On April 14, 2026, theStanford Institute for Human-Centered AIpublished its ninth annual AI Index report. Four hundred and twenty-three pages. Nine chapters. Fifteen top takeaways. Data from Epoch AI, McKinsey, the IEA, Zeki, Brookings, the NRC, and dozens of academic studies. It is the most comprehensive, independently sourced picture of AI’s trajectory available anywhere.[1]
It is also, almost entirely, a report aboutwhat. What was spent? What was built? What was measured? What countries produced how many models, filed how many patents, and deployed how many robots? The data is rigorous, the charts are clean, and the sourcing is transparent.
What the report does not do — what an annual data report is not designed to do — is explainwhyany of it is happening, orwhat breakswhen the underlying structure shifts. Stanford provides the measurement. The diagnosis requires a different kind of analysis.
This piece starts where the AI Index stops. It takes the five most consequential shifts between the 2025 and 2026 editions, layers on the analytical frameworks this publication has been building for the past year, and delivers the analysis the data requires, but the report doesn’t attempt. The data is Stanford’s. The diagnosis is mine.
Five shifts in twelve months
Reading the 2025 and 2026 AI Index reports back-to-back produces an uncanny effect. The format is nearly identical. The chapters follow the same structure. But the numbers have moved — in some cases, dramatically — and the direction of movement tells a story the report itself doesn’t narrate.
Shift 1: The U.S.-China model performance gap closed.
The 2025 report noted the gap was narrowing. On MMLU, MATH, and HumanEval, margins that had been 17–32 percentage points at the end of 2023 compressed to low single digits by late 2024.[2] The 2026 report delivers the conclusion: “U.S. and Chinese models have traded the lead multiple times since early 2025.”[3] The top U.S. model leads by 2.7% as of March 2026 — a margin that fluctuated throughout the year and briefly went to zero in February 2025, when DeepSeek-R1 matched the top American model.[4]
This is not a marginal update. It is a reclassification. Twelve months ago, the framing was “U.S. leads in models.” Today, the framing is “the lead changes hands monthly.”
Shift 2: AI talent stopped moving to the United States.
The number of AI researchers and developers moving to the U.S. dropped 89% between 2017 and 2025. The decline accelerated — down 80% in the last year alone.[5] Net talent flow to the U.S. fell from 324.6 (rolling average, 2022) to 26.0 in 2025.[6] India’s net outflow was -16.9. Canada went negative at -7.1. These are not fluctuations. The U.S. talent magnet is losing its field strength at exactly the moment the model performance advantage has evaporated.
A necessary qualification: the U.S. talentstockremains enormous — 220,520 AI authors and inventors, four times that of any other country. The flow is collapsing, but the base is still dominant. The real question is how long the stock sustains competitiveness without the flow. Talent is not a static asset. It depreciates through attrition, retirement, and the speed at which the frontier moves. A stockpile that isn’t replenished is a wasting asset on a long enough timeline.
The 2025 report lacked the data to flag this trend. It became visible only when the Zeki longitudinal dataset appeared in the 2026 edition.
Shift 3: Capex went exponential. Revenue followed, but on a different curve.
The 2025 report’s most cited finding was the cost decline: inference prices fell 280× in eighteen months for GPT-3.5-equivalent performance.[7] That finding survives in the 2026 edition. But the 2026 data introduces the other side of the ledger.
OpenAI’s annual compute spend rose from roughly $280 million in 2022 to $16.3 billion in 2025.[8] Anthropic went from around $420 million to $8.3 billion over the same period.[9] Google’s annual capital expenditure exceeded $150 billion in 2025, per Citi Research estimates cited in the report.[10] The Stargate venture announced a $100–500 billion range.[11] The HAI report’s 2025 data captures the acceleration, but the 2026 guidance announced since has escalated further — the Big Five hyperscalers’ aggregate 2026 capex is now projected at $660–690 billion, with CreditSights estimating as high as $750 billion.
Private AI investment in the U.S. alone reached $285.9 billion in 2025 (private investment events, including venture capital and private equity, per Quid — a distinct measure from infrastructure capex). That is more than 23 times China’s $12.4 billion in private investment, though, as the report notes, this likely understates China’s total AI spending given government-guided funds.[12]
Revenue is growing, too. OpenAI reached an estimated annualized $25 billion. Anthropic reached an estimated $19 billion.[13] The report positions these as “historically fast” growth trajectories — OpenAI outpacing Uber, Cheniere, and Moderna after crossing $1 billion in annual revenue.[14]
But the report does not perform the operation that matters: dividing the spend by the revenue and asking whether the structure is sustainable. It does not ask how any of these companies finance their compute. It does not check the depreciation schedules. It does not distinguish committed from disbursed. The numbers are presented. The financial architecture is absent.
Shift 4: “AI sovereignty” became an official policy category.
The 2025 report had no sovereignty section. The 2026 report introduces a new analytical framework that breaks sovereignty into five layers: infrastructure, data, models, applications, and talent.[15] Across those layers, the report tracks state-backed AI supercomputing clusters (Europe: 3 to 44 between 2018 and 2025), data localization measures (East Asia Pacific leads with 77; North America has 3), and model production by region (U.S.: 1,618 cumulative; China: 849; Europe and Central Asia: 666; South Asia: 21; Latin America: 2).[16]
The report maps Nvidia’s AI Factory partnerships and OpenAI’s Stargate country-level agreements geographically.[17] It notes that “private firms are playing an increasingly central role in building what many governments designate as national AI infrastructure.”[18]
This is data. It is useful data. But it describes the distribution of assets, not the structure of dependencies. It counts the switches without mapping who holds them.
Shift 5: The labor market cracked along generational lines.
The 2025 report noted productivity gains across several occupations. The 2026 report adds the cost side: employment for U.S. software developers aged 22–25 has fallen nearly 20% from its 2022 peak, even as headcount for older developers continues to grow.[19] Customer support agents show the same generational pattern.[20] AI agent deployment remains in single digits across nearly all business functions.[21] One-third of surveyed organizations expect workforce reductions in the coming year, with the reductions concentrated in service operations, supply chain, and software engineering.[22]
The productivity data is simultaneously positive (14–15% gains in customer support, 26% in software development, 50% in marketing output) and negative (the METR study found experienced developers became 19% slower with AI assistance, though the team has been unable to replicate the finding in later work).[23] The clearest pattern: gains are largest in structured, measurable work. For now, they shrink in tasks requiring deeper reasoning.[24]
These five shifts are not independent. The talent collapse feeds the performance convergence — if researchers stop flowing to the U.S. while China’s output scales, the gap closes. The convergence drives the capex arms race — when the model is no longer the differentiator, the infrastructure bet becomes the competitive move. The capex arms race creates the energy demand. And the energy demand, combined with the sovereignty imperative, drives the infrastructure buildout that the HAI sovereignty section now measures. The report presents these dynamics in separate chapters. The story is the chain between them.
What the data means — and what it doesn’t
Stanford’s contribution is the measurement. The five shifts above are real, sourced, and reproducible. What follows is the explanation the data requires.
The performance convergence is a capex problem, not a capability story
The U.S.-China model convergence sounds like a technology story. It isn’t. It’s a financial architecture story.
When models converge at the frontier — when the top four organizations are separated by just 22 Elo points (a chess-inspired rating system) on the Arena Leaderboard, and six sit within 79[25] — the competitive differentiator shifts from capability to cost, reliability, and distribution. The report notes this shift but doesn’t trace its financial consequence. If the model is no longer the moat, then the infrastructure beneath it is. And infrastructure is a capex game with a specific financial architecture.
Hotel Abileneand the capex series that followed it mapped this dynamic. When AI companies compete on infrastructure rather than intelligence, the financial dynamics change: depreciation schedules matter more than benchmark scores. Free cash flow after capex determines survival. The Commitment-vs-Spend Gap — the ratio between announced investment and actual capital expenditure — becomes the diagnostic metric.[26]
The HAI report provides the ingredients for this analysis. Google’s estimated $150 billion in annual capex (per Citi Research). OpenAI’s compute spend trajectory. The Stargate $100–500 billion range. But it does not apply the Depreciation Lens (has Google extended server useful life assumptions to flatter the income statement while capex rises?), the FCF Sustainability Test (is the incremental AI capex funded from operating cash flow or from debt?), or the Revenue Attribution Problem (when Google adds AI features to existing products, is the incremental revenue “AI revenue”?).[27]
The consumer surplus finding — $172 billion in annual value to U.S. consumers (estimated via willingness-to-accept survey, N=2,000), with a median per-user value that tripled between 2025 and 2026[28] — is the most revealing number in the entire report. If most of the value from generative AI accrues to consumers through free tools, and innovators historically capture only ~3% of total social returns (the Nordhaus finding the report itself cites[29]), then the capex-to-revenue conversion problem isn’t a timing issue. It’s a built-in feature of the economy. The money goes in at the infrastructure layer. The value emerges at the consumer level. The distance between those two layers is the entire AI financial problem.
The talent collapse validates the doom loop — across multiple countries simultaneously
The 89% decline in AI talent moving to the United States is not an immigration policy footnote. It is a systemic shift in the global talent market with direct implications for every country in the AI Realist’s national analysis series.
For Japan, the HAI data confirms what the doom loop predicts.[30] Japan has 6,280 AI authors and inventors — fewer than Singapore’s 6,610, despite having more than 20 times the population. Japan’s net talent flow: 0.0.[31] Not positive, not negative. Zero. The doom loop’s terminal state: a system that neither produces nor attracts the talent it needs, and has reached equilibrium at a level far below what its economic scale would suggest.
For India, the data confirms the services equilibrium thesis. India has the largest net outflow of any country tracked (-16.9).[32] Fifty thousand AI researchers and developers are in the system; the system exports them. The HAI data makes the “building for everyone but themselves” diagnosis quantitative.
For France, the model count is the verdict. France produced one notable model in 2025. Europe as a whole produced two.[33] The total cumulative model count for Europe and Central Asia (666) is less than half of China’s (849).[34]Mistral Succeeded. France’s AI Strategy Didn’twas a diagnosis of a sixty-year state apparatus pattern — the tax credit system, the Grandes Écoles pipeline, the state investment bank. The HAI data is the annual physical.
The per capita talent density numbers reveal a different cut: Switzerland leads the world at 110.5 per 100,000 inhabitants, Singapore at 109.5.[35] These are countries that appear in the HAI data as punching above their weight — and they’re the same countries the sovereignty analysis identifies as facing the trilemma most acutely. High talent density, deep foreign platform dependency, no viable path to full-stack sovereignty.
The sovereignty section describes the map. It doesn’t show the switches.
HAI’s new five-layer sovereignty framework (infrastructure, data, model, application, talent) is a taxonomic contribution. It organizes the conversation. This is not a failing of the report — HAI is a measurement exercise, and a rigorous one. The gap exists because the questions that matter most for investment and policy decisions require a different kind of analysis, one that starts where measurement ends. Specifically: mapping not whohaswhat, but who cantake it away.
The coercion stack — the three-switch model published inAccess, Disable, Destroy— operates at a different analytical level.[36] HAI counts state-backed supercomputing clusters. The coercion stack asks: if the chips inside those clusters are fabricated by one Taiwanese foundry (which the HAI report itself flags as a dependency[37]), and the cloud services running on those chips are operated by companies subject to U.S. jurisdiction, and the models trained on that compute are licensed under terms that permit geographic restrictions — then what is the “sovereignty” of the cluster?
The report maps Nvidia AI Factory partnerships across dozens of countries.[38]Every Country Needs Sovereign AI. Jensen Is Selling It.mapped the same partnerships and asked the question HAI doesn’t: is the Nvidia AI Factory model sovereignty or capture? The “closed orbit” framework — Nvidia’s ecosystem as a black hole that routes all activity back to Nvidia hardware[39] — directly explains the pattern the HAI map displays. More countries are standing up “sovereign” compute infrastructure. Nearly all of it runs on Nvidia silicon, managed through Nvidia’s software stack, with maintenance dependencies that create ongoing operational leverage.
The data localization count (77 measures in East Asia Pacific, 3 in North America) is similarly descriptive without being diagnostic. The Entity Test asks a different question: does localization create legal sovereignty, or does it create a data residency requirement that the CLOUD Act’s compelled disclosure provision (18 U.S.C. § 2713) renders moot?[40] A data center in Frankfurt running on U.S.-controlled cloud infrastructure, using U.S.-fabricated chips, operated by a subsidiary of a U.S. parent company, does not become “sovereign” because German data localization law requires the bits to stay in Germany. The bits comply. The legal exposure doesn’t.
HAI acknowledges that “open-source development is starting to redistribute participation.” But the data supports the closed orbit thesis: open source redistributesactivitywithout redistributingsovereignty, because the hardware dependency persists through every fork, every download, and every deployment.[41]
The labor market data confirms the process thesis, not the tool thesis
The productivity studies the report compiles — a 26% gain in software development, 14–15% in customer support, 50% in marketing output[42] — are study-level findings. They describe outcomes in specific contexts with specific implementations. The report presents them as evidence that “AI’s productivity effects are highly context dependent” and that “gains are strongest when work can be divided into well-defined, repeatable tasks.”[43]
The observation is correct. But “context dependent” is a description, not an explanation.AI Tools Work. Your Engineering Process May Not.supplied the mechanism: AI coding tools amplify existing organizational strengths and weaknesses.[44] Organizations with specification discipline, test-driven development practices, and senior engineer judgment as governance layers get the gains. Organizations without those structures end up with debt.
The junior developer employment decline — nearly 20% among ages 22–25 [45] — is the predictable consequence. There is no evidence that AI replaces junior developers. It is evidence that organizations, when given a tool that produces code faster, reduce the hiring they perceive as most substitutable. The METR finding that experienced developers were 19%slowerwith AI assistance[46] is not a contradiction — it’s the same thesis from the other end. Experienced developers, whose value lies in judgment and architectural decisions rather than code velocity, gain less from a tool that accelerates code generation. Their productivity metric is not lines per hour.
The learning penalty finding — software engineers who relied heavily on AI for learning showed no measurable speed improvement[47] — closes the loop. If junior developers are hired less, and the juniors who are hired learn less because they lean on AI tools, the senior talent pipeline degrades. The result is a doom loop in the labor market, identical in mechanism to the country-level doom loops in the national analysis series.
The energy data shows the demand. Nobody is showing the supply.
AI data center power capacity reached 29.6 GW, comparable to New York state at peak demand.[48] Grok 4’s estimated training emissions reached 72,816 tons of CO₂ equivalent.[49] Annual GPT-4o inference water consumption, per HAI estimates, may exceed the drinking water needs of 12 million people.[50] The IEA projects data center electricity consumption will roughly double between 2024 and 2030, with the U.S. accounting for the largest share.[51]
These are the demand numbers.
What no measurement exercise can do — not this one, not any — is apply the Nuclear Delivery Test: reactor status, fuel type, distance. Every hyperscaler nuclear announcement has a credibility score against those three questions, and most score poorly on them.[52] The 29.6 GW demand figure appears without any assessment of whether the supply pipeline — SMRs that require HALEU fuel that doesn’t exist at a commercial scale, sited in locations that don’t overlap with existing datacenter clusters, on timelines that don’t close before 2030 — will produce operational power.[53]
The bridge technology question — what powers AI datacenters through 2030 while the nuclear press releases age — is the analytical gapThe Half-Life of a Press Releasefills. The answer is natural gas, and the companies that benefit from the nuclear-datacenter timeline mismatch are not the ones in the press releases.[54]
The energy efficiency data is worth noting: DeepSeek V3’s training emissions were 597 tons, compared to Grok 4’s 72,816.[55] Per-query inference energy for Claude 4 Opus (5–6 Wh) is a quarter of DeepSeek V3.2 (23 Wh).[56] These efficiency spreads are enormous — two orders of magnitude at the training level, four to one at inference — and they directly affect the economic viability of the Reasoning Tax framework.[57] A reasoning model that generates 40,000 output tokens per query at 23 Wh per query falls into a different economic and environmental category than one that generates the same tokens at 5 Wh per query.
What was the thesis is now data
The most useful exercise in reading the 2026 report alongside a year of published analysis is watching arguments that were thesis-stage become data-confirmed — and noting where the data complicate the thesis rather than confirm it.
Confirmed: Model convergence shifts competition to infrastructure.Published in “Hotel Abilene” and “Open Source, Closed Orbit.” HAI 2026 data: four companies within 22 Elo points at the frontier, six within 79; competitive pressure “shifting toward cost, reliability, and domain-specific performance.”[58]
Confirmed: European AI sovereignty is a policy aspiration without industrial substance.Published in the France piece and subsequent sovereignty analysis. HAI 2026 data: Europe produced two notable models in 2025. France produced one. Europe and Central Asia’s cumulative model count (666) is under half of China’s. Meanwhile, Europe has 44 state-backed supercomputing clusters, second only to China. Infrastructure without intelligence.[59]
Confirmed: The U.S. structural advantage is at the chip and cloud layers, not the model layer.Published inAccess, Disable, Destroy. HAI 2026 data: model performance gap at 2.7%; TSMC fabricates nearly every leading AI chip; the U.S. hosts 5,427 data centers, more than 10 times any other country.[60] The advantage is physical and legal, not algorithmic.
Complicated: The process-vs-tool distinction is blurrier than “AI Tools Work” assumed.The METR study found experienced developers became 19% slower with AI assistance — consistent with the thesis that AI tools amplify existing process weaknesses.[61] But the 2026 report adds a wrinkle: METR has been unable to replicate the finding, primarily because developers now refuse to work without AI tools.[62] If the tool has become indispensable even though it measurably slows people down, the organizational-process explanation may be necessary but not sufficient. The tool is changing the process faster than the process can be measured. This doesn’t break the thesis: specification discipline and test-driven development still separate organizations that get gains from those that don’t. But it means the baseline has shifted under the measurement.
Three of these were published months before the HAI data confirmed them. That’s what the analysis does: it identifies the forces before the measurement catches up. The doom loops, the coercion stack, the capex frameworks — these aren’t predictions. They’re descriptions of systems that produce predictable outcomes. The HAI report provided 423 pages of those outcomes. The frameworks were already waiting for the data.
Where the data pushes back
An honest reading requires asking whether the 2026 reportcontradictsanything this publication has argued. Two findings deserve scrutiny.
The first is the productivity J-curve. U.S. productivity growth reached 2.7% in 2025, nearly double the prior decade’s average.[63] A study of 12,000 European firms found a 4% boost in labor productivity from AI adoption.[64] The OECD projects annual labor productivity growth of 0.2 to 1.3 percentage points for G7 economies over the next decade.[65] These are not vendor claims. They are peer-reviewed macro findings. The capex thesis — that the financial structure of AI investment contains risks the market underprices — does not assume AI doesn’t work. But the bull case is accumulating data. A note of intellectual honesty: the consumer surplus finding used earlier in this piece to frame the value-capture gap, and the J-curve hypothesis cited here as the strongest challenge to the capex thesis, come from the same researcher — Erik Brynjolfsson, who sits on HAI’s steering committee. His data supports both readings. The bear case says the gap between value created and value captured is permanent. The bull case says it closes as organizations learn to monetize AI. Both are defensible; the data has not yet resolved the question. The risks in the financial architecture remain (depreciation manipulation, debt-funded capex, commitment-vs-spend gaps), but the “what if it works?” scenario is no longer speculative. It’s showing up in the productivity statistics.
The second is France.Mistral Succeeded. France’s AI Strategy Didn’t.argued that France cannot build frontier AI. The HAI adoption data shows France ranking fifth globally in population-level AI diffusion at 44% — ahead of the United Kingdom, Germany, and the United States.[66] France can’t build. But it adopts more aggressively than the countries that can. This doesn’t break the thesis. It sharpens it. France is a structurally enthusiastic consumer of technology; it is structurally incapable of producing it. The sixty-year state apparatus pattern optimizes for adoption and regulation, not for production. The HAI data makes that distinction crisper than the original piece did.
Neither finding overturns a published thesis. But both add complexity that the next iteration of the analysis must absorb.
Where the data stops
The HAI AI Index is the best longitudinal dataset on AI’s trajectory. It is also, by design, a measurement exercise. It cannot diagnose causal mechanisms. It cannot trace legal pathways. It cannot run financial sustainability tests. It cannot model coercion scenarios.
These are not criticisms. They are boundary conditions. The AI Index measures what can be counted. This publication explains what the counts mean.
Five questions the data raises but doesn’t answer:
Thefinancial architectureof AI investment — depreciation schedules, FCF sustainability, debt vs. equity financing of capex, the gap between committed and disbursed capital — is the difference between knowing how much was spent and knowing whether the spend survives a revenue disappointment.
Thelegal architectureof sovereignty — CLOUD Act pathways, entity test analysis, the distinction between data residency and data access jurisdiction. Without it, counting localization measures tells you nothing about whether they provide legal protection.
Thecoercion topology— who holds the off switch at each layer, through what legal and commercial mechanism, and on what timeline. Without it, a map of AI asset distribution is mistaken for an understanding of AI power.
Theprocess layerbeneath productivity statistics — what separates the enterprises that achieve 26% gains from those where experienced developers slow down. Without it, outcomes are measured, but mechanisms are invisible.
Thesupply-side energy analysis— reactor status, fuel availability, geographic mismatch, bridge technology economics — is the difference between knowing the demand and knowing whether the demand gets met.
Stanford measures the field. This publication maps the forces underneath it. The 2026 data confirms that the forces we’ve been mapping are real — and the two places the data pushes back make the analysis sharper, not weaker.
You can wait twelve months for the 2027 report to tell you what happened. Or you can read the structural diagnosis here every week, while there’s still time to act on it.
Notes
[1] Stanford Institute for Human-Centered Artificial Intelligence, “Artificial Intelligence Index Report 2026,” April 2026. Ninth edition. 423 pages, 9 chapters. All HAI citations in this piece reference this edition unless noted as the 2025 edition. Available athttps://hai.stanford.edu/ai-index/2026-ai-index-report.
[2] Stanford HAI AI Index Report 2025, Chapter 2, “Technical Performance.” MMLU gap: 17.5 pp; HumanEval gap: 31.6 pp at end of 2023; all compressed to single digits by end of 2024.
[3] HAI 2026, Top Takeaways, #2.
[4] HAI 2026, Top Takeaways, #2. Arena Elo ratings as of March 2026: Anthropic 1,503; xAI 1,495; Google 1,494; OpenAI 1,481; Alibaba 1,449; DeepSeek 1,424. Chapter 2 Highlights.
[5] HAI 2026, Chapter 1, Section 1.8. Source: Zeki Data, 2025. The 89% figure is cited in both Top Takeaways (#7) and Chapter 1 Highlights (#9).
[6] HAI 2026, Section 1.8, “Mobility,” Figure 1.8.6. Rolling 12-month average.
[7] HAI 2025, Chapter 1, Report Highlights #6. Cost to query a model scoring 64.8 on MMLU fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024.
[8] HAI 2026, Section 4.2, Figure 4.2.20. Source: Epoch AI, 2026. OpenAI compute spend includes R&D, inference, and unattributed categories. The 2022 figure is approximately $280 million; the 2025 figure is $16.3 billion.
[9] HAI 2026, Section 4.2, Figure 4.2.20. Anthropic's 2022 figure is approximately $420 million; its 2025 figure is $8.3 billion.
[10] HAI 2026, Section 4.2, “Capital Expenditures.” Google “reporting more than $150 billion in annual capex in 2025.” Source: Citi Research, Figure 4.2.21.
[11] HAI 2026, Section 4.1, timeline entry January 21, 2025. “Between $100 billion and $500 billion” by 2029. Post-HAI data: Q4 2025 and Q1 2026 earnings guidance from the Big Five hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) produced aggregate 2026 capex projections of $660–690 billion (CreditSights, Futurum, multiple analyst estimates, February 2026). Amazon’s $200 billion single-year commitment was the largest in corporate history. CreditSights subsequently raised its aggregate estimate to ~$750 billion. Morgan Stanley projects hyperscaler borrowing exceeding $400 billion in 2026, more than double 2025’s $165 billion. Capex-to-revenue ratios for major hyperscalers reached 45–57% levels previously associated with industrial or utility companies, not technology firms.
[12] HAI 2026, Top Takeaways, #7 and Chapter 4 Highlights, #2. The report explicitly notes that “private investment figures likely understate China’s total AI spending, as government guidance funds have deployed an estimated $184 billion into AI firms between 2000 and 2023.”
[13] HAI 2026, Section 4.2, Figure 4.2.18. Source: Epoch AI, 2026. Described as “annualized revenue estimates” from “direct company statements or established media reporting” — i.e., these are not from audited filings. The report cautions that figures are “directional rather than precise.”
[14] HAI 2026, Section 4.2, Figure 4.2.19.
[15] HAI 2026, Section 8.3, “AI Sovereignty.” The five layers: infrastructure sovereignty, data sovereignty, model sovereignty, application sovereignty, and talent sovereignty.
[16] HAI 2026, Section 8.3. Supercomputing clusters: Figure 8.3.1 (Epoch AI data). Data localization measures: Figure 8.3.3 (Ferracane et al., 2026). Model production: Figure 8.3.4 (Epoch AI cumulative model releases, 2018–2025). Note these are all publicly documented models, not just “notable” models per Epoch AI criteria — a broader count than Chapter 1’s notable model dataset.
[17] HAI 2026, Section 8.3, Figure 8.3.2. Geographic map of Nvidia AI Factory and OpenAI Stargate country-level partnerships.
[18] HAI 2026, Section 8.3, “Infrastructure Sovereignty.”
[19] HAI 2026, Section 4.4, “Workforce Impact,” Figure 4.4.29. Source: Brynjolfsson et al., 2025. “Employment for software developers ages 22–25 had fallen close to 20% from its 2022 peak.” Also cited in Top Takeaways (#9).
[20] HAI 2026, Section 4.4, Figure 4.4.29. Customer service agents show a parallel generational pattern.
[21] HAI 2026, Section 4.3, “Deployment Stages,” Figure 4.3.7. “Scaled use was in the single digits for nearly all functions.”
[22] HAI 2026, Chapter 4 Highlights, #8. McKinsey survey.
[23] HAI 2026, Section 4.4, “Productivity Trends,” Figure 4.4.27. Customer support: Brynjolfsson et al., 2025 (14–15%). Software development: Cui et al., 2025 (26%). Marketing: Ju & Aral, 2025 (50%). METR: Becker et al., 2025 (-19%). The report notes METR “has not been able to replicate the results in a later study, primarily due to a growing reluctance among developers to work without AI.”
[24] HAI 2026, Section 4.4. “Gains are strongest when work can be divided into well-defined, repeatable tasks with clear quality monitoring.”
[25] HAI 2026, Chapter 2 Highlights, #2. Arena Elo ratings as of March 2026 cited in note [4]: Anthropic (1,503) to OpenAI (1,481) spans 22 points; Anthropic (1,503) to DeepSeek (1,424) spans 79 points.
[26] Frameworks published inHotel Abilene(March 2026),Cloud vs. Clout(March 2026), andTrain, Deploy, Write Down(March/April 2026), The AI Realist, www.airealist.ai. The Commitment-vs-Spend Gap is defined in the capex-finance vertical as the ratio of announced AI investment commitments to actual capital expenditure in the most recent filing.
[27] The Depreciation Lens, FCF Sustainability Test, and Revenue Attribution Problem are analytical frameworks published across the CapEx-Finance series. Specific definitions: Microsoft extended the useful life of servers from 4 to 6 years in 2022 (the reference case for the Depreciation Lens); Meta, Google, and Amazon followed suit. The Revenue Attribution Problem asks whether “AI revenue” is genuinely new or reclassified revenue from products that now contain AI features.
[28] HAI 2026, Section 4.2, Highlight: “What Is Generative AI Worth?,” Figure 4.2.22. Brynjolfsson et al., 2026. Consumer surplus grew from $112B to $172B annually. The median value per user tripled, from $3.40 to $11.40.
[29] HAI 2026, Section 4.2. “This pattern is consistent with findings by Nordhaus (2004) that innovators historically capture only ~3% of total social returns from major technologies.”
[30]Japan Built the Bullet Train. Why Can’t It Build an LLM?, The AI Realist. The doom loop: Low AI salaries → talent leaves → companies can’t build AI → companies buy from U.S. → domestic ecosystem stays small → no market pressure to raise salaries → low AI salaries.
[31] HAI 2026, Section 1.8, Figure 1.8.6. Japan’s 2025 net flow: 0.0. Total AI authors and inventors: 6,280 (Figure 1.8.1). Singapore: 6,610.
[32] HAI 2026, Section 1.8, Figure 1.8.6. India 2025 net flow: -16.9.
[33] HAI 2026, Chapter 1, Section 1.1, Figure 1.1.1. France: 1 notable model in 2025. Europe total: 2 notable models in 2025. Note: this is based on the Epoch AI “notable models” dataset, which applies criteria such as state-of-the-art performance and high citation counts. The broader model count in Chapter 8 (666 cumulative for Europe and Central Asia) covers all publicly documented models.
[34] HAI 2026, Section 8.3, Figure 8.3.4.
[35] HAI 2026, Section 1.8, Figure 1.8.2. Switzerland: 110.5 AI authors and inventors per 100,000 inhabitants. Singapore: 109.5.
[36]Access, Disable, Destroy, The AI Realist, March 2026. The coercion stack: three layers (chips, cloud, models), mapping who holds the off switch, through what legal/commercial mechanism, and on what timeline.
[37] HAI 2026, Top Takeaways, #3. “A single company, TSMC, fabricates almost every leading AI chip, making the global AI hardware supply chain dependent on one foundry in Taiwan.”
[38] HAI 2026, Section 8.3, Figure 8.3.2.
[39]Open Source, Closed Orbit, The AI Realist, March 2026, 6,439 words, 100 footnotes. The “black hole” framework: Nvidia’s ecosystem is centripetal — every contribution routes activity back to Nvidia hardware, converting open-source adoption into hardware lock-in.
[40] CLOUD Act compelled disclosure provision, 18 U.S.C. § 2713. The Entity Test is published in The AI Realist's sovereignty vertical. Key factors: ownership chain, incorporation jurisdiction, personnel with technical access, contractual dependencies, and management control. If any factor creates a link to U.S. jurisdiction, the compelled disclosure provision may apply. Note: data localization (where data physically resides) is distinct from data governance (who controls processing decisions under GDPR’s controller/processor framework). The Entity Test applies to both, because the question is jurisdiction over the entity, not the location of the bits.
[41]Open Source, Closed Orbit. Open-weight models can be downloaded, fine-tuned, and deployed on any hardware. But the most performant deployment requires Nvidia hardware, Nvidia’s inference stack (TensorRT-LLM, NIM), and Nvidia’s optimization libraries. The fork is free. The performance depends on the hardware.
[42] See note [23].
[43] HAI 2026, Section 4.4, “Productivity Trends.”
[44]AI Tools Work. Your Engineering Process May Not., The AI Realist, March 2026, 5,380 words, 24 footnotes.
[45] See note [19].
[46] See note [23]. Becker et al., 2025 (METR). -19% speed for experienced open-source developers using AI assistance.
[47] HAI 2026, Section 4.4, Figure 4.4.27. Shen and Tamkin, 2025. “Software engineers who relied heavily on AI for learning showed no measurable speed improvement.”
[48] HAI 2026, Top Takeaways, #10.
[49] HAI 2026, Section 1.4, Figure 1.4.3. Epoch AI estimate.
[50] HAI 2026, Section 1.4, Figure 1.4.8. “Annual estimates for GPT-4o inference range from about 1.3 to 1.6 kiloliters [presumably billion liters], which, at the high end, exceeds the annual drinking water needs of 12 million people.” The units, as stated in the report, appear internally inconsistent with the figure and comparison — verify against the original source (de Vries and Gao, 2025) before citing the absolute number. The order-of-magnitude comparison (AI inference water use ≈ drinking water for 12M people) appears consistent between the figure and the text.
[51] HAI 2026, Section 1.4, Figure 1.4.13. Source: IEA, 2025. Note: “Data in this chart reflects IEA projections rather than observed consumption.”
[52]The Half-Life of a Press Release, The AI Realist, March 2026, 4,896 words, 88 footnotes. The Nuclear Delivery Test: three questions (reactor status, fuel type, distance) producing a credibility score for any hyperscaler-nuclear deal announcement.
[53] The four-layer mismatch framework: geography mismatch (nuclear sites ≠ datacenter sites), timeline mismatch (NRC licensing path exceeds the 2025–2028 crisis window), cost mismatch (nuclear construction costs have risen faster than inflation for four decades), fuel supply mismatch (HALEU commercial production at scale does not exist as of 2026).
[54]The Half-Life of a Press Release, bridge technology analysis. Natural gas peaker plants and combined-cycle gas can be permitted and built in 18–36 months, compared to 5–20 years for nuclear.
[55] HAI 2026, Section 1.4, Figure 1.4.3. DeepSeek V3: 597 tons CO₂eq. Grok 4: 72,816 tons CO₂eq. Both figures are Epoch AI estimates. The DeepSeek figure is dramatically lower than peer models at comparable parameter count; this may reflect differences in training efficiency, hardware utilization, energy grid carbon intensity, or disclosure methodology. Treat the absolute number as directional.
[56] HAI 2026, Section 1.4, Figure 1.4.5. Claude 4 Opus: 5.13 Wh per medium-length prompt. DeepSeek V3.2: 23.13 Wh. Source: Jegham et al., 2025. A medium-length prompt is approximately 1,000 input tokens and 1,000 output tokens.
[57] The Reasoning Tax: total cost multiplier for reasoning-heavy workloads versus standard inference. Published in the AI tooling vertical of The AI Realist. The formula accounts for token inflation (reasoning models generate 10–40× more output tokens per query) and price premium.
[58] HAI 2026, Chapter 2 Highlights, #2.
[59] HAI 2026, Chapter 1 Highlights, Figures 1.1.1 and 8.3.1 and 8.3.4.
[60] HAI 2026, Top Takeaways, #2 and #3.
[61] See note [23]. Becker et al., 2025 (METR). The -19% finding for experienced open-source developers was the most widely cited negative productivity result in the field.
[62] HAI 2026, Section 4.4. The report notes METR “has not been able to replicate the results in a later study, primarily due to a growing reluctance among developers to work without AI, and that developers in late 2025 were likely sped up by AI relative to the original study period.”
[63] HAI 2026, Section 4.4, Figure 4.4.28. Brynjolfsson, 2026. 2.7% U.S. productivity growth in 2025, framed through the “J-curve” hypothesis — organizations absorb the costs of adopting AI before larger productivity gains materialize.
[64] HAI 2026, Section 4.4, Figure 4.4.28. Aldasoro et al., 2026. Study of 12,000 European firms (2019–2024). 4% increase in labor productivity from AI adoption; 5.9 percentage point gain for every 1% spent on training.
[65] HAI 2026, Section 4.4, Figure 4.4.28. Filippucci et al. (OECD, 2025). Projected annual gains: +0.4 to +1.3 pp (U.S./UK) vs. +0.2 to +0.8 pp (Italy/Japan).
[66] HAI 2026, Section 4.3, Highlight: “Measuring Signals of AI Diffusion,” Figure 4.3.12. Source: Microsoft AI Economy Institute, 2025. France: 44.0% AI diffusion (second half 2025), ranked 5th globally. United States: 28.3%, ranked 24th. UAE: 64.0%, ranked 1st. Singapore: 60.9%, ranked 2nd.