15% year-over-yearKLA Corporation · Q4 FY2026Epoch AI tracks the largest AI data centre in the world by computing capacity, and the record has changed hands repeatedly since August 2024 — Colossus, Anthropic and Amazon's New Carlisle, Microsoft's Fairwater Atlanta, Meta's Prometheus. Each time, the bar has risen at the same pace: a doubling every seven monthsEpoch AI · June 2026The largest AI data centre in Epoch AI's tracker, Colossus 2 in Memphis, Tennessee, holds a computing capacity of 1,112,000 H100-equivalent AI chipsEpoch AI · AI data centers tracker, August 2026Nvidia's accelerators have pushed thermal design power from 700W for the H100 to over 1,000W for the B200 and B300, and air can no longer carry the heat away. TrendForce expects liquid cooling in server racks to reach 47% of server racks by 2026TrendForce · November 2025 forecast for 2026Lumentum makes the optical components that link AI accelerators inside data centres, and reported fiscal fourth-quarter 2026 results of $1.01 billion in net revenue, up 109% year-over-yearLumentum Holdings · Q4 FY2026Lam Research, whose machines etch and deposit the layers inside AI chips, reported fiscal 2026 revenue of $23.23 billion — up 26% from $18.44 billion in fiscal 2025 — as CEO Tim Archer credited AI-driven demand for record margins and EPS $23.23 billion in fiscal 2026 revenueLam Research · FY2026 (ended June 28, 2026)A McKinsey survey of employers cited in Stanford's 2026 AI Index found anticipated AI-driven workforce cuts concentrated in service operations, supply chain and software engineering, while almost half of organizations expect little or no change overall. Still, one-third of employers expect AI to shrink their workforce in the year aheadMcKinsey survey (Stanford AI Index 2026) · 2025 surveyMeta's purchases of property and equipment nearly doubled in 2025, from $37.3bn to $69.7bn — 34.7% of its revenueMeta FY2025 Form 10-K · FY2025Meta's cash capital expenditure went the other way in the first quarter of 2026 — from $21.38 billion the quarter before, a fall to $19.0 billionMeta 10-Q, SEC EDGAR · Q1 2026Meta's most recently reported quarterly purchases of property and equipment reached $30.12 billionMeta, SEC EDGAR (XBRL filings) · Q2 2026In a randomised trial, 16 experienced open-source developers believed AI tools had made them 20% faster. Measured against issues they finished without AI, they were 19% slowerMETR · July 2025Micron breaks out the memory it sells into AI data centres as its own business unit. In the quarter to 28 May 2026 that unit took in $11.52 billion, against $1.53 billion in the same quarter a year before — Micron's core data centre revenue grew 7.5× in a yearMicron 8-K, SEC · quarter to 28 May 2026Micron plans to add up to 60,000 HBM wafers per month of capacity, reaching roughly 100,000 wafers/month by end-2026, per TrendForce -- still far below Samsung's and SK hynix's respective ~150,000-200,000 wafers/month Micron's HBM wafer capacity is set to reach ~100,000 wafers/month by end-2026, per TrendForceTrendForce · September 2026MIT researchers tracking corporate AI adoption found that in 2025, 11% of S&P 500 firms had AI deeply integrated into core business processes and a further 10% were using it in the production of goods or services -- both figures have surged since 2022, when deep integration stood at just 5% of S&P 500 firmsMIT researchers (Yu et al.) · 2025Modine Manufacturing, the thermal-management maker whose liquid-cooling systems keep hyperscale AI server racks from overheating, saw hyperscaler demand keep surging — first-quarter fiscal 2027 Data Centers segment sales reached $348.6 million in Data Centers segment sales, up 90% year-over-yearModine Manufacturing · Q1 FY2027Monolithic Power Systems makes the chips that step grid power down to the voltages AI GPUs need. Its Enterprise Data segment, which sells into AI servers, grew far faster than the rest of the company in the quarter to 30 June 2026, when total revenue hit a record $980.6 million and Enterprise Data segment revenue reached $380.6 million in AI server chip revenue, up 164.3% year-over-yearMonolithic Power Systems · Q2 2026Marvell designs the custom accelerators, optics and switching silicon that wire AI clouds together, and that business has crowded out everything else it sells. In the quarter to 2 May 2026, data centre revenue of $1.83 billion out of $2.42 billion total came to 76% of Marvell's revenueMarvell Technology · Q1 FY2027 (to 2 May 2026)Marvell's record quarter to 2 May 2026 looks very different depending on which profit line you read. Non-GAAP net income was $718.0 million. Under GAAP — after $207.6 million of stock compensation, $225.2 million of amortisation on acquired intangibles and a $250.7 million contingent-consideration revaluation — net income was $34.5 million, under 5% of the adjusted figureMarvell Technology · Q1 FY2027 (to 2 May 2026)Marvell's revenue hit a record $2.74 billion in the quarter to 1 August 2026, up 37% year-on-year, led by an AI data-centre business that grew even faster — accelerating to 46% year-on-year, to $2.17 billion in Data Center revenueMarvell Technology, Inc. · quarter ended 1 August 2026Marvell's data infrastructure chip business set a fiscal second-quarter revenue record in the three months to 1 August 2026, with net revenue up 37% year-over-year to $2.739 billion, as Data Center segment revenue growth accelerated to 46% $2.739 billion in record Q2 revenueMarvell Technology · Q2 FY2027Microsoft told investors in April 2026 that its AI business had reached an annual run rate of $37 billionMicrosoft FY2026 Q3 earnings call · April 2026Microsoft's cash capital expenditure rose by more than half between the third and fourth quarters of 2025, reaching $29.88 billion in a single quarterMicrosoft 10-Q, via SEC XBRL · Q4 2025Microsoft's most recently reported quarterly cash spending on property and equipment reached $35.80 billionMicrosoft, SEC EDGAR (XBRL filings) · Q2 2026Microsoft's cash spending on property and equipment reached, in the first quarter of 2026 alone, $30.88 billionMicrosoft 10-Q, via SEC XBRL · Q1 2026A randomised online experiment gave 1,174 adults aged 25 to 45 an incentivised business problem-solving task, some with a generative-AI assistant and some without. Unaided, the higher-education group beat the lower-education group by 0.548 standard deviations. With AI the gap fell to 0.139 — about three-quarters of the education gap closedNBER Working Paper 34851 · May 2026 revisionMenaka Hampole and co-authors linked resume data to firms' AI-application adoption to isolate AI's true effect on jobs, not just correlation. In their preferred instrumental-variables specification, a one-standard-deviation rise in AI exposure cuts within-firm employment share by 14%NBER · Working Paper 33509, 2025A National Bureau of Economic Research working paper, drawing on the Federal Reserve Banks of Atlanta and Richmond's CFO Survey of roughly 750 corporate executives, found that by the end of 2025 more than 40% of US companies still had not invested in AI at all, as of year-end 2025NBER (Atlanta/Richmond Fed CFO Survey) · March 2026 (NBER WP 34984)A Danish register study spanning 25,000 workers across 7,000 workplaces in 11 AI-exposed occupations found widespread chatbot adoption and new AI-related tasks, but a difference-in-differences design ruled out earnings or hours effects larger than 2%, at either the worker or workplace level, two years after ChatGPT's launch no earnings or hours effect above 2%, two years after ChatGPT's launchNBER Working Paper 33777 · revised Mar 2026A survey of nearly 6,000 senior business executives across the US, UK, Germany and Australia — run by the Bank of England, the Federal Reserve Bank of Atlanta, the Bundesbank and Macquarie University — found that on average 69% of businesses already use AINBER · Working Paper 34836, Feb 2026Firms across the US, UK, Germany and Australia expect AI to lower their employment by about 0.7% over the next three years, with the steepest expected declines in the UK (-1.4%) and US (-1.2%) Firms expect AI to cut employment ~0.7% over the next 3 yearsNBER Working Paper 34836 · Feb 2026, rev. Mar 2026Workplace generative AI use pays off in wages, not just output. Analyzing 2024 Current Population Survey data and controlling for demographics, occupation and industry, NBER researchers Alexander Bick, Adam Blandin and David Deming found that, compared with non-users, daily generative-AI users earn 40% more, and occasional users 22% moreNBER (Bick, Blandin & Deming) · Sep 2024, rev. Feb 2025NBER's 'Mind the Gap' working paper finds 43% of US workers used AI for their job in early 2026, versus 32% of workers surveyed across Germany, UK, France, Italy, Sweden and the Netherlands 43% of US workers used AI for work in 2026, versus 32% across EuropeNBER Working Paper 34995 · Jan-Feb 2026Nebius Group, the AI-infrastructure and GPU-cloud operator building out data centres for AI training and inference, spent $5.7 billion on capital expenditure in the quarter to 30 June 2026, driven mainly by GPU purchases and data-centre expansion. Across the first half of 2026 it spent a total of $8.1 billion on AI infrastructure capital expenditureNebius Group · H1 2026Nebius Group, the AI-infrastructure and GPU-cloud operator, keeps raising its own build-out target for AI data-centre power. In its second-quarter 2026 shareholder letter, up from a forecast of more than 4 gigawatts the quarter before, the company said it now expects to end 2026 with 5 gigawatts of contracted AI data-centre powerNebius Group · Q2 2026 shareholder letterNebius rents out AI cloud capacity and is scaling from a small base. Revenue in the second quarter of 2026 reached $582.3 million against $105.1 million in the same quarter of 2025 — a 454% year-on-year rise in revenueNebius Group · Q2 2026NRG Energy agreed on principal commercial terms with a global cloud and AI hyperscaler to build a 1.2 GW combined-cycle natural gas power plant in Texas to supply a new data centre 1.2 GW gas power plant for a hyperscaler's Texas data centreNRG Energy · Q2 2026In July 2025, Nvidia became the first US company in history to reach a $4 trillion market capitalisationAl Jazeera · July 2025Nvidia's Data Center segment brought in $193.7bn of the $215.9bn it made in its 2026 financial year — 89.7% of the company's revenueNvidia FY2026 Form 10-K · FY2026, ended 25 January 2026Nvidia's Data Center segment generated $89.0 billion in revenue in its fiscal second quarter ended July 26, 2026, up 117% from a year earlier and now 92% of total company revenue $89.0 billionNvidia · Q2 FY2027Nvidia reported revenue of $96.2 billion for its fiscal second quarter ended July 26, 2026, up 106% from a year earlier and 18% from the prior quarter $96.2 billionNvidia · Q2 FY2027Nvidia spent $18.5 billion on research and development in its 2026 financial year — three times its $6.04 billion of capital expenditureNvidia FY2026 Form 10-K · FY2026Onto Innovation's inspection systems qualify the 2.5D packaging and high-bandwidth memory built into every AI accelerator, and that demand pushed its Specialty Devices and Advanced Packaging segment to an all-time high. Total company revenue in the quarter ended 30 June 2026 reached $343.1 million in quarterly revenue, up 35.3% year-over-yearOnto Innovation · Q2 2026Across January to May 2026, Epoch AI's capabilities index puts the best open-weight models behind the frontier closed models by an average of four monthsEpoch AI · Jan-May 2026Epoch AI's new AI Chip Users explorer estimates that OpenAI's training-and-inference computing capacity, in H100-equivalents, grew over 2024 and 2025 combined to 17 times OpenAI's 2023 computing capacity, per Epoch AI's H100-equivalent estimateEpoch AI · September 2026Oracle's remaining performance obligations — cloud work it has signed but not yet delivered — stood, when its financial year closed on 31 May 2026, at $638 billion of contracted revenue, up 363%Oracle · Q4 FY2026, ended 31 May 2026Oracle's cash capital expenditures jumped from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026, an increase the company attributes directly to the expansion of its data centres $55.7 billion in Oracle FY2026 capex, up from $21.2 billionOracle · FY2026Oracle's capital expenditures totaled $28.5 billion in fiscal Q1 2027 (three months ended August 31, 2026), more than tripling from $8.5 billion in the year-ago quarter $28.5 billion in Oracle's fiscal Q1 2027 capital expendituresOracle Corporation (SEC 8-K) · Q1 FY2027 (Aug 31, 2026)Oracle's cloud infrastructure business, the unit selling AI training and inference capacity, more than doubled year-over-year in the three months to 31 August 2026, with revenue reaching $7.4 billion in Oracle's AI cloud infrastructure revenue, up 121% year-over-yearOracle Corporation · Q1 FY2027 (ended 31 Aug 2026)Oracle's remaining performance obligations — the cloud contracts it has signed but not yet delivered — jumped by $209 billion over the past year to close its fiscal 2027 first quarter, ended 31 August 2026, at a record $664 billion in Oracle's contracted cloud backlogOracle Corporation · Q1 FY2027 (ended 31 Aug 2026)Stanford's AI Index reports that organisational AI adoption kept climbing in 2025, reaching 88% of surveyed organisationsStanford HAI AI Index · 2025 figurePalantir Technologies grew to 1,049 customers in the twelve months ended June 30, 2026, up from 849 a year earlier — a 23.6% increase — as more organizations adopted its AI-centered software platforms, according to the company's own Q2 2026 Form 10-Q filed with the SEC Palantir's customer count grew to 1,049 in the year to June 2026Palantir Technologies · Q2 2026 10-QAI chip performance per dollar has been improving by approximately 1.37× per yearEpoch AI · 2026PJM's Dominion zone in Virginia serves the largest concentration of data centres in the world, and their power draw shows up hardest on the coldest nights. Winter peak electricity demand there reached 25,413 megawatts in the 2025–26 winter — 45% higher than the 2019–20 winterEIA · May 2026Palantir positions its whole product line as an AI platform but reports no AI revenue at all — its 2025 annual report splits revenue 54% government, 46% commercialPalantir FY2025 Form 10-K · FY2025Scale is not the only driver of AI progress. Epoch AI measures how much training compute it takes to reach a fixed level of language-model performance, and finds that requirement falling fast: pre-training compute efficiency doubles every 7.6 monthsEpoch AI · Feb 2026Quanta Services builds the transmission lines, substations and high-voltage switchgear that connect new AI data centres to the grid, and it told investors that rising demand from data centres, electrification and grid modernisation is driving orders. Its total backlog reached a record $53.4 billion, Quanta Services' record grid-construction backlog as of 30 June 2026Quanta Services · Q2 2026Reddit's Other revenue segment, which includes data-licensing deals with AI companies such as Google and OpenAI, grew 24% year-over-year to $43 million in Q2 2026, per the company's SEC-filed earnings release, even as total revenue rose 61% to $805 million Reddit's AI-licensing 'Other revenue' grew 24% YoY to $43 million in Q2 2026Reddit, Inc. · Q2 2026Rambus makes the memory-interface chips and IP that connect AI accelerators to high-speed DDR5 and HBM memory, and record demand for that silicon drove the company's product revenue for the quarter ended 30 June 2026 up 22% year-over-year and 13% sequentially, to $99.2 million in record product revenueRambus · Q2 2026Samsung is aiming to lift its HBM production capacity from roughly 170,000 wafers a month to around 250,000 wafers a month by the end of 2026, a roughly 47% increase, as it races to supply memory for AI accelerators Samsung's planned HBM capacity by end-2026: ~250,000 wafers/month, +47%TrendForce · Dec 2025Samsung Electro-Mechanics signed its largest single MLCC supply agreement in company history, a one-year contract worth KRW 1.07 trillion for AI server components running through 2027 KRW 1.07 trillion, its largest-ever single MLCC dealTrendForce · September 2026US cloud providers signed multi-year supply agreements that stop memory makers raising their prices, so TrendForce expects the third-quarter 2026 increase to fall on everyone else: server DRAM contract prices up another 13–18% in one quarterTrendForce · 9 July 2026Soaring AI-server memory prices lifted SK hynix to 79.32 trillion won of revenue and 60.54 trillion won of operating profit in the second quarter of 2026 — an operating margin of 76%SK hynix · Q2 2026Sandisk, the NAND flash memory maker spun off from Western Digital, said its Datacenter end-market revenue for the fiscal year ended 3 July 2026 reached $5.153 billion, up 437% from $960 million in fiscal 2025Sandisk Corporation · FY2026Snowflake bundles an AI coding agent, CoCo, directly into its data platform, and tracks how many customer accounts actually use it every week rather than just license it. Averaged over the last four weeks of the quarter ended 31 July 2026, weekly active accounts reached Snowflake's CoCo AI agent: 9,100+ weekly active accounts, up more than 2,000 in the quarter aloneSnowflake · Q2 FY2027Snowflake's push to embed AI across its data platform kept revenue accelerating for a third straight quarter. In the three months ended 31 July 2026 — its fiscal 2027 second quarter — the company reported revenue of $1.55 billion, up 35% year on year, and product revenue of $1.49 billion, up 37%Snowflake · Q2 FY2027Every advanced AI chip is designed with EDA software before a single wafer gets fabricated, and Synopsys, the industry's largest EDA vendor, said AI-driven chip complexity pushed its revenue for the quarter ended 30 April 2026 up 41.9% year-over-year from $1.604 billion, to $2.276 billion in quarterly EDA software revenueSynopsys · Q2 FY2026SpaceX now reports an AI segment alongside Space and Connectivity. In the quarter to 30 June 2026, advertising on X brought in $367 million — less than a year earlier — while the segment's AI solutions and infrastructure line, covering Grok API access, subscriptions, data licensing and cloud, reached $2.19 billion, seven times its level a year earlierSpaceX 10-Q · Q2 2026SpaceX listed on Nasdaq in June 2026, and its first quarterly report as a public company discloses a metric no rocket company used to publish: installed GPUs multiplied by their all-in power draw, before any cooling or facility overhead. At 30 June 2026 that came to 1.4 gigawatts of installed GPU power draw, 3.5× a year earlierSpaceX 10-Q · 30 June 2026Super Micro Computer assembles and ships AI servers rather than designing the chips inside them. For the financial year ended 30 June 2026 it reported net sales of $39.1 billion, up from $22.0 billion a year earlierSuper Micro Computer 8-K · FY ended 30 June 2026AI data centres need firm, round-the-clock power, and nuclear plants are being locked up years in advance to supply it. In June 2025, independent power producer Talen Energy expanded its agreement with Amazon Web Services to deliver up to 1,920 megawatts of carbon-free nuclear power from the Susquehanna plant through 2042Talen Energy · June 2025Teradyne, whose systems test semiconductor chips before they ship, reported quarterly revenue of $1.33 billion for the quarter ended June 28, 2026, up 104% from $652 million a year earlier, with $1.12 billion of that revenue in its Semiconductor Test segment Teradyne's revenue up 104% year-over-year to $1.33 billionTeradyne Inc. · Q2 2026Terminal-Bench 2.0 tests whether AI agents can chain together the kind of multi-step, unsupervised terminal work a developer does in a day — compiling code, training models, setting up servers. Stanford's 2026 AI Index reports agent accuracy on it nearly quadrupled in a year, reaching 77.3% by early 2026, up from 20% in February 2025Stanford AI Index / Terminal-Bench · early 2026Tesla's own quarterly filing attributes its capital spending mainly to global AI and operational infrastructure. Across the first half of 2026 that spending more than doubled against the same six months of 2025, from $3.89 billion to $8.28 billion of capital expenditureTesla Form 10-Q · Q2 2026Tesla reports its AI training compute the way it reports factory capacity — in installed megawatts. At the end of the second quarter of 2026 both Texas clusters, Cortex 1 and Cortex 2, were in production, and the compute installed there had more than doubled in six months, to over 205 MW of AI training capacityTesla Q2 2026 Update · Q2 2026Tesla's driver-assistance software is one of the few AI products consumers pay for directly, and the company discloses the subscriber count every quarter. At the end of the second quarter of 2026 it reported 1.48 million active FSD subscriptions, up 56% in a yearTesla Q2 2026 Update · Q2 2026Compute used to train frontier language models has grown roughly 5× per year since 2020Epoch AI · 2026TrendForce forecasts worldwide shipments of 800G-and-higher optical transceivers will jump 2.6 times, from 24 million units in 2025 to nearly 63 million units in 2026, as AI data centers strain the supply of the EML and CW lasers these modules require shipments jumping to 2.6 times 2025's level, near 63 million units in 2026TrendForce · December 2025 forecastTrendForce raised its 2026 global AI server shipment growth forecast to nearly 31% year-over-year, citing a roughly 90% jump in combined capex from the nine largest cloud service providers Global AI server shipments forecast to grow nearly 31% in 2026TrendForce · 2026 forecast, published Aug 2026TrendForce expects CoWoS-L to remain the mainstream advanced-packaging platform for AI chips through 2028, with Google adopting EMIB-T packaging and both AWS and Microsoft adopting CoWoS-L for their own AI chips in 2027, as Nvidia's high-end GPU shipments grow roughly 30% year-over-year in 2026 Nvidia's high-end GPU shipments growing about 30% year-over-year in 2026TrendForce · September 2026TrendForce projects the market for co-packaged and near-packaged optics — the technology replacing copper links inside AI data centers — will grow from about $100 million in 2025 to over $39 billion by 2030 $39 billion by 2030TrendForce · June 2026TrendForce forecasts global data-center power demand capacity will reach 161 gigawatts in 2026, up about 31% year-over-year, with AI servers driving an estimated 33.4% of that total and their share potentially exceeding 40% in 2027 161 gigawatts of global data-center power demand in 2026TrendForce · September 2026TrendForce forecasts that global foundry industry revenue will grow 24.8% year-over-year to approximately $218.8 billion in 2026, with TSMC posting the largest increase of around 32% Global foundry revenue forecast to grow 24.8% to $218.8 billionTrendForce · published March 2026TrendForce sees a slow shift away from merchant GPUs inside AI servers as Google, Meta and other North American cloud providers scale up in-house chips. It expects GPUs to still lead 2026 AI server shipments at 69.7%, while pegging the ASIC-based share at 27.8% in 2026, the highest since 2023TrendForce · 2026ETrendForce's tally of the industry's biggest cloud spenders keeps climbing. Combining Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu, it now forecasts their combined 2026 capital expenditure will exceed $886.7 billion, with the five North American hyperscalers accounting for nearly 90% of itTrendForce · 2026EAI servers have repriced memory faster than any forecast could keep up with. In a single quarter TrendForce raised its estimate for 2026 global memory market revenue from $551.6 billion to $889.3 billion — 61% above its own forecast from three months earlierTrendForce · May 2026 forecastSLC NAND flash contract prices rose an estimated 130% to 150% cumulatively in the first half of 2026 as supply tightened, according to TrendForce SLC NAND prices up 130-150% in H1 2026TrendForce · H1 2026TSMC's most advanced process nodes — 7-nanometer and finer, the technology AI accelerators and other leading-edge chips require — made up 77% of the foundry's total wafer revenue in the quarter ended June 30, 2026 77% of TSMC's wafer revenue from 7nm-and-below chipsTSMC Q2 2026 earnings release · Q2 2026TrendForce reports TSMC is racing to keep up with AI chip orders from Nvidia, AMD and Broadcom -- monthly 3-nanometer wafer starts, already near 150,000 in the first half of 2026, could reach 180,000 wafers a month by early Q4 2026TrendForce · Aug 2026TSMC's revenue for August 2026 reached NT$514.81 billion, up 53.3% year-over-year and 10.1% from July, taking its January-through-August 2026 total to NT$3,386.87 billion, up 39.3% from the same period in 2025 NT$514.81 billion in August revenue, up 53.3% year-over-yearTSMC · August 2026TSMC's capital expenditure reached about US$15.7 billion in the second quarter of 2026, up 66.9% year-over-year, as the foundry ramps AI chip capacity TSMC spent about US$15.7 billion on capex in Q2 2026 alone, up 67% year-over-yearTSMC (SEC Form 6-K, Ex-99.2) · Q2 2026Every advanced AI chip needs 2.5D packaging before it ships, and TSMC's CoWoS lines have been the industry's tightest bottleneck. TrendForce expects the shortfall between what customers want and what TSMC can supply to ease as capacity ramps toward 200,000 wafers a month industry-wide, narrowing by the end of 2026 from TSMC's CoWoS supply-demand gap narrowing from about 20% to roughly 10% by the end of 2026TrendForce · June 2026 forecastTSMC's high performance computing platform, the category that carries AI accelerators, accounted for 58% of its 2025 revenueTSMC 4Q25 management report · 2025In the first quarter of 2026, TSMC's share of revenue across the world's ten largest foundries reached 72%TrendForce · Q1 2026Stanford's AI Index found the United States trailing most of the world's advanced economies on population-level AI adoption. Citing Microsoft AI Economy Institute telemetry for the second half of 2025, the index ranked the US 24th out of the top 30 economies worldwide, with an AI usage rate of 28.3% of Americans, still in the bottom half of the global rankingStanford AI Index (Microsoft data) · H2 2025Across US businesses surveyed between December 2025 and May 2026, AI use ran between 17% and 20% — and even among firms with 250 or more employees, only 37% reported using AIUS Census Bureau, BTOS · December 2025 – May 2026Counting the notable AI models released in 2025, Stanford's AI Index puts the United States on 59, against China's 35Stanford HAI, AI Index 2026 · 2025US data centres used 176 terawatt-hours of electricity in 2023 — on Berkeley Lab's measurement, 4.4% of the country's totalLawrence Berkeley National Laboratory · 2023Berkeley Lab projects that by 2028 US data centres will draw between 6.7% and 12% of all US electricityLawrence Berkeley National Laboratory · Projection to 2028Generative AI reached 53% adoption in three years, faster than the personal computer or the internet — but unevenly: Singapore sits at 61% and the UAE at 64%, while Stanford's AI Index finds the US is 24th for generative AI adoption, at 28.3%Stanford AI Index · 2026 reportEIA forecasts US electricity use rising 1% in 2026 and 3% in 2027, driven largely by large computing facilities including data centres — which would be the first four-year run of demand growth since 2007EIA Short-Term Energy Outlook · Jan 2026Video generation models can now produce visually convincing footage, but Stanford's 2026 AI Index finds they still struggle to get the physics and consistency right. On VBench-2.0, which scores videos for human-aligned faithfulness to reality rather than visual polish alone, no model evaluated in early 2026 topped a total score of Veo 3 leads the VBench-2.0 video-realism benchmark at 66.7% — no AI video model tested tops 67%Stanford AI Index (VBench-2.0) · early 2026Vertiv, which builds data-centre power and cooling, ended 2025 with an order backlog of $15.0 billion — up 109% in a yearVertiv Q4 2025 results release · Q4 2025Vertiv Holdings, which makes the power and thermal systems AI data centres run on, reported net sales of $3.274 billion for the second quarter of 2026, up 24% year-over-year, citing intensifying demand for AI and general-purpose compute infrastructure Vertiv's Q2 2026 net sales: $3.274 billion, up 24% year-over-yearVertiv Holdings · Q2 2026Vistra, the power generator that has signed long-term nuclear supply deals with Meta and Amazon for their AI data centres, said in its second-quarter 2026 earnings release that it has made an initial commitment of up to $1.0 billion to Helix Digital Infrastructure, a new AI-power venture with Nvidia and KKRVistra Corp. · Q2 2026Vistra signed 20-year power purchase agreements with Meta Platforms in January 2026 to supply a total of 2,609 MW of carbon-free power and capacity from its Perry, Davis-Besse and Beaver Valley nuclear plants in PJM, with full delivery of operating capacity by 2027 and uprate capacity by 2034 2,609 MW of nuclear power under 20-year Vistra-Meta PPAsVistra Corp. · January 2026Stanford's AI Index, citing Brynjolfsson et al., finds AI's labour-market effect landing on hiring pipelines rather than on broad job losses. By September 2025, employment for US software developers aged 22 to 25 had fallen close to 20% from its 2022 peakStanford AI Index · 2026 reportWhile Amazon, Microsoft, Meta, Alphabet and Oracle have all sharply increased AI capital spending, Apple's cash paid for property, plant and equipment in the nine months to 27 June 2026 actually fell, to Apple's nine-month capex of $6.8 billion, down 28% from $9.47 billion a year earlierApple Inc. (SEC 10-Q) · Nine months ended 27 Jun 2026Epoch AI tracks how much AI compute a dollar buys, measured in constant 2025 dollars across the chips companies actually purchased each quarter. Since 2023, AI chip price-performance has grown 49% a year, doubling every 1.7 yearsEpoch AI · August 2026AI data centres worldwide reached a power capacity comparable to New York State at peak demand in 2025, which Stanford's AI Index puts at 29.6 gigawattsStanford HAI, AI Index 2026 · 2025AI hiring intensity doesn't track the size of a country's AI industry. Stanford's 2026 AI Index, drawing on Lightcast's job-postings database across dozens of countries, found the 2025 leader for AI-skill job-posting share by a wide margin was Singapore, at 4.69% of all job postings — nearly double the 2.6% share in the United StatesStanford AI Index · 2025 data (2026 AI Index)AI models can now win gold at the International Mathematical Olympiad, yet Stanford's 2026 AI Index found they still fumble a task most children master early. Across ClockBench's 180 clock designs and 720 questions, the top model, GPT-5.4 High, read analog clocks correctly only 50.6% of the time, versus 90.1% for humansStanford AI Index · 2026 AI Index ReportStanford's 2026 AI Index reports that organizations' use of generative AI specifically, not just AI overall, grew sharply in 2025: the share of organizations that regularly use generative AI in at least one business function reached 79%, up from 71% in 2024Stanford HAI AI Index (McKinsey Survey) · 2025Stanford's 2026 AI Index compared generative AI's adoption curve after its mass-market debut to that of earlier transformative technologies, and found generative AI reached 53% population adoption within three years, faster than the PC or the internetStanford HAI AI Index · 2026OpenAI does not disclose GPT-4o's resource footprint, so Stanford's 2026 AI Index Report turned to independent researchers' estimates of the water needed to cool the servers behind it — as much as 1.58 million kilolitres a year — more than 1.2 million people drink annuallyStanford HAI, AI Index · 2026 AI Index ReportAI training's carbon footprint has scaled by orders of magnitude. Stanford's 2026 AI Index estimates that training AlexNet in 2012 produced about 0.01 tons of CO2 equivalent, while training xAI's Grok 4 in 2025 produced roughly 72,816 tons of CO2 equivalent — over 1,150 times the lifetime emissions of an average carStanford HAI, AI Index · 2026 ReportFrontier AI models are catching up fast to benchmarks meant to resist them. A year ago they scored under 10% on Humanity's Last Exam, a 2,700-question test designed to be hard for AI and favorable to human experts. Now accuracy has reached 38.3% on Humanity's Last Exam, up from under 10% a year earlierStanford AI Index 2026, citing CAIS · 2026 AI Index ReportThe information sector — telecom, media, tech services — led US corporate AI hiring in 2025. Lightcast data cited in Stanford's 2026 AI Index found AI skills required in 13.2% of the sector's job postings, up from 7.8% in 2024Lightcast, via Stanford AI Index · 2025 data, pub. 2026McKinsey's State of AI survey, cited in Stanford's 2026 AI Index, asked organizations already using AI where they had seen the clearest gains. The single most commonly cited outcome in 2025 was improved innovation, named by 64% of respondents, more than any other organizational outcome measuredMcKinsey, via Stanford HAI · 2025AI agents attempting OSWorld's real-world computer tasks across Ubuntu, Windows and macOS — file operations, multi-app workflows — historically topped out at just 1% to 12% success. Stanford's 2026 AI Index reports the best model, Claude Opus 4.5, now reaches 66.3% accuracy, within 6 percentage points of the 72.35% human baseline 66.3% accuracy on OSWorld's real-computer-task benchmarkStanford HAI 2026 AI Index (Epoch AI data) · early 2026Global private investment in AI reached $344.7 billion in 2025, up 127.5% year-over-year, with generative AI companies alone capturing $170.9 billion of that total $344.7 billion in global private AI investment in 2025Stanford HAI AI Index 2026 (citing Quid) · 2025 full yearStanford's 2026 AI Index reports that four out of five US high school and college students now use generative AI for schoolwork, even though only half of middle and high schools have any AI policy and just 6% of teachers describe those policies as clear four out of five US students use AI for schoolworkStanford HAI AI Index 2026, Ch. 7 Education · 2026 reportSWE-bench Verified hands an AI model a real GitHub issue and a codebase and checks whether the patch it writes actually fixes the bug. Stanford's 2026 AI Index reports the leading model, Claude 4.5 Opus in high-reasoning mode, had solved about 76.8% of the benchmark's issues as of February 2026Stanford AI Index · Feb 2026TrendForce says a shortage of critical optoelectronic components — EML and CW lasers — is reshaping the AI optical transceiver supply chain, and projects the global market to grow from $16.5 billion in 2025 to $26 billion in 2026, an AI optical transceiver market up more than 57% year-over-yearTrendForce · April 2026Applied Materials sells the deposition, etch and inspection machines chipmakers must install before a single AI chip exists, so its results run a step ahead of the chip cycle. Its third fiscal quarter, ended 26 July 2026, was the biggest in the company's history: Applied Materials booked a record $9.12bn of revenue, up 25% year over yearApplied Materials 8-K (SEC) · Quarter ended 26 July 2026Applied Materials' Semiconductor Systems segment — the deposition, etch and inspection tools chipmakers install before fabricating a single AI chip — reported fiscal third-quarter 2026 revenue up 26.5% year-over-year to $7.04 billion in Applied Materials' Semiconductor Systems segment revenueApplied Materials, Inc. · Q3 FY2026 (ended 26 Jul 2026)AMD's Data Center segment, the one selling Instinct GPUs and EPYC server chips, took $16.6 billion of the company's record $34.6 billion of revenue in 2025 — 48% of all AMD revenueAMD FY2025 results · year ended 27 Dec 2025Amkor Technology, the outsourced chip packaging and test house scaling up advanced packaging for AI processors, reported record second-quarter 2026 net sales of $1.90 billion, up 26% year-over-year $1.90 billion in record Q2 net sales, +26% YoYAmkor Technology · Q2 2026Amazon's cash capital expenditure across the whole of 2025 came to $131.81 billion in one yearAmazon 10-K, SEC EDGAR · FY2025Amazon's most recently reported quarterly cash capital expenditure reached $54.21 billionAmazon, SEC EDGAR (XBRL filings) · Q2 2026Amazon's cash capital expenditure in the first quarter of 2026 was $44.2 billion in a single quarterAmazon 10-Q, SEC EDGAR · Q1 2026Arista Networks, whose switches carry AI training and inference traffic inside hyperscaler data centres, reported total revenue for the quarter to 30 June 2026 of $3.036 billion, its first quarter above $3 billion and up 37.7% year-over-yearArista Networks · Q2 2026Amphenol's Communications Solutions segment, whose interconnects and cabling carry AI-related IT datacom traffic, reported net sales of $5.38 billion in the quarter ended 30 June 2026, up 85% year-over-year $5.38 billionAmphenol Corporation 10-Q · Q2 2026Applied Digital had signed long-term hyperscaler leases for 1,410 megawatts of AI data-centre capacity across five campuses as of May 31, 2026, worth approximately $36 billion in contracted revenue over the initial 15-year lease terms 1.41 gigawatts of AI data-centre capacity leased for $36 billionApplied Digital Corp. · as of May 31, 2026AppLovin's revenue grew 53% year-over-year to $1.924 billion in the second quarter of 2026, powered by its AI-driven Axon ad-matching engine, with net income up 55% to $1.267 billion AppLovin revenue: $1.92 billion in Q2 2026, up 53% year-over-yearAppLovin Corporation · Q2 2026Arm's total royalty revenue grew 11% year-over-year to $671 million in its fiscal fourth quarter of 2026, and the company said its data centre royalty revenue within that more than doubled year-over-year on rising Arm-based server chip deployments by hyperscalers Arm data-centre royalties more than doubled YoYArm Holdings · Q4 FY2026Arm Holdings' royalty revenue grew 22% year-over-year to $715 million in its fiscal first quarter of 2027, as data-center royalties more than doubled $715 million in Arm royalty revenue, up 22% year-over-yearArm Holdings · Fiscal Q1 2027 (Apr–Jun 2026)Every advanced AI chip needs packaging and testing before it ships, and ASE Technology, the world's largest outsourced semiconductor packaging and test house, is racing to add capacity for it. Equipment spending in its packaging and test segment jumped in the quarter to 30 June 2026, from $942 million a year earlier to $1.646 billion in ASE's advanced-packaging equipment spending, up 74.7% year-over-yearASE Technology Holding · Q2 2026Every advanced AI chip is printed on an ASML machine, and the company says AI demand has its customers accelerating their expansion plans. On 15 July 2026 ASML said it would lift its 2026 low-NA EUV capacity of around 65 systems, for 2027, by 30% more EUV machines a yearASML Q2 2026 results · 15 July 2026ASML's CEO says ongoing AI-related investment is driving demand for advanced logic and memory chips, and it shows up in the numbers -- the lithography maker's second-quarter 2026 total net sales climbed roughly 21% year-over-year, from 7.7 billion euros to 9.3 billion eurosASML Holding · Q2 2026Astera Labs makes the connectivity chips that link AI accelerators inside data-centre racks, and demand pushed its revenue for the quarter ended 30 June 2026 up 104% year-on-year to $392.4 millionAstera Labs 10-Q (SEC) · Q2 2026Astera Labs, whose PCIe and CXL connectivity chips link AI accelerators inside hyperscaler racks, posted record second-quarter 2026 revenue of $392.4 million, up 104% year-over-year, at a GAAP gross margin of 73.3% GAAP gross margin for Astera Labs on record $392.4 million quarterly revenueAstera Labs, Inc. · Q2 2026 (ended 28 Jun 2026)Broadcom reported $20 billion of AI revenue in its 2025 financial year, out of $63.89 billion in total — 31.3% of everything it soldBroadcom FY2025 Form 10-K · FY2025Broadcom's entire 2025 financial year produced $20 billion of AI revenue. In the single quarter to 3 May 2026, its AI semiconductor revenue — custom accelerators and AI networking — reached $10.8 billion in one quarter, up 143% year on yearBroadcom · Q2 FY2026 (to 3 May 2026)Broadcom's custom AI accelerators and networking chips for hyperscale AI clusters just delivered their fastest growth quarter yet — in the fiscal quarter ended 2 August 2026, revenue reached $16.7 billion in AI semiconductor revenue, up 221% year-over-yearBroadcom Inc. · Q3 FY2026Alibaba is racing to keep pace with AI demand on its cloud platform. Capital expenditure in the quarter ended 30 June 2026 reached RMB67.68 billion ($9.98 billion), up 75% year-on-year from RMB38.68 billionAlibaba Group · quarter ended 30 Jun 2026Baidu's Apollo Go robotaxi service has expanded to 28 cities worldwide and its fleet has now driven over 240 million kilometres with nobody behind the wheel, out of more than 350 million autonomous kilometres in total Apollo Go's fully driverless total: 240 million+ kilometres across 28 citiesBaidu Q2 2026 results · Q2 2026Baidu's AI Cloud Infra revenue grew 50% year-over-year in the quarter to 30 June 2026, powered by GPU Cloud demand up 283%, and the company poured capital into keeping pace. Group-wide capital expenditure nearly tripled year-over-year to RMB11.39 billion ($1.68 billion), Baidu's Q2 2026 capital expenditure, up 199.7% year-over-yearBaidu · Q2 2026The Bureau of Labor Statistics built a new four-tier AI exposure classification for every US occupation, combining AI-capability measures with real-world Claude and Copilot usage data mapped onto job tasks. Summing 2025 base-year employment across its 831 rated occupations, BLS's own data shows 53.2 million US workers, 31% of total employment, hold jobs rated 'Very high' AI exposureUS Bureau of Labor Statistics · 2025 employment projectionsBLS's 2024-34 employment projections show AI adoption fueling outsized growth in the occupations that build and secure it: data scientist jobs are projected to grow 33.5%, more than ten times the 3.1% projected growth rate for all US occupations combined 33.5% projected growth in US data scientist jobs through 2034U.S. Bureau of Labor Statistics · 2024-34 projectionsConstellation Energy's newly acquired Calpine subsidiary signed a new agreement in February 2026 to connect and power a CyrusOne data centre next to its Freestone Energy Center in Texas — plus an exclusive option on a second phase of equal size — for an initial 380 megawatts of power committed to a single AI-era data centre in Texas (Calpine-CyrusOne deal)Constellation Energy Corporation · 24 Feb 2026AI benchmarks typically score just one model on a single run, but a new study testing 21 LLMs across 16 benchmarks spanning coding, reasoning, medicine and agentic tasks found that correcting for both single-model and single-run bias reveals real achievable performance is 82% better than standard single-model benchmark scoresarXiv (Fowler et al.) · Jun 2026Cerebras Systems, the wafer-scale AI chip maker whose cloud inference revenue nearly quadrupled year-over-year, disclosed in its second-quarter 2026 earnings release that it now has 600 MW of data center capacity under contractCerebras Systems · Q2 2026Cadence Design Systems, whose software chipmakers use to design AI accelerators, reported $1.584 billion of revenue in its second quarter of 2026, up 24.2% from $1.275 billion a year earlier, with quarter-end backlog reaching a record $8.1 billion as CEO Anirudh Devgan cited accelerating demand for AI-driven design tools Cadence's Q2 2026 revenue: $1.584 billion, up 24.2% year-over-yearCadence Design Systems · Q2 2026Long-dated nuclear supply is being locked up now for demand that only arrives at the end of the decade. In the quarter to 30 June 2026 Constellation Energy signed power purchase agreements running 15 to 20 years with investment-grade customers, beginning between 2029 and 2032: Constellation signed another 920MW of long-term nuclear supply, none of it starting before 2029Constellation Energy 8-K (SEC) · Q2 2026A Census Bureau working paper using its 2026 BTOS AI supplement found that among US firms that adopted AI between November 2025 and January 2026, the scope of use stayed narrow: 57% of AI-adopting firms use it in three or fewer business functionsUS Census Bureau · April 202644% of US firms using AI report it augmenting workers' existing tasks rather than replacing them, versus just 5% reporting AI substituting tasks outright, according to Census Bureau research on business AI adoption 44%U.S. Census Bureau · 2026 BTOS AI supplementAmong U.S. firms where AI is substituting worker tasks, the share reporting a 'large number' of tasks replaced rose from 2.5% in the Dec 2023-Feb 2024 Census BTOS AI supplement to 7% in the Nov 2025-Jan 2026 supplement Share of AI-task-substituting US firms reporting large-scale task replacement rose from 2.5% to 7%US Census Bureau BTOS AI Supplement · April 2026The Census Bureau's Business Trends and Outlook Survey found AI adoption climbing fastest among big employers: as of May 2026, 37% of firms with 250+ staff used AI, 32% at 100-249, under 20% at the smallestUS Census Bureau · May 2026The Census Bureau's Business Trends and Outlook Survey also breaks AI adoption down by industry — in the wave ending 3 May 2026, the Information sector (telecom, media, tech services) led every US industry, with 39.7% of firms using AI, more than double the 19.8% national averageU.S. Census Bureau · May 2026 (BTOS)A Census Bureau working paper covering AI use at over 117,000 US firms found AI-related employment decreases occurred at only 2% of firms, with job cuts linked to firm-wide AI deployment rather than individual worker AI use Only 2% of 117,000+ US firms cut jobs after adopting AI, Census Bureau findsUS Census Bureau (CES-WP-26-25) · Apr 202655% of U.S. workers say they have used AI on the job for at least one of 11 listed tasks, and among those who used it in the past week, 31% said it saved them one to two hours of work, per the Census Bureau's March 2026 Household Trends and Outlook Pulse Survey 55% of U.S. workers have used AI on the job, Census Bureau survey findsU.S. Census Bureau · March 2026 HTOPSA Census Bureau working paper using Quarterly Workforce Indicators administrative payroll data found that early-career employment — ages 22 to 24 — in the most AI-exposed industries fell, over the ten quarters since ChatGPT's release, by a regression-adjusted 12% decline in early-career (22-24) US employment in the most AI-exposed industries since ChatGPT's launchU.S. Census Bureau · April 2026 (CES-WP-26-27)A Census Bureau working paper on the 2026 AI supplement to the Business Trends and Outlook Survey found that, among firms using AI in any business function, Sales and Marketing was the single most common use case, ahead of Strategy and Business Development at 45% and IT at 41% 52% of AI-adopting firms use it in Sales and MarketingU.S. Census Bureau · CES WP, Apr 2026Nearly 750 executives in the Duke/Federal Reserve CFO Survey expect AI to barely touch total headcount — aggregate employment down less than 0.4% in 2026. What shifts is the mix, over the next three years: routine clerical work down more than 2 percentage points of employment shareNBER w34984, The CFO Survey · 2026Q1 survey waveChatGPT was estimated to have reached 100 million monthly users roughly two months after launchUBS, reported by The Register · February 2023The total computing power of the stock of AI chips is growing about 3.4× per yearEpoch AI · 2026Ciena, the pure-play optical networking supplier wiring hyperscale AI data centers together, just posted a record fiscal third quarter — revenue for the quarter ended 1 August 2026 reached $1.67 billion in quarterly revenue, up 37% year-over-yearCiena Corporation · Q3 FY2026CL-Bench, a new benchmark testing whether AI agents actually improve with experience across sequential real-world tasks, found that even the best system -- Claude Sonnet 4.6 using full-context in-context learning -- achieved only a 25.4% normalized gain over its own stateless baseline Best AI agent's continual-learning gain on CL-Bench: 25.4%Asawa et al. (UC Berkeley/Snorkel AI), arXiv · June 2026A UC Berkeley, Sao Paulo and Stanford study of 141 countries found that national Claude usage rises 12x for every 0.10-point increase in a country's AI exposure score, based on Anthropic's Economic Index data for February 2026 Claude usage rises 12x per +0.10 rise in national AI exposureUC Berkeley/Stanford study · arXiv, Aug 2026Cloudflare's capital expenditure fell to 7% of revenue in the quarter to 30 June 2026, down from 12% a year earlier, even as revenue grew 36% year-over-year to $696 million 7% of revenue, down from 12% a year earlierCloudflare Q2 2026 earnings release · Q2 2026Epoch AI asked LLM judges how long an experienced engineer would need, without AI, to reproduce each merged pull request from the 41 collaborators on OpenAI's public Codex repository. In the second quarter of 2026, work beyond what anyone could finish in a single unassisted day accounted for 8% of contributor-days, up from 2% a year earlierEpoch AI · Q2 2026Coherent's Datacenter & Communications segment, the optical transceivers that link GPUs inside AI data centers, brought in $1.615 billion in the quarter to 30 June 2026, up 58.6% from $1.018 billion a year earlier as demand for 800G and 1.6T transceivers accelerated Coherent's AI-datacenter segment revenue: $1.615 billion, up 58.6% year-over-yearCoherent Corp · Q4 FY2026 (June 2026)Coherent's Datacenter & Communications segment, which supplies optical transceivers for AI data centers, generated $1.62 billion in revenue in the quarter ended June 30, 2026, up 58.6% from a year earlier $1.62 billionCoherent Corp. · Q4 FY2026CoreWeave's capital expenditures for Q2 2026 (April-June) totaled approximately $6.42 billion, derived from its SEC 10-Q filings as the $14.12 billion six-month total less the $7.70 billion spent in Q1 2026 CoreWeave's Q2 2026 capital expenditures: approximately $6.42 billionCoreWeave, Inc. (SEC 10-Q) · Q2 2026Core Scientific has contracted about 1.1 gigawatts of leased data-center power capacity to AI and HPC customers, including CoreWeave and AMD, representing more than $24 billion of potential revenue, the company said in its second-quarter 2026 results Core Scientific's leased AI data-center capacity: 1.1 gigawatts, over $24 billion in contractsCore Scientific · Q2 2026Credo Technology, whose active electrical cables and optical DSPs connect AI accelerators inside hyperscaler racks, reported $479.0 million in quarterly revenue, up 114.7% year-over-yearCredo Technology · Q1 FY2027Credo Technology makes the active electrical cables and optical DSPs that connect AI accelerators inside hyperscaler racks, and demand for that interconnect gear pushed its revenue for the quarter ended 26 April 2026 up 157.0% year-over-year to $437.0 million in quarterly revenueCredo Technology · Q4 FY2026Salesforce brands its whole product line around AI agents and reports separately what customers actually pay for them. In the quarter to 30 April 2026, on total revenue of $11.13 billion, that came to $1.2 billion of Agentforce annual recurring revenue, up 205% year on yearSalesforce Q1 FY2027 results · Quarter to 30 April 2026Salesforce brands its whole product line around AI agents and reports separately what customers actually pay for them. In the quarter to 31 July 2026, Agentforce annual recurring revenue exceeded $1.5 billion in Agentforce ARR, up over 240% year-over-yearSalesforce Q2 FY2027 results · Q2 FY2027Salesforce's quarter to 31 July 2026 brought total revenue of $11.3 billion, up 11% year-over-year, with subscription and support revenue of $10.8 billion providing the bulk of the growth $11.3 billion in Q2 FY2027 revenueSalesforce · Q2 FY2027CoreWeave rents out AI computing capacity, and building that capacity is capital-intensive. Its first-quarter 2026 cash flow statement shows capital expenditure on property and equipment, including capitalised internal-use software, jumped to $7.7 billion in Q1 2026 capital spending, a 5.5x jump from $1.4 billion a year earlierCoreWeave · Q1 2026Datadog's own 10-Q filing shows AI is now a material swing factor in its growth: for the quarter ended 30 June 2026, its AI-native customer cohort — which includes its single largest customer — contributed high-single-digit percentage points of Datadog's total year-over-year revenue growthDatadog Q2 2026 10-Q · Q2 2026Dell's AI server business booked a record $60.9 billion in orders and exited its second fiscal quarter of 2027 with a record $95 billion backlog, on revenue of $16.4 billion $95 billion backlogDell Technologies · Q2 FY2027 (ended Aug 1, 2026)Dell's quarter to 28 May 2026 was the one where AI servers took over the business: total revenue hit a record $43.8 billion, up 88% year on year, and of that, $16.1 billion came from AI servers aloneDell Technologies 8-K · Q1 FY2027, ended 28 May 2026Data-center landlords are racing to keep pace with AI and cloud demand. Digital Realty's second-quarter 2026 financial supplement to investors disclosed a global construction pipeline of Digital Realty's 1,402 megawatts of IT capacity under construction, 54% already pre-leasedDigital Realty Q2 2026 supplement · Q2 2026Digital Realty ended Q2 2026 with a record $1.9 billion backlog of signed-but-not-commenced data center leases, driven by surging AI and hyperscale demand $1.9 billion data center lease backlogDigital Realty Trust · Q2 2026 (reported July 23, 2026)As cloud providers shifted from AI training to inference, conventional DRAM contract prices jumped 93% to 98% in a single quarter — and on TrendForce's count of the first quarter of 2026, global DRAM revenue hit $97 billion, up 81%TrendForce · Q1 2026Epoch AI's Capabilities Index scores frontier models on one aggregate scale, and fits a break in the trend at April 2024, when reasoning models and reinforcement learning took over frontier training. Before the break the frontier gained 8.3 index points a year; after it, 15.5 points a year, 1.85 times as fastEpoch AI · December 2025NBER asked 69 economists what the US economy would look like in 2050 if AI capabilities advance rapidly. Their median answer was 3.5% annual GDP growth, alongside a labour force participation rate falling from today's 62% to 55.0% by 2050, the level before women entered the workforce en masseNBER working paper 35046 · April 2026AI data centres are pushing US electricity generation to record highs. Total output from the country's power sector in the first half of 2026 rose 1.8% versus the first half of 2025, adding 37 billion kilowatt-hoursEIA, Short-Term Energy Outlook · Aug 2026 STEOEIA modelled what would happen if data-centre power demand grew faster than forecast. Its February 2026 baseline had US coal generation falling 9.3% between 2025 and 2027; in the higher-demand scenario, coal plants run on their existing spare capacity and the decline shrinks to a 5.0% fall in US coal generation instead of 9.3%EIA Short-Term Energy Outlook analysis · March 2026EIA's September 2026 Short-Term Energy Outlook forecasts commercial-sector electricity sales climbing 3.3% in 2026, accounting for 63% of that year's total US electricity-sales growth, with data-centre development cited as a key driver 63% of all US electricity-sales growth in 2026EIA Short-Term Energy Outlook · September 2026EIA's Annual Energy Outlook 2026 projects that in its High Electricity Demand case, US data-centre servers alone — before any cooling or other building overhead — could draw as much as 818 billion kWh of server electricity a year by 2050US Energy Information Administration · AEO2026Data-centre development is now the main driver of new US electricity demand, and the EIA's September 2026 Short-Term Energy Outlook forecasts US electricity generation climbing 2.2% in 2026 to a record, before rising a further 1.7% in 2027, to 4,368 billion kilowatt-hours of US electricity generation in 2026U.S. EIA · Sep 2026 STEOEIA's Annual Energy Outlook 2026 estimates that data-centre servers alone, before any cooling or other building load, already accounted for 7% of all US commercial-sector electricity in 2025US Energy Information Administration · AEO2026, May 2026EIA's September 2026 Short-Term Energy Outlook, which attributes record US electricity demand to data-center development, forecasts the average residential electricity price rising from 17.30 cents per kilowatt-hour in 2025 to 18.20 cents in 2026 and 18.59 cents in 2027 18.20 cents per kilowatt-hourEIA · September 2026 STEOEIA's September 2026 Short-Term Energy Outlook projects the West South Central region (Texas, Oklahoma, Louisiana, Arkansas) will drive nearly 20% of total U.S. electricity sales growth in 2026 and nearly 40% in 2027, propelled largely by data center development despite Texas's grid-connection pause for new data center projects West South Central region accounts for nearly 40% of 2027 U.S. electricity sales growth (EIA)EIA Short-Term Energy Outlook · September 2026On 3 August 2026 the governor of Texas paused new data-centre development while projects already under review are assessed. In the same month's Short-Term Energy Outlook the EIA cut its forecast for Texas electricity load growth in 2027 to 6% growth in Texas load, down from 14% forecast a month earlierEIA Short-Term Energy Outlook · August 2026Texas paused new data-centre development on 3 August 2026 while it reviews the projects already in the queue, and the EIA cut its forecast for the state in response: Texas electricity load growing 6% in 2027, not the 14% forecast a month earlierEIA Short-Term Energy Outlook · August 2026Data centers are reshaping state-level power demand — the EIA reports commercial electricity sales in Virginia, home to the world's largest concentration of data centers in PJM's Dominion zone, rose by nearly 30.0 million megawatthours between 2019 and 2025, the fastest growth of any US state except TexasU.S. Energy Information Administration · 2019-2025NBER's Firm Data on AI paper found a striking gap in expectations: while executives surveyed expect AI to cut their firms' employment by 0.7% over the next three years, US employees surveyed separately in December 2025 predicted AI would raise US firms' employment about 0.5% over the next three yearsNBER (Yotzov et al.) · March 2026AI inference is consuming NAND flash faster than suppliers can allocate it. Enterprise SSD contract prices rose roughly 80% in the first quarter of 2026, and on TrendForce's count the top five brands' combined revenue hit a record $18.46 billion — enterprise SSD revenue rose 86.1% in a single quarterTrendForce · Q1 2026Epoch AI modeled the total cost of ownership of a typical 1-gigawatt AI data center built by a US hyperscaler around Nvidia's GB200 NVL72 systems, and found the up-front capital expenditure alone comes to $38 billion for one hyperscale AI data centerEpoch AI · May 2026Epoch AI's AI Data Centers tracker now covers an estimated 46% of the world's deployed AI computing capacity, up from 26% as of March 2026, spanning 87 sites with a combined 13.2 gigawatts of IT power and about 13.7 million H100-equivalent GPUs, as of 21 September 2026 46% of the world's deployed AI computing capacityEpoch AI · 21 September 2026Investment in AI-related data center construction, compute hardware and networking equipment reached about 0.8% of US GDP in the first quarter of 2026, pushing total computing-infrastructure investment to roughly 1.5% of GDP — more than double the 2015-2022 average of about 0.7%, per Epoch AI AI infrastructure investment: ~0.8% of US GDP in Q1 2026Epoch AI · Q1 2026The 84 AI data centers tracked by Epoch AI have a combined real-world power footprint of 16.9 gigawatts as of August 2026, more than New York City's 11 GW peak demand 16.9 GWEpoch AI · August 2026Epoch AI tested three commercial AI detectors — Pangram, GPTZero, Originality.ai — pinned to their June 2026 versions. Against plain machine-generated text they catch nearly everything, but shown five samples of a specific author's writing and told to imitate the style, AI detectors missed 13% of style-imitated AI passages, versus under 1% of plain AI textEpoch AI · Jul 2026Epoch AI rebuilt the cost structure of three AI companies — Anthropic, MiniMax and Z.ai — from IPO filings and reporting. Staff, despite the industry's headline pay packages, take under a quarter of spending. Research and inference compute together take 54% to 62% of an AI company's costsEpoch AI · 2025 accountsEpoch AI finds a $2,500 consumer gaming GPU (RTX 5090) can now run open models matching frontier AI performance from 6 to 12 months earlier, with the lag ranging from 6.3 months on the Artificial Analysis Intelligence Index to 12.4 months on LM Arena Elo 6 to 12 monthsEpoch AI · 2026Epoch AI fits a trend through the models whose context windows ranked among the ten longest on their release date, and finds that since mid-2023 the frontier context window has been getting about 30× longer every yearEpoch AI · since mid-2023Frontier models have started finding software flaws at scale. Counting the high- and critical-severity CVEs disclosed by 21 major vendors — Microsoft, Google, Apple and Cisco among them — Epoch AI puts July 2026 at about 2,500 serious CVEs in July 2026, five times the monthly record set before April 2026Epoch AI · July 2026Epoch AI's tracked data-center dataset captures only 26% of global AI computing capacity as of March 2026, covering 5.2 million H100-equivalent chips versus an estimated 20.2 million H100-equivalent chips sold through 2025 Epoch AI's data-center dataset covers just 26% of global AI computing capacity as of March 2026Epoch AI · March 2026Epoch AI tracks the largest AI data centre in the world by IT power capacity, not raw compute -- and on that measure the record has roughly doubled every 10 months since mid-2024, climbing from Google's 237-megawatt Papillion facility to xAI's Colossus 2, now at 946 megawatts of IT power capacityEpoch AI · Sept 2026Epoch AI had frontier models repeatedly play the board game Earthborne Rangers to test whether they improve with practice, and found GPT-5.5 and Claude Opus 4.8 both mismanaged its core fatigue mechanic at 2.1 fatigue per round, versus 0.6 for expert humans and 3.5 for purely random playEpoch AI · July 2026Epoch AI finds the frontier of its Capabilities Index has advanced 14 points a year for reasoning models since September 2024, more than double the 6 points a year non-reasoning models achieved from GPT-4 to GPT-4.5 14 points a year, vs 6 for non-reasoning modelsEpoch AI · September 2026Epoch AI tracked 333 compute estimates for notable AI models released between 2010 and May 2024, and found that training compute for the frontier — the top 10 models by compute at release — has been growing each year to reach 5.3x the prior year's levelEpoch AI · 2010–May 2024 dataOpenAI, Anthropic and xAI are the three labs most associated with sparking the AI boom, yet Epoch AI estimates the trio together controlled well under 4 million H100-equivalent GPUs at the end of 2025 — only 20% to 30% of the world's total operational AI computing powerEpoch AI · End of 2025Epoch AI's FrontierMath Erdős benchmark tasks AI models with proving or disproving 68 significant unsolved Erdős problems in Lean, each within a fixed $300, 72-hour budget. Running five frontier models in August 2026, only one produced any verified solution: GPT-6 Astra solved 2 of 68 problems (3%) — every other model tested scored 0%Epoch AI · Aug 2026Epoch AI audited its FrontierMath benchmark after OpenAI flagged suspiciously high scores, and its 12 June 2026 v2 release corrected errors in 42% of the original problems, leaving a verified set of 338 problems across four tiers 42% of FrontierMath's original problems contained errorsEpoch AI · June 2026 (v2 release)Epoch AI modeled the power budget of a frontier AI data centre built around Nvidia's GB200 architecture, and found the GPUs themselves -- the part everyone pictures when they think 'AI compute' -- account for only about 40% of total facility power at peakEpoch AI · Dec 2025Epoch AI estimates Huawei will produce less than 4% as much AI compute, measured in H100-equivalents, as Nvidia in 2026 -- and without access to foreign high-bandwidth memory, that share could fall to around 1% by 2028 under 4% of Nvidia's AI compute in 2026Epoch AI · Sep 2026AI capability is getting cheaper to reach, not just bigger. Epoch AI tracked what it cost in tokens to hit the same score — about 27% accuracy on its FrontierMath benchmark — first with o4-mini in April 2025, then with GPT-5.2 in December 2025, and found the same capability got ~3x cheaper in eight months, a ~5-10x-a-year declineEpoch AI · Feb 2026Epoch AI and Ipsos asked 469 US workers who had used AI for work in the previous seven days whose subscription they were on. In computer, engineering and science jobs 67% use a tool their employer provides — everywhere else the free tier wins: 46% of US workers using AI at work are on a free plan, against 37% given a tool by their employerEpoch AI / Ipsos · Fieldwork 10-19 July 2026Epoch AI and Ipsos surveyed 1,106 employed US adults between 10 and 19 July 2026, on the Ipsos KnowledgePanel, and found 20% say AI now handles a task they used to hand off to a human coworker or contractorEpoch AI / Ipsos · Jul 2026Epoch AI compared how OpenAI's and Anthropic's flagship models handle very long prompts, and found GPT-5.6 Terra and Sol slow down quadratically as context grows while Claude Sonnet 5 and Opus 5 stay almost linear -- the marginal latency to process the next 10,000 tokens for GPT-5.6 Terra rose from 0.071 seconds at a 100,000-token prompt to 1.66 seconds at a 10-million-token promptEpoch AI · Sept 2026Epoch AI's new AI Chip Users explorer estimates Meta Superintelligence Labs held about 996,000 H100-equivalent GPUs of AI computing capacity as of December 2025, putting it behind Google DeepMind and Anthropic among the frontier labs it tracks Meta's ~996,000 H100-equivalent GPUsEpoch AI · December 2025Epoch AI and METR's new MirrorCode benchmark found the top AI model correctly rebuilt real software from behavior alone in 56% of cases, up from about 30% for leading models roughly eight months earlier Top AI model now solves 56% of Epoch AI's toughest long-horizon coding tests, up from about 30% eight months agoEpoch AI / METR (MirrorCode benchmark) · Jul 2026The share of US adults using AI at least six days a week more than doubled in six months, rising from 8.2% in March 2026 to 18.7% in August 2026 Near-daily US AI use rose from 8.2% to 18.7% in six monthsEpoch AI / Ipsos polling · Mar–Aug 2026Equinix, the world's largest data-center operator, ties its buildout of hyperscale capacity directly to demand it links explicitly to the adoption of AI. Its capital expenditure in the quarter ended 30 June 2026 reached Equinix's Q2 2026 capital expenditures: $1.578 billion, up 59.6% year-over-yearEquinix · Q2 2026Eaton's Electrical Global segment, which sells the power distribution and switching gear AI data centres depend on, ended the second quarter of 2026 with total backlog up 103% year-over-year Eaton Electrical Global backlog up 103% year-over-yearEaton · Q2 2026Asked what three years of AI had actually changed inside their own firms, nearly 6,000 senior executives across the US, UK, Germany and Australia mostly reported nothing: nine in ten saw no impact on employment or productivityNBER working paper 34836 · February 2026Fabrinet, the Thai contract manufacturer that assembles the optical transceivers linking Nvidia's and other hyperscalers' AI clusters, just posted a record quarter — revenue for the fiscal fourth quarter, ended 26 June 2026, hit $1.32 billion in quarterly revenue, up 45% year-over-yearFabrinet · Q4 FY2026Of the notable AI models released in 2025, Stanford's AI Index counts over 90% as coming from industryStanford HAI, AI Index 2026 · 2025Across 42 notable language models, Epoch AI estimates frontier training costs have risen from roughly $2 million for GPT-3 to up to nearly $390 million per runEpoch AI · 2024Epoch AI's Open Problems benchmark collects research mathematics problems that have resisted serious attempts by professional mathematicians. Expanded to 50 problems on 31 July 2026, the count two weeks later — with the newest AI solution still marked provisional — stood at 4 solved by AI, 1 by a human, 45 still openEpoch AI FrontierMath · 17 August 2026Alphabet told investors in its second-quarter 2026 results that Gemini models now process 22 billion API tokens a minute, and that the Gemini app has reached 950 million monthly active usersAlphabet 8-K · Q2 2026Brynjolfsson et al. asked US adults what they'd need to be paid to give up generative AI for a month, and Stanford's 2026 AI Index reports their estimate that total US consumer surplus grew from $112 billion in 2025 to $172 billion in annual consumer surplus by early 2026Stanford AI Index (Brynjolfsson et al.) · early 2026In a randomised field experiment across 66 firms and 7,137 knowledge workers, a generative AI tool built into everyday work apps saved the people who used it two hours a week on email, with no detected change in their tasksNBER working paper 33795 · revised Nov 2025Mentions of generative AI skills in US AI job postings grew 111% from 2024 to 2025, even as their share of total AI postings slipped, as employer demand pivots toward agentic-AI skills, per Lightcast data cited in Stanford's AI Index 111%Stanford HAI · 2025Across 5,179 customer-support agents, a generative AI assistant raised issues resolved per hour by 14% on average — and by 34% for the least experienced staffBrynjolfsson, Li & Raymond, NBER · April 2023GE Vernova builds the gas turbines and grid equipment that electricity supply runs on, and AI data centres are now one of its largest sources of new demand. Its gas equipment under contract rose from 100 GW to 116 GW during the second quarter of 2026, and data-centre orders through the first half of the year reached over $5 billion of data-centre orders, more than double the whole of 2025GE Vernova Q2 2026 results · First half of 2026GE Vernova's Electrification segment, which sells the transformers and switchgear data centres need to connect to the grid, booked over $5 billion in data-centre equipment orders in the first half of 2026 alone -- more than its entire 2025 total GE Vernova's H1 2026 data-centre orders: over $5 billion, beating all of 2025GE Vernova · Q2 2026A large-language-model analysis of 18,797 work tasks across 124 economies found automation-exposed task shares ranging from 3.3% in the least-exposed economy to 61.6% in the most, rising with income but varying widely within income groups 3.3% to 61.6% across 124 economiesGlobal Automation Atlas (Garg, Crosta & Baier) · 2026Alphabet's most recently reported quarterly cash spending on property and equipment reached $44.92 billionAlphabet, SEC EDGAR (XBRL filings) · Q2 2026Alphabet's cash capital expenditure in the first quarter of 2026 reached $35.67 billion in a single quarterAlphabet 10-Q, SEC EDGAR · Q1 2026Alphabet's quarterly cash capital expenditure has roughly tripled in two years, from $12.01 billion in the first quarter of 2024 to $35.67 billion in the first quarter of 2026Alphabet 10-Q, SEC EDGAR · Q1 2024 - Q1 2026Epoch AI estimates the single largest owner of AI computing power worldwide is not a chipmaker or a cloud reseller but Google, which held, measured in H100-equivalents as of the fourth quarter of 2025, about a quarter of all global AI compute, mostly its own TPUs rather than Nvidia GPUsEpoch AI · Q4 2025Epoch AI's Capabilities Index scores language models on one aggregate scale, and nothing since has held first place as long as GPT-4 did after its March 2023 release — roughly a year. The second-longest lead, by OpenAI's o1, lasted a little over three months, less than a third of GPT-4's run at the topEpoch AI Capabilities Index · July 2026OpenAI's GPT-6 Astra scores 97.6% on Epoch AI's FrontierMath Tier 4 (v2) math benchmark, versus 87.8% for Anthropic's Claude Fable 5.1 GPT-6 Astra: 97.6% on FrontierMath Tier 4 (v2), vs Claude Fable 5.1's 87.8%Epoch AI · September 2026Traditional computing powers like Germany, Japan and France now play marginal roles in AI infrastructure. Epoch AI's tracking of global GPU cluster performance in May 2025 put China second at 15%, behind a United States hosting about three-quarters of the world's GPU cluster performanceEpoch AI · May 2025US private AI investment reached $285.9 billion in 2025 — 23.1 times China's $12.4 billion and 48.5 times the UK's $5.9 billion — per Stanford's 2026 AI Index 23.1 timesStanford AI Index (Quid data) · 2025High-bandwidth memory is taking over the memory industry's capacity: TrendForce expects HBM to account, among the top three DRAM suppliers by the end of 2026, for 22% of DRAM wafer input, up from 18% a year earlierTrendForce · June 2026 projectionAmazon, Google, Meta, Microsoft and Oracle together held, measured in H100-equivalents as of Q4 2025, an estimated 71% of the world's AI computeEpoch AI · Q4 2025A January 2026 IMF staff discussion note tracking the spread of AI-related job skills across US local labor markets found that in occupations highly exposed to AI but with little scope to complement it, five years after AI-skill demand appeared, employment was 3.6% lower than in other regionsIMF · Jan 2026About 20% of Finland's workforce holds jobs with high AI exposure but low complementarity — including software developers, finance professionals and administrative staff — putting them at greater risk of job displacement, per IMF analysis of 2023 EU labor data 20%IMF · 2026 Selected Issues PaperAn IMF staff analysis of Israel's labour market found that 23.1% of national employment sits in occupations with high AI exposure but low task complementarity -- the group most vulnerable to AI-driven displacement rather than augmentation 23.1% of Israeli jobs: high AI exposure, low complementarityIMF Selected Issues Paper · July 2026Epoch AI models the world's AI inference capacity, at fixed model size and context length, as growing 3.4× a year — but it puts token demand growth at around 10× a year, three times faster than supplyEpoch AI · May 2026Innodata's data engineers annotate, label and curate the training data behind frontier AI models, and demand for that work keeps climbing. Its second-quarter 2026 revenue hit a record $92.1 million, up 58% year-over-yearInnodata Inc. · Q2 2026Intel's Data Center and AI segment, the unit selling Xeon server chips and Gaudi/Crescent Island AI accelerators, reported revenue for the quarter ended 27 June 2026 up 59% year-over-year to $6.3 billion in quarterly segment revenueIntel Corporation · Q2 2026IREN, the bitcoin miner turned AI cloud operator, said its AI Cloud Services revenue grew roughly 8x year-over-year to $128.8 million in the fiscal year ended June 30, 2026, and the company now has $4 billion of ARR contracted for 2026 GPU capacity, with $1 billion already operatingIREN FY26 results release · FY2026IREN, a bitcoin miner turned AI cloud operator, told investors in its quarter ended 31 March 2026 that it now holds executed grid connection agreements across seven data-centre sites in Texas, Oklahoma and British Columbia for a combined 4,510 megawatts of AI data-centre power capacityIREN Ltd · Q3 FY2026 10-QIron Mountain is building AI-ready colocation data centres alongside its legacy records-storage business, but its own quarterly filing shows growth capital expenditure on that data-centre build-out fell in the six months to 30 June 2026, from $953.0 million a year earlier, to $880.1 million in data-centre growth capex, down 8%Iron Mountain (SEC 10-Q) · H1 2026KLA Corporation, whose process-control and inspection tools qualify AI chip production before wafers ship, posted fiscal fourth-quarter 2026 results of $3.66 billion in revenue, up 15% year-over-yearKLA Corporation · Q4 FY2026Epoch AI tracks the largest AI data centre in the world by computing capacity, and the record has changed hands repeatedly since August 2024 — Colossus, Anthropic and Amazon's New Carlisle, Microsoft's Fairwater Atlanta, Meta's Prometheus. Each time, the bar has risen at the same pace: a doubling every seven monthsEpoch AI · June 2026The largest AI data centre in Epoch AI's tracker, Colossus 2 in Memphis, Tennessee, holds a computing capacity of 1,112,000 H100-equivalent AI chipsEpoch AI · AI data centers tracker, August 2026Nvidia's accelerators have pushed thermal design power from 700W for the H100 to over 1,000W for the B200 and B300, and air can no longer carry the heat away. TrendForce expects liquid cooling in server racks to reach 47% of server racks by 2026TrendForce · November 2025 forecast for 2026Lumentum makes the optical components that link AI accelerators inside data centres, and reported fiscal fourth-quarter 2026 results of $1.01 billion in net revenue, up 109% year-over-yearLumentum Holdings · Q4 FY2026Lam Research, whose machines etch and deposit the layers inside AI chips, reported fiscal 2026 revenue of $23.23 billion — up 26% from $18.44 billion in fiscal 2025 — as CEO Tim Archer credited AI-driven demand for record margins and EPS $23.23 billion in fiscal 2026 revenueLam Research · FY2026 (ended June 28, 2026)A McKinsey survey of employers cited in Stanford's 2026 AI Index found anticipated AI-driven workforce cuts concentrated in service operations, supply chain and software engineering, while almost half of organizations expect little or no change overall. Still, one-third of employers expect AI to shrink their workforce in the year aheadMcKinsey survey (Stanford AI Index 2026) · 2025 surveyMeta's purchases of property and equipment nearly doubled in 2025, from $37.3bn to $69.7bn — 34.7% of its revenueMeta FY2025 Form 10-K · FY2025Meta's cash capital expenditure went the other way in the first quarter of 2026 — from $21.38 billion the quarter before, a fall to $19.0 billionMeta 10-Q, SEC EDGAR · Q1 2026Meta's most recently reported quarterly purchases of property and equipment reached $30.12 billionMeta, SEC EDGAR (XBRL filings) · Q2 2026In a randomised trial, 16 experienced open-source developers believed AI tools had made them 20% faster. Measured against issues they finished without AI, they were 19% slowerMETR · July 2025Micron breaks out the memory it sells into AI data centres as its own business unit. In the quarter to 28 May 2026 that unit took in $11.52 billion, against $1.53 billion in the same quarter a year before — Micron's core data centre revenue grew 7.5× in a yearMicron 8-K, SEC · quarter to 28 May 2026Micron plans to add up to 60,000 HBM wafers per month of capacity, reaching roughly 100,000 wafers/month by end-2026, per TrendForce -- still far below Samsung's and SK hynix's respective ~150,000-200,000 wafers/month Micron's HBM wafer capacity is set to reach ~100,000 wafers/month by end-2026, per TrendForceTrendForce · September 2026MIT researchers tracking corporate AI adoption found that in 2025, 11% of S&P 500 firms had AI deeply integrated into core business processes and a further 10% were using it in the production of goods or services -- both figures have surged since 2022, when deep integration stood at just 5% of S&P 500 firmsMIT researchers (Yu et al.) · 2025Modine Manufacturing, the thermal-management maker whose liquid-cooling systems keep hyperscale AI server racks from overheating, saw hyperscaler demand keep surging — first-quarter fiscal 2027 Data Centers segment sales reached $348.6 million in Data Centers segment sales, up 90% year-over-yearModine Manufacturing · Q1 FY2027Monolithic Power Systems makes the chips that step grid power down to the voltages AI GPUs need. Its Enterprise Data segment, which sells into AI servers, grew far faster than the rest of the company in the quarter to 30 June 2026, when total revenue hit a record $980.6 million and Enterprise Data segment revenue reached $380.6 million in AI server chip revenue, up 164.3% year-over-yearMonolithic Power Systems · Q2 2026Marvell designs the custom accelerators, optics and switching silicon that wire AI clouds together, and that business has crowded out everything else it sells. In the quarter to 2 May 2026, data centre revenue of $1.83 billion out of $2.42 billion total came to 76% of Marvell's revenueMarvell Technology · Q1 FY2027 (to 2 May 2026)Marvell's record quarter to 2 May 2026 looks very different depending on which profit line you read. Non-GAAP net income was $718.0 million. Under GAAP — after $207.6 million of stock compensation, $225.2 million of amortisation on acquired intangibles and a $250.7 million contingent-consideration revaluation — net income was $34.5 million, under 5% of the adjusted figureMarvell Technology · Q1 FY2027 (to 2 May 2026)Marvell's revenue hit a record $2.74 billion in the quarter to 1 August 2026, up 37% year-on-year, led by an AI data-centre business that grew even faster — accelerating to 46% year-on-year, to $2.17 billion in Data Center revenueMarvell Technology, Inc. · quarter ended 1 August 2026Marvell's data infrastructure chip business set a fiscal second-quarter revenue record in the three months to 1 August 2026, with net revenue up 37% year-over-year to $2.739 billion, as Data Center segment revenue growth accelerated to 46% $2.739 billion in record Q2 revenueMarvell Technology · Q2 FY2027Microsoft told investors in April 2026 that its AI business had reached an annual run rate of $37 billionMicrosoft FY2026 Q3 earnings call · April 2026Microsoft's cash capital expenditure rose by more than half between the third and fourth quarters of 2025, reaching $29.88 billion in a single quarterMicrosoft 10-Q, via SEC XBRL · Q4 2025Microsoft's most recently reported quarterly cash spending on property and equipment reached $35.80 billionMicrosoft, SEC EDGAR (XBRL filings) · Q2 2026Microsoft's cash spending on property and equipment reached, in the first quarter of 2026 alone, $30.88 billionMicrosoft 10-Q, via SEC XBRL · Q1 2026A randomised online experiment gave 1,174 adults aged 25 to 45 an incentivised business problem-solving task, some with a generative-AI assistant and some without. Unaided, the higher-education group beat the lower-education group by 0.548 standard deviations. With AI the gap fell to 0.139 — about three-quarters of the education gap closedNBER Working Paper 34851 · May 2026 revisionMenaka Hampole and co-authors linked resume data to firms' AI-application adoption to isolate AI's true effect on jobs, not just correlation. In their preferred instrumental-variables specification, a one-standard-deviation rise in AI exposure cuts within-firm employment share by 14%NBER · Working Paper 33509, 2025A National Bureau of Economic Research working paper, drawing on the Federal Reserve Banks of Atlanta and Richmond's CFO Survey of roughly 750 corporate executives, found that by the end of 2025 more than 40% of US companies still had not invested in AI at all, as of year-end 2025NBER (Atlanta/Richmond Fed CFO Survey) · March 2026 (NBER WP 34984)A Danish register study spanning 25,000 workers across 7,000 workplaces in 11 AI-exposed occupations found widespread chatbot adoption and new AI-related tasks, but a difference-in-differences design ruled out earnings or hours effects larger than 2%, at either the worker or workplace level, two years after ChatGPT's launch no earnings or hours effect above 2%, two years after ChatGPT's launchNBER Working Paper 33777 · revised Mar 2026A survey of nearly 6,000 senior business executives across the US, UK, Germany and Australia — run by the Bank of England, the Federal Reserve Bank of Atlanta, the Bundesbank and Macquarie University — found that on average 69% of businesses already use AINBER · Working Paper 34836, Feb 2026Firms across the US, UK, Germany and Australia expect AI to lower their employment by about 0.7% over the next three years, with the steepest expected declines in the UK (-1.4%) and US (-1.2%) Firms expect AI to cut employment ~0.7% over the next 3 yearsNBER Working Paper 34836 · Feb 2026, rev. Mar 2026Workplace generative AI use pays off in wages, not just output. Analyzing 2024 Current Population Survey data and controlling for demographics, occupation and industry, NBER researchers Alexander Bick, Adam Blandin and David Deming found that, compared with non-users, daily generative-AI users earn 40% more, and occasional users 22% moreNBER (Bick, Blandin & Deming) · Sep 2024, rev. Feb 2025NBER's 'Mind the Gap' working paper finds 43% of US workers used AI for their job in early 2026, versus 32% of workers surveyed across Germany, UK, France, Italy, Sweden and the Netherlands 43% of US workers used AI for work in 2026, versus 32% across EuropeNBER Working Paper 34995 · Jan-Feb 2026Nebius Group, the AI-infrastructure and GPU-cloud operator building out data centres for AI training and inference, spent $5.7 billion on capital expenditure in the quarter to 30 June 2026, driven mainly by GPU purchases and data-centre expansion. Across the first half of 2026 it spent a total of $8.1 billion on AI infrastructure capital expenditureNebius Group · H1 2026Nebius Group, the AI-infrastructure and GPU-cloud operator, keeps raising its own build-out target for AI data-centre power. In its second-quarter 2026 shareholder letter, up from a forecast of more than 4 gigawatts the quarter before, the company said it now expects to end 2026 with 5 gigawatts of contracted AI data-centre powerNebius Group · Q2 2026 shareholder letterNebius rents out AI cloud capacity and is scaling from a small base. Revenue in the second quarter of 2026 reached $582.3 million against $105.1 million in the same quarter of 2025 — a 454% year-on-year rise in revenueNebius Group · Q2 2026NRG Energy agreed on principal commercial terms with a global cloud and AI hyperscaler to build a 1.2 GW combined-cycle natural gas power plant in Texas to supply a new data centre 1.2 GW gas power plant for a hyperscaler's Texas data centreNRG Energy · Q2 2026In July 2025, Nvidia became the first US company in history to reach a $4 trillion market capitalisationAl Jazeera · July 2025Nvidia's Data Center segment brought in $193.7bn of the $215.9bn it made in its 2026 financial year — 89.7% of the company's revenueNvidia FY2026 Form 10-K · FY2026, ended 25 January 2026Nvidia's Data Center segment generated $89.0 billion in revenue in its fiscal second quarter ended July 26, 2026, up 117% from a year earlier and now 92% of total company revenue $89.0 billionNvidia · Q2 FY2027Nvidia reported revenue of $96.2 billion for its fiscal second quarter ended July 26, 2026, up 106% from a year earlier and 18% from the prior quarter $96.2 billionNvidia · Q2 FY2027Nvidia spent $18.5 billion on research and development in its 2026 financial year — three times its $6.04 billion of capital expenditureNvidia FY2026 Form 10-K · FY2026Onto Innovation's inspection systems qualify the 2.5D packaging and high-bandwidth memory built into every AI accelerator, and that demand pushed its Specialty Devices and Advanced Packaging segment to an all-time high. Total company revenue in the quarter ended 30 June 2026 reached $343.1 million in quarterly revenue, up 35.3% year-over-yearOnto Innovation · Q2 2026Across January to May 2026, Epoch AI's capabilities index puts the best open-weight models behind the frontier closed models by an average of four monthsEpoch AI · Jan-May 2026Epoch AI's new AI Chip Users explorer estimates that OpenAI's training-and-inference computing capacity, in H100-equivalents, grew over 2024 and 2025 combined to 17 times OpenAI's 2023 computing capacity, per Epoch AI's H100-equivalent estimateEpoch AI · September 2026Oracle's remaining performance obligations — cloud work it has signed but not yet delivered — stood, when its financial year closed on 31 May 2026, at $638 billion of contracted revenue, up 363%Oracle · Q4 FY2026, ended 31 May 2026Oracle's cash capital expenditures jumped from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026, an increase the company attributes directly to the expansion of its data centres $55.7 billion in Oracle FY2026 capex, up from $21.2 billionOracle · FY2026Oracle's capital expenditures totaled $28.5 billion in fiscal Q1 2027 (three months ended August 31, 2026), more than tripling from $8.5 billion in the year-ago quarter $28.5 billion in Oracle's fiscal Q1 2027 capital expendituresOracle Corporation (SEC 8-K) · Q1 FY2027 (Aug 31, 2026)Oracle's cloud infrastructure business, the unit selling AI training and inference capacity, more than doubled year-over-year in the three months to 31 August 2026, with revenue reaching $7.4 billion in Oracle's AI cloud infrastructure revenue, up 121% year-over-yearOracle Corporation · Q1 FY2027 (ended 31 Aug 2026)Oracle's remaining performance obligations — the cloud contracts it has signed but not yet delivered — jumped by $209 billion over the past year to close its fiscal 2027 first quarter, ended 31 August 2026, at a record $664 billion in Oracle's contracted cloud backlogOracle Corporation · Q1 FY2027 (ended 31 Aug 2026)Stanford's AI Index reports that organisational AI adoption kept climbing in 2025, reaching 88% of surveyed organisationsStanford HAI AI Index · 2025 figurePalantir Technologies grew to 1,049 customers in the twelve months ended June 30, 2026, up from 849 a year earlier — a 23.6% increase — as more organizations adopted its AI-centered software platforms, according to the company's own Q2 2026 Form 10-Q filed with the SEC Palantir's customer count grew to 1,049 in the year to June 2026Palantir Technologies · Q2 2026 10-QAI chip performance per dollar has been improving by approximately 1.37× per yearEpoch AI · 2026PJM's Dominion zone in Virginia serves the largest concentration of data centres in the world, and their power draw shows up hardest on the coldest nights. Winter peak electricity demand there reached 25,413 megawatts in the 2025–26 winter — 45% higher than the 2019–20 winterEIA · May 2026Palantir positions its whole product line as an AI platform but reports no AI revenue at all — its 2025 annual report splits revenue 54% government, 46% commercialPalantir FY2025 Form 10-K · FY2025Scale is not the only driver of AI progress. Epoch AI measures how much training compute it takes to reach a fixed level of language-model performance, and finds that requirement falling fast: pre-training compute efficiency doubles every 7.6 monthsEpoch AI · Feb 2026Quanta Services builds the transmission lines, substations and high-voltage switchgear that connect new AI data centres to the grid, and it told investors that rising demand from data centres, electrification and grid modernisation is driving orders. Its total backlog reached a record $53.4 billion, Quanta Services' record grid-construction backlog as of 30 June 2026Quanta Services · Q2 2026Reddit's Other revenue segment, which includes data-licensing deals with AI companies such as Google and OpenAI, grew 24% year-over-year to $43 million in Q2 2026, per the company's SEC-filed earnings release, even as total revenue rose 61% to $805 million Reddit's AI-licensing 'Other revenue' grew 24% YoY to $43 million in Q2 2026Reddit, Inc. · Q2 2026Rambus makes the memory-interface chips and IP that connect AI accelerators to high-speed DDR5 and HBM memory, and record demand for that silicon drove the company's product revenue for the quarter ended 30 June 2026 up 22% year-over-year and 13% sequentially, to $99.2 million in record product revenueRambus · Q2 2026Samsung is aiming to lift its HBM production capacity from roughly 170,000 wafers a month to around 250,000 wafers a month by the end of 2026, a roughly 47% increase, as it races to supply memory for AI accelerators Samsung's planned HBM capacity by end-2026: ~250,000 wafers/month, +47%TrendForce · Dec 2025Samsung Electro-Mechanics signed its largest single MLCC supply agreement in company history, a one-year contract worth KRW 1.07 trillion for AI server components running through 2027 KRW 1.07 trillion, its largest-ever single MLCC dealTrendForce · September 2026US cloud providers signed multi-year supply agreements that stop memory makers raising their prices, so TrendForce expects the third-quarter 2026 increase to fall on everyone else: server DRAM contract prices up another 13–18% in one quarterTrendForce · 9 July 2026Soaring AI-server memory prices lifted SK hynix to 79.32 trillion won of revenue and 60.54 trillion won of operating profit in the second quarter of 2026 — an operating margin of 76%SK hynix · Q2 2026Sandisk, the NAND flash memory maker spun off from Western Digital, said its Datacenter end-market revenue for the fiscal year ended 3 July 2026 reached $5.153 billion, up 437% from $960 million in fiscal 2025Sandisk Corporation · FY2026Snowflake bundles an AI coding agent, CoCo, directly into its data platform, and tracks how many customer accounts actually use it every week rather than just license it. Averaged over the last four weeks of the quarter ended 31 July 2026, weekly active accounts reached Snowflake's CoCo AI agent: 9,100+ weekly active accounts, up more than 2,000 in the quarter aloneSnowflake · Q2 FY2027Snowflake's push to embed AI across its data platform kept revenue accelerating for a third straight quarter. In the three months ended 31 July 2026 — its fiscal 2027 second quarter — the company reported revenue of $1.55 billion, up 35% year on year, and product revenue of $1.49 billion, up 37%Snowflake · Q2 FY2027Every advanced AI chip is designed with EDA software before a single wafer gets fabricated, and Synopsys, the industry's largest EDA vendor, said AI-driven chip complexity pushed its revenue for the quarter ended 30 April 2026 up 41.9% year-over-year from $1.604 billion, to $2.276 billion in quarterly EDA software revenueSynopsys · Q2 FY2026SpaceX now reports an AI segment alongside Space and Connectivity. In the quarter to 30 June 2026, advertising on X brought in $367 million — less than a year earlier — while the segment's AI solutions and infrastructure line, covering Grok API access, subscriptions, data licensing and cloud, reached $2.19 billion, seven times its level a year earlierSpaceX 10-Q · Q2 2026SpaceX listed on Nasdaq in June 2026, and its first quarterly report as a public company discloses a metric no rocket company used to publish: installed GPUs multiplied by their all-in power draw, before any cooling or facility overhead. At 30 June 2026 that came to 1.4 gigawatts of installed GPU power draw, 3.5× a year earlierSpaceX 10-Q · 30 June 2026Super Micro Computer assembles and ships AI servers rather than designing the chips inside them. For the financial year ended 30 June 2026 it reported net sales of $39.1 billion, up from $22.0 billion a year earlierSuper Micro Computer 8-K · FY ended 30 June 2026AI data centres need firm, round-the-clock power, and nuclear plants are being locked up years in advance to supply it. In June 2025, independent power producer Talen Energy expanded its agreement with Amazon Web Services to deliver up to 1,920 megawatts of carbon-free nuclear power from the Susquehanna plant through 2042Talen Energy · June 2025Teradyne, whose systems test semiconductor chips before they ship, reported quarterly revenue of $1.33 billion for the quarter ended June 28, 2026, up 104% from $652 million a year earlier, with $1.12 billion of that revenue in its Semiconductor Test segment Teradyne's revenue up 104% year-over-year to $1.33 billionTeradyne Inc. · Q2 2026Terminal-Bench 2.0 tests whether AI agents can chain together the kind of multi-step, unsupervised terminal work a developer does in a day — compiling code, training models, setting up servers. Stanford's 2026 AI Index reports agent accuracy on it nearly quadrupled in a year, reaching 77.3% by early 2026, up from 20% in February 2025Stanford AI Index / Terminal-Bench · early 2026Tesla's own quarterly filing attributes its capital spending mainly to global AI and operational infrastructure. Across the first half of 2026 that spending more than doubled against the same six months of 2025, from $3.89 billion to $8.28 billion of capital expenditureTesla Form 10-Q · Q2 2026Tesla reports its AI training compute the way it reports factory capacity — in installed megawatts. At the end of the second quarter of 2026 both Texas clusters, Cortex 1 and Cortex 2, were in production, and the compute installed there had more than doubled in six months, to over 205 MW of AI training capacityTesla Q2 2026 Update · Q2 2026Tesla's driver-assistance software is one of the few AI products consumers pay for directly, and the company discloses the subscriber count every quarter. At the end of the second quarter of 2026 it reported 1.48 million active FSD subscriptions, up 56% in a yearTesla Q2 2026 Update · Q2 2026Compute used to train frontier language models has grown roughly 5× per year since 2020Epoch AI · 2026TrendForce forecasts worldwide shipments of 800G-and-higher optical transceivers will jump 2.6 times, from 24 million units in 2025 to nearly 63 million units in 2026, as AI data centers strain the supply of the EML and CW lasers these modules require shipments jumping to 2.6 times 2025's level, near 63 million units in 2026TrendForce · December 2025 forecastTrendForce raised its 2026 global AI server shipment growth forecast to nearly 31% year-over-year, citing a roughly 90% jump in combined capex from the nine largest cloud service providers Global AI server shipments forecast to grow nearly 31% in 2026TrendForce · 2026 forecast, published Aug 2026TrendForce expects CoWoS-L to remain the mainstream advanced-packaging platform for AI chips through 2028, with Google adopting EMIB-T packaging and both AWS and Microsoft adopting CoWoS-L for their own AI chips in 2027, as Nvidia's high-end GPU shipments grow roughly 30% year-over-year in 2026 Nvidia's high-end GPU shipments growing about 30% year-over-year in 2026TrendForce · September 2026TrendForce projects the market for co-packaged and near-packaged optics — the technology replacing copper links inside AI data centers — will grow from about $100 million in 2025 to over $39 billion by 2030 $39 billion by 2030TrendForce · June 2026TrendForce forecasts global data-center power demand capacity will reach 161 gigawatts in 2026, up about 31% year-over-year, with AI servers driving an estimated 33.4% of that total and their share potentially exceeding 40% in 2027 161 gigawatts of global data-center power demand in 2026TrendForce · September 2026TrendForce forecasts that global foundry industry revenue will grow 24.8% year-over-year to approximately $218.8 billion in 2026, with TSMC posting the largest increase of around 32% Global foundry revenue forecast to grow 24.8% to $218.8 billionTrendForce · published March 2026TrendForce sees a slow shift away from merchant GPUs inside AI servers as Google, Meta and other North American cloud providers scale up in-house chips. It expects GPUs to still lead 2026 AI server shipments at 69.7%, while pegging the ASIC-based share at 27.8% in 2026, the highest since 2023TrendForce · 2026ETrendForce's tally of the industry's biggest cloud spenders keeps climbing. Combining Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu, it now forecasts their combined 2026 capital expenditure will exceed $886.7 billion, with the five North American hyperscalers accounting for nearly 90% of itTrendForce · 2026EAI servers have repriced memory faster than any forecast could keep up with. In a single quarter TrendForce raised its estimate for 2026 global memory market revenue from $551.6 billion to $889.3 billion — 61% above its own forecast from three months earlierTrendForce · May 2026 forecastSLC NAND flash contract prices rose an estimated 130% to 150% cumulatively in the first half of 2026 as supply tightened, according to TrendForce SLC NAND prices up 130-150% in H1 2026TrendForce · H1 2026TSMC's most advanced process nodes — 7-nanometer and finer, the technology AI accelerators and other leading-edge chips require — made up 77% of the foundry's total wafer revenue in the quarter ended June 30, 2026 77% of TSMC's wafer revenue from 7nm-and-below chipsTSMC Q2 2026 earnings release · Q2 2026TrendForce reports TSMC is racing to keep up with AI chip orders from Nvidia, AMD and Broadcom -- monthly 3-nanometer wafer starts, already near 150,000 in the first half of 2026, could reach 180,000 wafers a month by early Q4 2026TrendForce · Aug 2026TSMC's revenue for August 2026 reached NT$514.81 billion, up 53.3% year-over-year and 10.1% from July, taking its January-through-August 2026 total to NT$3,386.87 billion, up 39.3% from the same period in 2025 NT$514.81 billion in August revenue, up 53.3% year-over-yearTSMC · August 2026TSMC's capital expenditure reached about US$15.7 billion in the second quarter of 2026, up 66.9% year-over-year, as the foundry ramps AI chip capacity TSMC spent about US$15.7 billion on capex in Q2 2026 alone, up 67% year-over-yearTSMC (SEC Form 6-K, Ex-99.2) · Q2 2026Every advanced AI chip needs 2.5D packaging before it ships, and TSMC's CoWoS lines have been the industry's tightest bottleneck. TrendForce expects the shortfall between what customers want and what TSMC can supply to ease as capacity ramps toward 200,000 wafers a month industry-wide, narrowing by the end of 2026 from TSMC's CoWoS supply-demand gap narrowing from about 20% to roughly 10% by the end of 2026TrendForce · June 2026 forecastTSMC's high performance computing platform, the category that carries AI accelerators, accounted for 58% of its 2025 revenueTSMC 4Q25 management report · 2025In the first quarter of 2026, TSMC's share of revenue across the world's ten largest foundries reached 72%TrendForce · Q1 2026Stanford's AI Index found the United States trailing most of the world's advanced economies on population-level AI adoption. Citing Microsoft AI Economy Institute telemetry for the second half of 2025, the index ranked the US 24th out of the top 30 economies worldwide, with an AI usage rate of 28.3% of Americans, still in the bottom half of the global rankingStanford AI Index (Microsoft data) · H2 2025Across US businesses surveyed between December 2025 and May 2026, AI use ran between 17% and 20% — and even among firms with 250 or more employees, only 37% reported using AIUS Census Bureau, BTOS · December 2025 – May 2026Counting the notable AI models released in 2025, Stanford's AI Index puts the United States on 59, against China's 35Stanford HAI, AI Index 2026 · 2025US data centres used 176 terawatt-hours of electricity in 2023 — on Berkeley Lab's measurement, 4.4% of the country's totalLawrence Berkeley National Laboratory · 2023Berkeley Lab projects that by 2028 US data centres will draw between 6.7% and 12% of all US electricityLawrence Berkeley National Laboratory · Projection to 2028Generative AI reached 53% adoption in three years, faster than the personal computer or the internet — but unevenly: Singapore sits at 61% and the UAE at 64%, while Stanford's AI Index finds the US is 24th for generative AI adoption, at 28.3%Stanford AI Index · 2026 reportEIA forecasts US electricity use rising 1% in 2026 and 3% in 2027, driven largely by large computing facilities including data centres — which would be the first four-year run of demand growth since 2007EIA Short-Term Energy Outlook · Jan 2026Video generation models can now produce visually convincing footage, but Stanford's 2026 AI Index finds they still struggle to get the physics and consistency right. On VBench-2.0, which scores videos for human-aligned faithfulness to reality rather than visual polish alone, no model evaluated in early 2026 topped a total score of Veo 3 leads the VBench-2.0 video-realism benchmark at 66.7% — no AI video model tested tops 67%Stanford AI Index (VBench-2.0) · early 2026Vertiv, which builds data-centre power and cooling, ended 2025 with an order backlog of $15.0 billion — up 109% in a yearVertiv Q4 2025 results release · Q4 2025Vertiv Holdings, which makes the power and thermal systems AI data centres run on, reported net sales of $3.274 billion for the second quarter of 2026, up 24% year-over-year, citing intensifying demand for AI and general-purpose compute infrastructure Vertiv's Q2 2026 net sales: $3.274 billion, up 24% year-over-yearVertiv Holdings · Q2 2026Vistra, the power generator that has signed long-term nuclear supply deals with Meta and Amazon for their AI data centres, said in its second-quarter 2026 earnings release that it has made an initial commitment of up to $1.0 billion to Helix Digital Infrastructure, a new AI-power venture with Nvidia and KKRVistra Corp. · Q2 2026Vistra signed 20-year power purchase agreements with Meta Platforms in January 2026 to supply a total of 2,609 MW of carbon-free power and capacity from its Perry, Davis-Besse and Beaver Valley nuclear plants in PJM, with full delivery of operating capacity by 2027 and uprate capacity by 2034 2,609 MW of nuclear power under 20-year Vistra-Meta PPAsVistra Corp. · January 2026Stanford's AI Index, citing Brynjolfsson et al., finds AI's labour-market effect landing on hiring pipelines rather than on broad job losses. By September 2025, employment for US software developers aged 22 to 25 had fallen close to 20% from its 2022 peakStanford AI Index · 2026 report
LIVE FACTS
262Facts published, each traced to a filing
72AI companies tracked
12AI models ranked
4Languages
6Sends a week
Sourced fromSEC EDGARStanford AI IndexEpoch AITrendForceU.S. Census Bureau
Menu

FACT PROVENANCE

Every fact has a receipt

Every number we publish carries its source and its date — visible, on the card, every time.

Today's facts · 24 September 2026

  • Primarystraight from the source
  • Reportedvia a named outlet
  • Estimatea named analyst's projection
Adoption6/day by emailReported

Stanford's 2026 AI Index reports that four out of five US high school and college students now use generative AI for schoolwork, even though only half of middle and high schools have any AI policy and just 6% of teachers describe those policies as clear four out of five US students use AI for schoolwork

SourceStanford HAI AI Index 2026, Ch. 7 Education·2026 report
Open this fact →
Compute6/day by emailPrimary

Epoch AI's AI Data Centers tracker now covers an estimated 46% of the world's deployed AI computing capacity, up from 26% as of March 2026, spanning 87 sites with a combined 13.2 gigawatts of IT power and about 13.7 million H100-equivalent GPUs, as of 21 September 2026 46% of the world's deployed AI computing capacity

SourceEpoch AI·21 September 2026
Open this fact →

Browse every fact — the full archive, filterable by topic and source tier.

FOUNDING MEMBERS · THE DAILY RECEIPTS & THE WEEKLY BRIEF

Free, six days a week

The Daily Receipts, every weekday: all six sourced AI facts for the day — the free site only shows four — the market open, and one part of the site worth your time. The Weekly Brief every Saturday. Free, and every figure traces to its filing.

Sign up now and you're a founding member — you'll hear Insider and Pro pricing before anyone else.

Edition of 24 September 2026

Stanford's 2026 AI Index reports that four out of five US high school and college students now use generative AI for schoolwork, even though only half of middle and high schools have any AI policy and just 6% of teachers describe those policies as clear four out of five US students use AI for schoolwork

Stanford's 2026 AI Index reports that four out of five US high school and college students now use generative AI for schoolwork, even though only half of middle and high schools have any AI policy and just 6% of teachers describe those policies as clear four out of five US students use AI for schoolwork

Read this morning's edition →

Free forever. Six emails a week — The Daily Receipts on weekdays, The Weekly Brief on Saturday — drop either in one click.

THE INTELLIGENCE CLUB

Everything you need to navigate the AI decade

Inspired by the best of investing research and AI intelligence platforms — rebuilt around one obsession: facts you can act on.

Daily AI Facts

The signal, minus the noise. Every fact sourced and dated — receipts on the card, every time.

262 facts on file

Go to today's facts

Model & Chip Leaderboards

Who's really winning? Live rankings of the models, the chips, and the companies behind them.

Top: Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)

Go to the leaderboard

Investing Intelligence

The AI share of revenue — published only where the company itself discloses one. No estimates.

Go to AI exposure

AI Stock Screener

Every AI-exposed listing in one table, each price stamped with the minute it was captured.

72 listings tracked

Go to the screener

Weekly Video Briefings

The AllFactsAI channel — the week in AI, decoded in minutes, every week.

Go to the channel

The Members' Club

A community of investors comparing notes: early access, member calls and AMAs.

See the three tiers

Not sure where to start? Start free with The Daily Receipts →

Menu

THE AI VALUE CHAIN

Follow the money through the boom

Every layer of the AI build-out, from the power grid to the software that has to pay for it all.

  1. Compute is now bounded by electricity and heat, not by demand for chips. Grid interconnects, turbines, transformers and liquid cooling decide how fast the boom can physically go.

    What to watch
    Interconnection queues, power purchase deals, cooling attach rates.

    Price this layer →
    VRTVertivETNEatonGEVGE VernovaCEGConstellation Energy

  2. The machines that make the machines. Lithography and deposition tools gate every new fab, with lead times measured in years.

    What to watch
    Equipment bookings, export-control changes, fab build announcements.

    Price this layer →
    ASMLASMLAMATApplied MaterialsLRCXLam ResearchKLACKLA

  3. Advanced logic and high-bandwidth memory are the scarcest inputs in the chain. Advanced packaging capacity has repeatedly been the binding constraint on accelerator supply.

    What to watch
    Advanced packaging capacity, HBM sold-out announcements, node transitions.

    Price this layer →
    TSMTSMCSKHYSK HynixMUMicron005930.KSSamsung

  4. GPUs and custom silicon convert capital into compute. This layer captures the largest share of AI spending — and sets the price of intelligence.

    What to watch
    Data-center revenue, backlog, custom-silicon share shift, gross margin.

    Price this layer →
    NVDANvidiaAMDAMDAVGOBroadcomMRVLMarvell

  5. Training runs span tens of thousands of accelerators, so the fabric between them is as decisive as the silicon itself. Optics scale with cluster size, not with chip count.

    What to watch
    Switch port shipments, optical transceiver upgrade cycles.

    Price this layer →
    ANETArista NetworksCOHRCoherentCRDOCredoCIENCiena

  6. Where the compute physically lives. Pre-leasing has moved years ahead of construction, turning shells and land banks into strategic assets.

    What to watch
    Pre-leasing rates, megawatts under construction, vacancy.

    Price this layer →
    EQIXEquinixDLRDigital RealtyIRMIron MountainCRWVCoreWeave

  7. Four balance sheets fund most of the boom. Their capex guidance is the single most important number in the entire chain — it is everyone else's revenue.

    What to watch
    Capex guidance revisions, cloud backlog, depreciation schedules.

    Price this layer →
    MSFTMicrosoftGOOGLAlphabetAMZNAmazonMETAMeta

  8. Where capital becomes capability. Mostly private, so public-market exposure runs through their partners and investors — which is exactly where mispricing hides.

    What to watch
    Benchmark leadership, cost per token, funding rounds, partner economics.

    Price this layer →
    private · MSFTOpenAIprivate · GOOGL/AMZNAnthropicGOOGLGoogle DeepMindSPCXxAI

  9. The layer that has to justify everything below it. If AI revenue does not show up here, the capex cycle above eventually resets.

    What to watch
    Seat expansion, AI attach rates, pricing power, net revenue retention.

    Price this layer →
    PLTRPalantirNOWServiceNowCRMSalesforceTSLATesla
Menu

AI EXPOSURE

DISCLOSED FIGURES ONLY · 21 SEPTEMBER 2026

How much AI is really in there?

We publish the AI share of revenue only where the company itself reports one — and say so plainly where it does not.

AI share of revenue

7

No AI revenue figure disclosed

20
89.7%OF REVENUE

Nvidia

Accelerators

AI share of revenueCompany disclosed

Data Center segment revenue of $193.74bn on $215.94bn total. FY2026, ended 25 January 2026.

Data Center is not a pure AI line — it includes networking and non-AI accelerated computing — but it is the closest figure Nvidia reports.

Nvidia FY2026 Form 10-K ↗
From the filingsDerived from filings
Revenue
$215.94B
R&D / revenue
8.6%
Capex / revenue
2.8%

Fiscal year ended 2026-01-25. Capex intensity measures spending on AI, not selling into it — it runs high for hyperscalers and low for the companies supplying them.

Where we exercise judgementOur read
Chain criticalityVery high
Moat durabilityHigh

The purest expression of the trade, and the company with the least anywhere to hide if demand turns.

Only 7 of these 27 companies tell you what share of their revenue is AI. That silence is itself the finding: the further you get from the silicon, the less anyone is willing to put a number on it.

AI revenue shares are the companies' own disclosures, linked on each card. Revenue, R&D and capex come from SEC XBRL filing data. Criticality and moat are our editorial judgement and are labelled as such — they are not folded into any number. Nothing here is a rating or investment advice; see the methodology.

The whole card, company by company

Every figure above, plus the cash flows, margins and valuation behind it. Nvidia's card is open to everyone; the rest are Insider.

Nvidia

AI COMPANY FACT CARD

Nvidia

NVDA · Accelerators · FY2026, ended 25 January 2026

Sample card
71.1%
Gross margin
60.4%
Operating margin
44.8%
Cash flow margin
89.7%
AI share of revenue

Revenue & free cash flow

Billions of dollars per fiscal year, with revenue growth under each. Levels, not growth rates: one year off a small base makes a growth chart unreadable.

27.0
3.8
60.9
27.0
130.5
60.9
215.9
96.7
FY2023
baseline
FY2024
+126%
FY2025
+114%
FY2026
+65%
RevenueFree cash flow

From the filings

From exposure

Nvidia FY2026 Form 10-K, via SEC XBRL

Revenue
$215.94B
Free cash flow
$96.68B
Cash & equivalents
$10.61B
Shares outstanding
24.3bn
Open the filing ↗

Revenue by market

The company's own market split, latest fiscal year against the one before. Bars are scaled to the largest line, not to total revenue.

Data Center$193.7B+68%
Gaming$16.0B+41%
Professional Visualization$3.2B+70%
Automotive$2.4B+39%
OEM and Other$0.6B+59%

Revenue by region

United States69.3%
Taiwan19.6%
China (incl. Hong Kong)9.1%
Other2.0%

Revenue by customer billing location. Nvidia discloses that 76% of the revenue billed to Taiwan is attributable to end customers based in the United States and Europe — billing location is not the end customer's location.

Nvidia FY2026 Form 10-K, revenue by region ↗

Quality

Our read

Our judgement, kept out of every measured figure.

Strengths — more is better

Chain criticalityVery high
Moat durabilityHigh
Disclosure qualityHigh

Risks — more is worse

Customer concentrationVery high

Two direct customers each accounted for 22% and 14% of Compute & Networking revenue in FY2026 — up from 12% and 11% a year earlier. Nvidia names neither. That segment is most of the company, so this is close to company-wide concentration, though it is not reported as such.

The read

Our read

The purest expression of the trade, and the company with the least anywhere to hide if demand turns.

  • Reports a segment figure at all — most of the chain does not.
  • Free cash flow grew with revenue rather than lagging it.
  • Capex intensity of 2.8% while selling to buyers spending ten times that.
  • Data Center is not a pure AI line; it is an upper bound.
  • Almost the whole business now rides on one end market.
  • Cash and equivalents are small next to the market's expectations.

What the price assumes

Our read

The growth today's quote already implies, on our other assumptions.

No live quote available, so this is left blank rather than computed from a stale price.

Terminal multiple
25×
Discount rate
12%

Our valuation

Our read

Low, base and high cases. A range, never a target.

$130 – $208

Base case $165

$165
$130
$208

Free cash flow grew 59% in FY2026. All three cases assume that decelerates sharply, because five more years at anything close to the recent rate is not a forecast anyone can defend. Net cash is cash and equivalents less total debt; marketable securities are excluded rather than estimated, which makes the band conservative.

Sources on every figureallfacts.ai · Not financial advice
Open the Nvidia card ↗

26 more companies have a card of their own. Every one is Insider-only.

Unlock with InsiderSee every company ↗

Company of the week

Free to read this week

Meta

META · Hyperscalers

It has told the market it will spend up to $145 billion this year to support its AI efforts, and has never disclosed a single dollar of AI revenue.

Revenue grew 28%
$60.80bnFiled
Operating income fell 8%
$18.78bnFiled
Capital spending, one quarter
$31.08bnFiled
Read the deep dive →All deep dives ↗

AllFactsAI holds no position in Meta (META), as at 4 September 2026.

Menu

THE MODEL LEADERBOARD

SNAPSHOT · 23 SEPTEMBER 2026

Smartest per dollar — and who owns it

Intelligence against blended price, every lab tagged with the listed company that actually gives you exposure to it.

Artificial Analysis Intelligence Index vs blended $ / 1M tokens

Up and to the left is better

45505560$0.5$1$2$5$10$20Blended price per 1M tokens (log scale) — cheaper is leftIntelligence indexClaude Opus 5.5Muse Spark 1.3MiMo-V2.6-Pro
Closed / API onlyOpen weightsWeights not classifiedValue frontier
Artificial Analysis Intelligence Index and blended price per 1M tokens, 23 September 2026
ModelLabIntel.$ / 1MListed proxy
Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)frontierAnthropic57.6$8.00GOOGL · AMZN
Intelligence57.6$ / 1M tokens$8.00closedHow you get exposurePrivate. Backed by Google and Amazon; runs on TPU and Trainium.GOOGLAMZN

Most of the frontier is still private, and the listed exposure sits in the compute, the clouds and the backers underneath it. The exceptions are the labs owned outright by a listed company — Google, Meta, and xAI, inside SpaceX since their 2026 merger.

Scores and prices: Artificial Analysis, read 23 September 2026 — artificialanalysis.ai. One source only; other leaderboards score differently and are not comparable. Proxy notes describe exposure, not ownership, and are not investment advice.

Menu

THE AI STOCK SCREENER

MARKET OPEN · 14:57 UTCPRICES 15-MIN DELAYED · EODHD₩ RESTATED AT 1,365.78/$ · 23 Sept

Every layer of the chain — priced

72 listed companies, each placed in the layer of the AI value chain where its AI revenue is actually recognised. Sort by any column; a figure we do not have prints as a dash, never as a zero.

S&P 500

 

7,686.91

-0.25%

NASDAQ

 

26,792.98

-0.53%

DOW

 

51,305.12

-0.40%

RUSSELL 2000

IWM ETF

$280.61

-0.46%

VIX

 

15.71

+3.49%

Compare stocks

#30-Day
1
Nvidia
Accelerators$222.49-1.34%+5.43%+8.29%+13.45%$5.36T$18.34M28.1$164.79 – $235.20
2
Apple
Hyperscalers$336.61-0.12%+1.39%+8.60%+15.09%$4.91T$6.10M38.6$224.07 – $339.79
3
Alphabet
Hyperscalers$338.76+0.28%-1.47%-2.88%-2.10%$4.14T*$6.44M17.0$198.63 – $402.12
4
Microsoft
Hyperscalers$494.54-1.21%+2.10%+2.73%+37.23%$3.67T$4.25M27.5$352.17 – $537.65
5
Amazon
Hyperscalers$246.50-1.11%+1.35%-4.88%+6.40%$2.66T$8.16M19.8$198.79 – $284.02
6
TSMC
Foundry & Memory$443.64-0.66%+6.91%+9.18%+1.58%$2.30T$1.58M—$224.65 – $476.28
7
Meta Platforms
Hyperscalers$765.44+2.87%+10.60%+33.21%+33.54%$1.96T*$13.50M28.8$524.82 – $777.09
8
Broadcom
Accelerators$348.36-1.87%+4.75%-0.87%-6.92%$1.66T$5.17M44.4$286.95 – $479.90
9
Tesla
Applications$377.73-0.63%+6.16%+8.93%+1.22%$1.49T$7.44M349.8$298.32 – $489.88
10
Micron Technology
Foundry & Memory$1,057.22-1.37%+15.69%+17.73%+2.24%$1.19T$6.60M23.9$115.60 – $1,213.37
11
AMD
Accelerators$616.43+0.30%+19.92%+34.56%+18.25%$1.01T$8.58M158.5$151.14 – $623.77
12
SpaceX
Model Labs$147.86-0.34%-1.67%+9.90%-4.00%$721.38B*$28.05M—$108.27 – $211.39
13
ASML Holding
Semi Equipment$1,719.89-1.42%+8.89%+0.26%-0.90%$662.88B$324,785—$721.10 – $1,986.87
14
Intel
Foundry & Memory$123.76+0.94%+21.33%+40.50%-6.87%$624.23B$41.04M—$23.50 – $140.94
15
Palantir Technologies
Applications$192.37+0.30%+10.01%+9.04%+68.98%$462.25B*$7.54M165.8$107.27 – $207.18
16
Oracle
Hyperscalers$137.44-4.93%+0.98%+1.48%-7.91%$415.58B$28.90M21.5$114.99 – $324.63
17
Lam Research
Semi Equipment$302.45-1.57%+14.25%-0.83%-17.93%$378.46B$1.49M52.5$96.39 – $432.87
18
Applied Materials
Semi Equipment$464.33-2.12%+14.20%-2.03%-19.37%$368.49B$858,02240.1$155.39 – $722.23
19
Dell Technologies
Data Centers$535.63-2.58%-2.39%+26.93%+26.88%$342.80B*$1.78M31.1$110.54 – $588.40
20
Arm Holdings
Accelerators$319.96-3.79%+36.31%+39.27%-7.39%$340.45B$2.25M326.5$104.55 – $439.46
21
Sandisk
Foundry & Memory$1,787.23-1.62%+19.51%+21.66%-5.11%$261.68B$2.82M24.4$44.40 – $2,335.00
22
Arista Networks
Networking & Optics$203.06-0.20%+3.00%+8.14%+25.80%$256.10B$695,17164.1$116.13 – $210.50
23
GE Vernova
Power & Cooling$943.95-0.83%+2.89%+1.03%-10.01%$251.40B$501,98927.1$546.95 – $1,174.86
24
KLA
Semi Equipment$184.44-1.82%+12.25%+3.46%-21.79%$240.98B$1.70M50.3$83.94 – $301.37
25
Marvell Technology
Accelerators$255.39-2.11%+13.58%+13.79%-5.69%$223.95B$3.46M84.3$62.18 – $316.35

Market capDERIVED and P/EDERIVED are computed from SEC filings and the current price

Menu

THE AI CAPEX TRACKER

FROM SEC FILINGS · 23 SEPTEMBER 2026

The number that moves everything

Four balance sheets fund most of the boom.

Combined quarterly capex · big four

$165.05B+212% vs Q2 24

$0B$90B$180BQ2 24$52.85BQ3 24Q4 24Q1 25Q2 25Q3 25Q4 25Q1 26Q2 26$165.05B
MSFT$35.80Blatest
GOOGL$44.92Blatest
AMZN$54.21Blatest
META$30.12Blatest
Data table
QuarterMSFTGOOGLAMZNMETATotal
Q2 24$13.87B$13.19B$17.62B$8.17B$52.85B
Q3 24$14.92B$13.06B$22.62B$8.26B$58.86B
Q4 24$15.80B$14.28B$27.83B$14.43B$72.34B
Q1 25$16.75B$17.20B$25.02B$12.94B$71.91B
Q2 25$17.08B$22.45B$32.18B$16.54B$88.25B
Q3 25$19.39B$23.95B$35.09B$18.83B$97.26B
Q4 25$29.88B$27.85B$39.52B$21.38B$118.63B
Q1 26$30.88B$35.67B$44.20B$19.00B$129.75B
Q2 26$35.80B$44.92B$54.21B$30.12B$165.05B

SEC filings (10-Q, 10-K) via the SEC XBRL API, read 23 September 2026 — every quarter traced to the filing that reported it (EDGAR). These are cash purchases of property and equipment and exclude finance leases, which Microsoft in particular uses heavily — their headline capex figure is higher than the one here. They are also total company capex, not AI capex: AI infrastructure drives the growth, but fulfilment centres and ordinary facilities are in these numbers too. The chart stops at Q2 2026, the last quarter all four have filed.

Full quarterly table, SEC filings and CSV →

Menu

THE MONEY LOOP

The money goes out. Then it comes back

Nvidia invests in the companies that buy its chips, and those chips come back as Nvidia's revenue.

Nvidia

NVDA · Supplier

OpenAI

Customer

Anthropic

Customer

Out of Nvidia

  • 27 February 2026equityreported

    Forbes, quoting OpenAI's own announcement ↗

    Part of a $110bn round at a $730bn pre-money valuation.

Back to Nvidia

Around the loop

COMPANY ANNOUNCEMENTS AND NAMED PRESS REPORTS · 1 AUGUST 2026

Menu

THE CHANNEL

Watch the boom get decoded

AllFactsAI on YouTube is where it started: the AI gold rush, the chips, the money and the winners — explained for investors in minutes, not hours.

Latest from the channel

Watch on YouTube

New videos every week

Menu

MEMBERSHIP

Three ways in. One club.

Spectator is live today: The Daily Receipts every weekday and The Weekly Brief every Saturday, both free.

FREE

Spectator

Available now

  • First look at the Saturday film, and the AI stock of the day
  • The Daily Receipts (weekdays) and The Weekly Brief (Saturdays) — free, sign-up required
  • The full AI stock screener
  • Value chain, money loop and exposure map
Join the club + get the free brief

MEMBER

Insider

In development

  • Everything in Spectator
  • AI company fact cards, every company
  • The complete fact library, searchable
  • Every past issue of both briefs, searchable
  • Fact alerts when a tracked number moves
  • Watchlists, leaderboard and capex history
Free brief + founding list

ALL ACCESS

Pro

In development

  • Everything in Insider
  • Every card figure that changes at each earnings release
  • Pro Charts with indicators
  • Portfolio Connect and AI-exposure X-ray
  • Exclusive Pro videos
  • Data export and API
  • Custom tracked series
Free brief + founding list