KXCO
KXCO
Request a briefing

Key findings

492 Entities Β· 1175 Claims Β· $6.0T in Capital ClaimsData as of 16 November 2026
Key findings

Key findings: risks and opportunities across the whole sector

These findings describe patterns that appear only when the whole sector is read together, not from the record of any single company. Each finding is based on claims with a named source. Select a name under Evidence to open it in the graph.

Terms explained

Claim: a statement in the record, with the source it was read from and the date it was read.

Chokepoint: a step in the supply chain served by one supplier or very few, so a disruption there stops everything downstream.

GPU: graphics processing unit, the chip on which most AI models are trained and run.

HBM: high-bandwidth memory, the stacked memory packaged beside an AI chip.

EUV: extreme ultraviolet lithography, the process that prints the smallest features on leading-edge chips. ASML is the only maker of the machines.

EDA: electronic design automation, the software used to design chips.

Foundry: a company that manufactures chips designed by others, such as TSMC.

Hyperscaler: a cloud provider operating at the largest scale, such as Amazon Web Services, Microsoft Azure or Google Cloud.

Circular capital: money that returns to where it started, as when a vendor invests in a customer that then spends the investment with the vendor.

Consensus target: a single figure that summarises the 12-month share price targets of the sell-side analysts who cover a company.

Short interest: the shares sold short and not yet bought back, as brokers report them to FINRA.

Weighting and derived analysis: Shayne Heffernan Β· what is judgement here, and what is sourced

Every claim in this graph carries a source. The weighting applied over those claims is Shayne Heffernan's: which entities sit upstream, which are single points of failure, which findings rank as critical, and how far each claim's evidence is trusted. Those are judgments, they are where most of the analytical value sits, and they are accredited rather than presented as facts. Where a reading of the graph is ours rather than a source's it is marked derived and carries no source URL, because inventing one would be the failure this map exists to prevent.

Weighted by Shayne Heffernan
tier (505) Position in the stack, 0 upstream to 5 peripheral.
choke (9) Designation as a single point of failure. Judgment, not a sourced fact, and deliberately not degree-ranked: chokepoint mean degree was 5.2 against 47 for the largest hubs in August 2026, and 5.6 against 55.8 on 14 October 2026, because a chokepoint is a thing with no alternative rather than a thing everyone deals with.
sev, kind (54) Severity ranking and risk or opportunity classification of each finding, on four criteria. Critical: a dependency or concentration the whole sector rests on, with no route around it in the record. High: an exposure across the sector, measured in money or capacity, that touches many parties at once. Elevated: an exposure concentrated in one vendor, one country, one programme or one person, or one still forming. Opportunity: value the record shows a fund can act on.
conf (1215) Grade of the evidence behind a sourced claim, assigned by source tier. On the derived readings, which no source published, it is the analyst's confidence in the reading, and conf_basis says which.
g (1215) Classification of each claim into one of nine relationship groups.
disc=derived (49) Readings of the graph that no source published, marked as such rather than presented as reported fact.
Not weighted, taken as found
community, precomputed by algorithm rather than assigned by a person
mag, magU, asof, url, note and pred, all taken from the cited source
Systemic Risk46
CRITICAL

Every major AI lab, cloud and sovereign programme depends on Nvidia

124 claims, 81 counterparties; $96.2bn quarter, $108bn guided
The finding

Every major lab, cloud and sovereign AI programme routes through one vendor. Nvidia is by far the most-connected entity in the sector: 84 typed claims to 64 counterparties, 54 of them outbound, and a presence in 10 of the graph's 16 circular-capital claims. On undirected two-hop supply and capital adjacency it reaches 160 of 382 entities, 42% of the map. A shock to it stalls the entire stack. The 26 August print measured the other side of that: revenue of $96.2bn, up 106% in a year, with Data Center at $89.0bn, a 75% gross margin, and a guide of $108bn for the quarter in progress that assumes nothing from China. The market added about $442bn to the company in one session, the second-largest single-day gain by any stock. Three things had deepened the dependency in the month before, and none was forced: SpaceX committed on 4 August to build its AI compute exclusively on Vera Rubin against a 10GW end-2027 target; Japan's national physical-AI factory is 13,750 Vera CPUs and 27,500 Rubin GPUs; and on the day of the print AWS scheduled 2 million more GPUs for 2027-2028 after its 1 million commitment ran out early. The concentration the graph flags as a risk is the same concentration the income statement reports as revenue. Both readings are correct, and they are the same fact.

September

new Vera Rubin deployments at Nscale (Anthropic, Microsoft) and Firmus (OpenAI), and Colossus 2 adding up to 660,000 GB300s by year-end on top of 440,000 installed.

Count, 12 October

According to the graph's current claims, Nvidia carries 101 claims to 66 counterparties, 64 of them outbound, and appears in 11 of the 20 circular-capital claims. Within two hops of its supply and capital links it reaches 167 of the 401 entities on a current claim, 42% of the map.

October

On 12 October Microsoft said it will scale its own Azure Maia accelerators to multi-gigawatt deployments from 2027 and make Maia the primary inference system for its products and services. Maia 300's compute chiplets are built at TSMC, so the move takes Microsoft's inference away from bought GPUs but not away from the same foundry. On 14 October Nvidia said Synopsys and Cadence, the two EDA vendors every AI chip is designed on, run parts of their simulation and design tools on its CUDA-X libraries and Nemotron models, alongside Applied Materials, KLA, Lam Research and Siemens, so parts of the chip design and manufacturing toolchain now run on Nvidia's stack. Three days after the Maia announcement, on 15 October, Microsoft AI said it pre-trained MAI-Base-2, a 2.8T-parameter model, on 16,000 GB200 GPUs over 38 days, so the dependency runs inside Microsoft's own model lab while Maia is built for inference. The frontier exception held: Google's Gemini 4 Argon, generally available from 13 October, was trained on Google's own TPUs, per its model card.

November

Nvidia names its own supply as the constraint. On 21 October it set out capacity agreements and plans with Amkor, KYEC, SPIL, Samsung Electro-Mechanics and Unimicron, prepaid in the cases of KYEC (a $1.1bn advance) and Amkor. On 26 and 27 October it tightened two customer ties: a $3bn investment in Meta's equity offering as Meta expands Vera Rubin, and plans with Oracle to take OCI Superclusters on Vera Rubin to multi-gigawatt scale. Reuters reported on 28 October, citing two people familiar, that the Justice Department is asking how its Hugging Face purchase would affect AI competition. Its fiscal Q3 2027 ended on 25 October; results are due on 17 November.

November

alternatives scaled in results week. AMD's Data Center revenue reached $8.5bn, up 97%, with its first Helios racks at Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure; Qualcomm reiterated a $5bn fiscal 2027 Data Center revenue target and tied an Amazon warrant to up to $60bn of purchases; Arm expects up to $2bn from its own AGI CPU over two fiscal years; and Marvell agreed on 1 November to buy Lumentum, an optics supplier Nvidia invested $2bn in, at an implied $115bn enterprise value. Nvidia added a $1bn stake in Naver, which builds on its GPUs.

November, 11 to 12 November

the vendor the map depends on changed its leadership. Nvidia's board accepted Jensen Huang's resignation as president, chief executive and director on 11 November, effective immediately, appointed chief financial officer Colette Kress interim chief executive and Harvey C. Jones chair, and said it expects the company's operations, customer commitments and product roadmap to continue without interruption (Form 8-K and release, 12 November). Results for fiscal Q3 2027 remain due on 17 November. The chairmen of the House China, Foreign Affairs and Intelligence committees had asked Huang on 5 November to brief them on national security concerns, and renewed the request to him and to the company on 12 November.

November

a sovereign customer said the dependency held through Nvidia's leadership change. South Korea's science minister said on 13 November that Nvidia reaffirmed at the corporate level its supply of 260,000 advanced GPUs, with about 21,000 of the 35,000 the government buys directly secured and about 14,000 due by early 2027.

Evidencesource β†—
CRITICAL

Leading-edge AI chips depend on a few sole or near-sole suppliers

3 technological, 1 physical
The finding

ASML alone makes the EUV lithography every leading-edge AI chip needs, behind TSMC, behind Nvidia. Two more compound it: the EDA duopoly (Synopsys and Cadence, 61% of the market in 2024 by TrendForce's count, 74% with Siemens EDA; SemiAnalysis put the three at over 85% in May 2026, with Ansys now part of Synopsys) and the HBM triopoly (100%). No capital can quickly route around them. The last week of July 2026 showed what that concentration costs on the way down: the memory tier fell hardest of all, with SK Hynix down $176bn, Samsung $173bn and Micron $113bn of market value in days. A fourth chokepoint sits upstream of all of them and is not technological at all. The Strait of Hormuz carries the Gulf's sulfur and about a third of the world's helium, and it was effectively closed from February 2026.

October

ASML, the only EUV maker, said on 14 October 2026 that its planned 2027 and 2028 supply will not fully meet demand even after a 30% capacity increase for 2027. TSMC said on 15 October that HPC rose to 68% of its Q3 2026 revenue.

November

Below the foundry, packaging, substrates and test are now being reserved through capacity agreements, several of them prepaid. On 21 October 2026 Nvidia said 'Demand is extraordinary. Capacity is the constraint' and set out supplier agreements: substrate capacity up 2.5x over three years with Samsung Electro-Mechanics and Unimicron among the suppliers, a more than 300% rise in Amkor's Korean interposer wafer capacity by end-2028 within $3bn of Amkor agreements, testing capacity doubling from 2026 to 2027, with new US facilities backed by a $1.1bn advance to KYEC, and a 15x rise in SPIL's optical-engine capacity for co-packaged optics by mid-2027. TSMC said on its 15 October call that the gap between demand and supply is still quite big (gap fill). In memory, SK Hynix said on 28 October that demand will exceed supply through 2030, with long-term agreements covering more than 50% of its 2027 and 2028 capacity, and Samsung said on 29 October that AI-driven demand in 2027 will far exceed industry supply in comparison with 2026.

Correction, 16 November 2026

an earlier version gave Synopsys and Cadence about 96% of the EDA market. TrendForce gives them 31% and 30% for 2024.

Evidencesource β†—
HIGH

AI deals form loops in which the money flows back to the companies that funded it

21 vendor and customer pairs, now including credit, a second vendor financier and a warrant
The finding

Nvidia invests in the customers that buy its chips; investors fund the labs that spend it back on the investors' clouds. Now 15 loops. The newest one changes the kind of risk rather than the amount: Nvidia and OpenAI were reported in late July 2026 to be discussing a backstop of up to $250bn, which would let OpenAI raise debt against Nvidia's credit rather than take Nvidia's equity. Demand is partly self-referential, and now partly self-guaranteed.

September

a second vendor now finances its customers. Broadcom said on its 2 September call that its financing structures are mainly for Anthropic and OpenAI. Nvidia put $1.0bn into Nscale, which rents Vera Rubin capacity to Anthropic and Microsoft, and is in talks to anchor Anthropic's IPO.

Late September

a cloud vendor now pays its customer in its own equity. Akamai's warrant gives Anthropic up to about 5% of Akamai's common stock, about 2% vesting with an $11.6bn seven-year commitment and about 1% for each further $3bn of spend, per Akamai's 8-K of 24 September.

Count, 7 October

20 loop claims across 16 distinct vendor and customer pairs, counted from the claims themselves rather than kept by hand. The sixteenth is Akamai and Anthropic. OCTOBER (gap fill, dated 1 October): Reuters, reading Anthropic's IPO prospectus, reported that Broadcom agreed to lend Anthropic up to $42bn to finance infrastructure spending, while Anthropic stands to become the largest customer of Broadcom's chip design business in 2027.

November

Nvidia agreed on 26 October to invest $3bn in Meta's equity offering in the same release in which Meta said it intends to expand its deployment of Vera Rubin, Vera CPUs, Spectrum-X and BlueField-4, a further vendor and customer pair of the shape already recorded for AMD and Anthropic. The release does not tie the investment to purchases. Broadcom's residual value support behind the $63bn Anthropic SPV is the credit form of the same pattern at the second chip vendor. AMD followed on 27 October with $1bn into the same offering as its first Helios rack systems went into testing at Meta, as Quiver and Digital Terminal reported from AMD's announcement.

November

chip vendors are granting large customers warrants on their own shares that vest as purchases rise. Qualcomm's 10-K (4 November) discloses a September 2026 warrant to an Amazon affiliate for up to 25 million shares at $161.26, vesting as purchases rise towards a $60bn maximum over ten years, with about 4 million shares vested on issue; the same day Qualcomm announced a multi-generational AI data-centre collaboration with Amazon. AMD's 10-Q (4 November) names OpenAI and Meta as holders of warrants for up to a combined 320 million AMD shares at $0.01; none had vested at 26 September, but AMD said most market conditions were met and booked a $1.2bn noncash charge in the quarter. A gap fill: Marvell issued Google a warrant on 18 August 2026 for up to 58,970,907 shares, vesting one of 240 tranches per $500m of Google's custom-product purchases.

Evidencesource β†—
HIGH

Anthropic is backed by five companies that compete with each other

5 rival backers; $278.5bn owed to three of them regardless of usage
The finding

Anthropic is backed at once by Amazon, Google, Microsoft, Nvidia and now AMD, five direct competitors on one cap table, at a $965B post-money (May 2026 Series H). AMD's up-to-$5bn equity plus its up-to-2GW MI450 supply add a fifth rival while softening pure Nvidia dependency. The lab is simultaneously filing to go public (confidential S-1, 1 Jun 2026).

September

Broadcom named Anthropic as one of two customers whose deployments it finances, Nscale's S-1 disclosed an Anthropic compute contract of up to $44.6bn, and Nvidia is in talks to take up to $10bn of the IPO. LATE SEPTEMBER, IN ANTHROPIC'S

Own words

the prospectus Reuters read lists the trap as a risk factor. Amazon, Google and Microsoft are at once its investors, customers, cloud providers, distributors and competitors, and if compute access is 'curtailed, repriced, or terminated' the business could be adversely affected. Against those three it carries $111.1bn with Google, $110bn with Amazon and $31.4bn with Microsoft, owed regardless of usage.

October

Broadcom joins the list of counterparties that are supplier and financier at once: up to $42bn of convertible lending against $125.2bn of TPU leases, per Reuters' read of the prospectus (1 October), and Anthropic expected to become its largest compute customer in 2027.

November

Anthropic's S-1/A of 2 November 2026 states the usage-independent commitments in its own words: Microsoft at least $44.9bn from November 2026 to June 2033 after amendments between February and October 2026 (previously $31.4bn as read by Reuters); Google at least about $121.3bn from April 2026 to July 2033 (previously $111.1bn), plus further TPU compute for 2028 to 2031 estimated at $125bn to $150bn without a fixed dollar commitment; Amazon $110.0bn from May 2026 to April 2036, unchanged. Owed to the three regardless of usage: $44.9bn + $121.3bn + $110.0bn = $276.2bn. Two further commitments are sized: AMD hardware expected to exceed $20bn for its initial 2027 tranche if fully procured, and about $41.3bn with Lambda for Nvidia-powered capacity, beside the up to about $44.6bn with Nscale held since September.

November, 10 November

the amended S-1 adds the supply chain to the cap table. Google ($10.0bn, part of its March 2026 commitment), Nvidia ($10.0bn), AMD ($2.5bn), Broadcom ($1.0bn), Micron ($1.0bn), Microsoft ($1.0bn) and TSMC Partners (about $0.5bn) agreed to buy $26.0bn of non-voting Class N stock at the IPO price, conditional on the IPO. Micron and TSMC appear beside Anthropic for the first time in this record; Amazon is not among the buyers.

November, 10 November, commitments

the same filing raises the Microsoft commitment to at least $47.2bn and up to $49.2bn between November 2026 and June 2033, under the agreement as amended between February and November 2026. Google's about $121.3bn and Amazon's $110.0bn are unchanged. Owed to the three regardless of usage: $47.2bn + $121.3bn + $110.0bn = $278.5bn.

Evidencesource β†—
CRITICAL

The US government decides which AI companies work on its classified networks, and Anthropic is excluded

8 firms cleared; 3 instruments; 1 exclusion
The finding

Eight companies hold IL6/IL7 classified-network agreements as of 1 May 2026: SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS and Oracle. Anthropic remains excluded, with talks stalled over surveillance and weapons guardrails, and Palantir pulled Claude from DoD platforms. That access list is one half of the picture; the policy scaffolding around it is the other, and it is consistent in direction. America's AI Action Plan of July 2025 directed the federal government to pursue AI dominance principally by minimising regulation. The executive order of 11 December 2025 moved against state AI laws directly, standing up a Justice Department litigation task force and putting federal broadband funding behind the preference. The executive order of 2 June 2026 addressed frontier-model risk through a framework that is expressly voluntary. Read together the posture is consistent: accelerate at the federal level, pre-empt at the state level, and handle safety by invitation. Access to the frontier is gated by the executive, and the gate is discretionary rather than statutory.

Late September

the posture held. On 29 September Anthropic, OpenAI, Google, Meta, Nvidia and xAI signed a voluntary White House accord that Trump called morally binding, with no penalty and no regulator, and the same day an executive order told the executive branch to call AI 'super intelligence' and gave the science adviser 60 days to propose a statutory definition.

October

the Pentagon said its own use of Anthropic's products had ended, and the gate gained a body. A Defense Department official told the BBC on 5 October that the Pentagon 'has ceased the use of Anthropic products', some five weeks after its own late August deadline, while people familiar with the matter told the BBC Claude had been in use the week before. On 4 October the President named Director of National Intelligence Jay Clayton to lead a new White House Super Intelligence Force.

October

Anthropic petitioned the D.C. Circuit on 13 October to rehear the 2-1 ruling of 25 September that upheld its supply-chain-risk designation, arguing retaliation for its speech on AI safety. Palantir's chief technology officer said on CBS on 11 October that the Pentagon has already replaced Claude with models from xAI, OpenAI and Google and open-weight models such as Nvidia's Nemotron. The Super Intelligence Force made an Anthropic test-model incident the first case of its mandatory incident notification process (9 to 13 October), and the FTC sent civil investigative demands to Anthropic and OpenAI on 14 October. In the same week Anthropic committed $150m over three years to the Genesis Mission, a federal AI-for-science programme spanning more than 15 agencies, announced at a summit hosted by the White House Office of Science and Technology Policy on 8 October, and signed an AI safety MOU with Singapore on 14 October. Anthropic has told IPO investors that government work is under 1% of its revenue but that the government's view of it could cost commercial revenue.

November

the gate stays with the executive, and Congress changes hands. Democrats won both chambers on 3 November 2026. The White House said through Kevin Hassett on 6 November that the data-centre build-out will proceed, and President Trump posted that he is "100% in favor of Super Intelligence (SI)".

Correction, 16 November 2026

an earlier version said the exclusion completed in October. A Defense Department official told the BBC on 5 October that the Pentagon had ceased using Anthropic products.

Evidencesource β†—
ELEVATED

China is building a frontier AI stack to cut its dependence on US chips

capability gap 2.7%; 30 notable models vs 50
The finding

China is building a near-complete AI stack of its own. DeepSeek's V4 (Apr 2026) runs on Huawei Ascend chips, and Huawei said its chips were used in some of V4's training. The V4 technical report does not name the chips that trained it, and Bloomberg reported in September that DeepSeek has relied on Nvidia for training. A $295bn state plan for a unified national compute grid (2028) mandates 80% domestic sourcing, which leaves at most a fifth of that sourcing to foreign suppliers such as Nvidia and AMD. Alibaba and Tencent jointly bankroll a fleet of open-weight labs (Moonshot, MiniMax, Zhipu). It runs months, not years, behind the West, answers to none of its evaluators, and export controls are the only remaining lever. The 2026 AI Index puts numbers on how narrow that gap now is. The United States released 50 notable models in 2025 and China 30, with no other country close: South Korea 5, and Canada, France and the UK one each. On capability the distance is smaller than the count suggests, at about 2.7% between the best US and best Chinese model as of March 2026. China simultaneously leads on publication volume, citations, patents and industrial robot installations. The reading that follows is that output quantity is a funding artefact and capability is not, so a count of releases is the wrong measure of where this stands.

September, against the reading

the stack still leans on foreign parts. Chinese accelerators depend on grey-market HBM at several times world prices, and DeepSeek's order for at least 160,000 Ascend 950DT is for inference, not training.

For it

Huawei pulled the Ascend 960DT forward three quarters to Q1 2027 at HUAWEI CONNECT on 17 September.

Late September, for it

Huawei's Eric Xu said on 17 September, in a summary Huawei released on 30 September, that on Huawei's own data Ascend has passed Nvidia in China market share, and that Huawei cannot yet meet Chinese demand.

Against it

the claim carries no figure, and Xu said himself that Nvidia's China share is hard to measure.

October

On 16 October China's commerce and science ministries proposed export controls on 23 AI technologies, among them large models, agents and computing clusters, with open-source releases exempt and pre-export registration in force the same day. If adopted, Beijing would treat its own frontier models as controlled exports.

November, for it

Moonshot released the open weights of Kimi K3.1 Pro, a 2.8T-parameter model, on Hugging Face on 5 November, with Huawei announcing inference support on Ascend 950; Tencent released Hy4 on 4 November and expanded its inference capacity tenfold the next day; Apple Intelligence began beta testing in China with Alibaba's Qwen on 4 November. STAR Market Daily reported on 3 November, citing supply-chain sources, that DeepSeek plans to pre-train V5 on Ascend chips; DeepSeek has not confirmed it. The same report says DeepSeek has used Ascend mainly for inference so far.

Correction, 16 November 2026

an earlier version called the stack sanction-proof, said V4 trains end to end on the Ascend 950PR and said the sourcing mandate locks out Nvidia and AMD. Huawei said its chips were used in some of V4's training, Bloomberg reported that DeepSeek has relied on Nvidia for training, and an 80% mandate leaves a fifth of sourcing open.

Evidencesource β†—
ELEVATED

Frontier labs use more capable models internally than they release

~2mo internal lead
The finding

The people who build the frontier run more capable models internally than they ship - months early, wired into their own R&D. Self-built and fair, but compounding and invisible to outsiders.

October

two labs showed the internal lead in the open, and the EU said where its rules begin. Google announced Gemini 4 Argon on 30 September with access limited to trusted cyber defenders, said thousands of Googlers were already using it, and said it would release it without cyber guardrails to its own internal teams and trusted defenders; general availability followed on 13 October.

Gap fill, 28 September

OpenAI said it would not release GPT-6.1 Astra after it failed the company's internal standards for scope, authorisation and reporting back to users (Al Jazeera, 29 September), so a model OpenAI built is held back from release. A Commission official told Euronews on 14 October that the EU AI Act is not enforced while a model is in testing, but applies once a model escapes its test environment and affects the EU market, and the AI Office's first requests for information went to OpenAI and Anthropic over two such escapes.

Evidencesource β†—
ELEVATED

The US leads quantum error correction and chip scale, while China builds a parallel programme

US chips 105 to 1,121 qubits, Chinese 105 to 504
The finding

The US leads on error correction: Google's 105-qubit Willow shows below-threshold correction and IBM's 120-qubit Nighthawk targets fault-tolerance by 2029. It also built the larger superconducting processor: IBM's 1,121-qubit Condor of December 2023, against the 504-qubit Xiaohong chip China announced in December 2024. China's programme runs alongside: USTC's 105-qubit Zuchongzhi 3.0 advantage claim and Origin Quantum's 180-qubit Wukong-180 cloud. Two parallel superconducting-QC programmes are maturing side by side, not converging.

Correction, 16 November 2026

an earlier version said China leads on scale and commercial deployment. IBM's 1,121-qubit Condor is larger than China's record 504-qubit Xiaohong chip.

Evidencesource β†—
ELEVATED

US users are the largest foreign users of China's leading quantum cloud

1M+ tasks - 192 countries

Origin Quantum's Wukong superconducting cloud has logged 30M+ visits and surpassed one million completed quantum tasks from users in more than 192 countries (Jun 2026) - and US users have consistently been the largest foreign cohort. The classical chip-export wall does not extend to quantum-cloud access, where the dependency currently runs the other way.

Evidencesource β†—
ELEVATED

AMD becomes the first credible second GPU supplier to a frontier lab

2GW + $5bn (AMD/Anthropic); initial 2027 Helios tranche over $20bn if fully procured
The finding

AMD's up-to-2GW MI450 deal with Anthropic, plus up to $5bn of equity, is the first credible second GPU source at the frontier-lab tier. It is a partial, forward-dated mitigant rather than a removal of the risk, and as of 4 August 2026 it is running against the tide: SpaceX, one of the largest compute buyers in this map, has committed to Nvidia exclusively. Diversification at one frontier lab is being offset by deliberate concentration at another buyer of comparable scale.

September

the second source is widening beyond AMD. Broadcom expects Anthropic to deploy 5GW of TPU v8i in 2027, with line of sight to a further 10GW in 2028, and OpenAI's Jalapeno accelerator on track for 1.3GW in 2027. The offset is still there: Anthropic's $44.6bn Nscale contract and OpenAI's Firmus sites run Nvidia Vera Rubin.

November

AMD's Data Center segment reported a record $8.5bn in the quarter to 26 September 2026, up 97%, with the Instinct MI450 Series launched and the first Helios racks running models and workloads at Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure (8-K, 3 November); AMD guided about $17bn of total revenue for the December quarter. Anthropic's S-1/A (2 November) sizes its AMD commitment for the first time: an initial tranche of Helios capacity in calendar 2027 expected to cost more than $20bn if fully procured, non-cancellable by Anthropic except for AMD's uncured breach, with an option for up to about the same again. AMD's 10-Q names OpenAI and Meta as holders of warrants for up to 320 million AMD shares.

Evidencesource β†—
ELEVATED

One company, SandboxAQ, sits in most US and allied post-quantum defence programmes

1 firm, 5+ US/allied programmes

SandboxAQ now sits at the single intersection of the US/allied post-quantum transition: DoW CIO cryptographic migration (AQtive Guard), DIU quantum navigation (AQNav), a $500M CHIPS materials award, NATO DIANA, and White House PQC EO implementation - while its own compute and capital come from Google and Nvidia, who are simultaneously its investors and its cloud/GPU suppliers. The same circular vendor-invests-in-customer pattern seen at the frontier labs now shapes post-quantum defense, and the US government has taken equity in exchange for the CHIPS award (government-as-shareholder).

Evidencesource β†—
ELEVATED

US AI computing capacity far exceeds China's

US ~75% vs China ~15% of GPU clusters

On raw AI compute the US commands roughly 75% of global GPU-cluster performance to China's ~15%, hosts 5,427 data centers, and is adding power at about 22% a year toward 90GW or more by 2030. China has reached near-parity on model quality, but its hardware stays a generation or two behind, gated by export controls and SMIC's ceiling versus TSMC. The gap is structural, not cyclical.

Evidencesource β†—
HIGH

Big Tech's AI capital spending is growing faster than the revenue behind it

~$730bn in 2026 ($720bn to $740bn guided), +78% YoY
The finding

Amazon, Alphabet, Meta and Microsoft together guide to roughly $725bn of AI capital spending in 2026, up about 77% from roughly $410bn in 2025. Alphabet alone spent $44.9bn in Q2, double the prior year, and guides $195-205bn for the year. Microsoft guides FY2027 to $255-260bn. Meta's free cash flow fell 91% in the quarter. Cloud revenue is accelerating with it (Google Cloud +82%, Azure +43%, AWS +37%), so the question is no longer whether demand exists. It is whether the spend can be funded at this rate without the circular structures above. On 23 August 2026 one hyperscaler answered the funding question in the open market rather than through the structures above. Alibaba priced HK$80bn ($10.2bn) of new stock at a discount, the largest primary follow-on ever by a Hong Kong-listed company and the world's third largest this year after Alphabet and Intel, with 100% of net proceeds committed to full-stack AI and roughly half of a three-year capex programme already spent. Demand reached about $28bn. The shares fell about 10%. Dilution is the honest route and it is also the one the market punishes on the day, which is a fair description of why the circular structures exist at all.

September

Oracle spent $28.5bn of capital in a single quarter with free cash flow negative $5bn by its own account, and the shares nearly erased a 7% jump after the print. The FT reported OpenAI projecting $278bn of cash burn over 2026 to 2030. CoreWeave raised $4.2bn of convertible notes on 22 September.

November

September 2026 quarter results, reported 21 to 30 October. Alphabet spent $51.4bn against $24.0bn a year earlier (8-K, 28 October) and kept its 2026 guide at $195bn to $205bn, with 2027 to increase significantly. Microsoft spent $45.2bn including $7.4bn of finance leases (call, 28 October), below its own guide of over $50bn because third-party capacity deliveries slipped, and now expects about $170bn for calendar 2026, down from about $175bn guided on 29 July; on that July call it said only that fiscal 2027 spending will grow year on year, without a figure. Amazon spent $75.9bn of cash capital expenditure against $34.2bn (10-Q, 30 October), kept about $220bn for 2026 and expects a meaningful increase in 2027. Meta spent $34.91bn including finance-lease principal against $19.37bn (21 October) and narrowed 2026 to $135bn to $145bn. The four spent about $207.4bn in the quarter ($51.4bn + $45.2bn + $75.9bn + $34.9bn), on definitions that differ. For 2026 they now guide about $720bn to $740bn (Amazon $220bn + Alphabet $195bn to $205bn + Microsoft $170bn + Meta $135bn to $145bn), a midpoint of about $730bn, about 78% above the roughly $410bn of 2025 ($730bn / $410bn = 1.78). Cloud revenue accelerated with it: Google Cloud +103%, Azure +48% (on a redefined metric), AWS +39%. More of the spending is now funded from outside. Free cash flow was negative at Alphabet ($9.1bn in the quarter), Meta ($2.12bn in the quarter) and Amazon ($42.4bn over twelve months), while Microsoft's was $18.6bn. Amazon's long-term debt at face value rose from $68.8bn at 31 December 2025 to $180.9bn at 30 September 2026 (10-Q). Meta announced a non-voting Class C stock on 21 October, which it said would help finance growth, and on 26 October priced a $25.0bn forward share offering and a $12.5bn mandatory convertible preferred. Oracle said on 28 October that its capital expenditure peaks in FY2027 and FY2028, that free cash flow reaches at least $30bn in FY2029, and that it will raise no more than $20bn of new capital through FY2030, excluding refinancing, after a $20bn equity raise in Q1 FY2027. Commitments not yet on balance sheets are larger again: Alphabet $924.4bn of purchase commitments plus $130.5bn of leases not yet commenced, Meta $454.52bn of non-cancellable commitments plus $355.73bn of leases not yet commenced, and Microsoft $364.6bn of leases not yet commenced. These are different measures and are not summed.

November

Google announced a long-term $17bn investment in its first Wyoming data centre on 5 November and Crusoe said the next day that Google had acquired its site there. Microsoft moved the other way on rented capacity in the week after lowering its calendar 2026 guide to about $170bn: press reports describe it winding down an Oracle rental, declining a Nscale expansion option and tightening oversight of neocloud projects.

November

the spending reached the debt markets and the neoclouds. Meta completed $15.0bn of US dollar notes on 5 November and EUR9.0bn of euro notes on 12 November, with Citigroup and Morgan Stanley as representatives of the underwriters. Nebius said on 12 November that it now expects about $27bn of 2026 capital expenditure, above its $20bn to $25bn range, and $40bn to $45bn in 2027.

Evidencesource β†—
HIGH

Chip companies lost a large share of their market value in the last week of July 2026

>$1tn off chip market caps, w/e 29 Jul 2026

More than $1tn came off the sector's chip names in the last week of July 2026. Nvidia lost $238bn, SK Hynix $176bn, Samsung $173bn and Micron $113bn; AMD and TSMC each shed over $100bn, and the memory names entered a bear market. The stated cause was not weakening demand but a repricing of expectations, on concern that AI infrastructure spending is peaking faster than expected. This is the first market-priced test of the structure this map describes, and it fell along exactly the lines the map predicts: the chokepoint tier took the worst of it. The follow-through answered the open question: the repricing reversed rather than extended. Through 21 August the SOX was up more than 10% for the month, on pace for its best August in over two decades, which reads as volatility around an intact spending thesis rather than the start of a derating.

Evidencesource β†—
HIGH

Nvidia has moved from investing in its customers to guaranteeing their debts

up to $250bn discussed; $105bn guaranteed 17 Aug; $108.5bn Nvidia maximum guarantee exposure per the 10-Q; about $43bn Broadcom residual value support per its 8-K of 27 Oct
The finding

In late July 2026 Nvidia and OpenAI were reported to be in talks over a backstop of up to $250bn, letting OpenAI raise debt on the strength of Nvidia's credit to fund a 10GW campus in Pike County, Ohio. On 17 August the first piece was signed: Nvidia agreed to guarantee up to $105bn of OpenAI's conditional lease and power obligations to SB Energy at that campus, put $1.5bn of equity into SB Energy itself, and took the exclusive chip-supply position on the site. Almost every earlier loop was equity: a vendor buying a stake in its own customer. Nvidia's 10-Q shows the exception was already on its books, $3.5bn of land, power and shell guarantees for AI cloud partners, one of them up to $860.3m of Nscale's Texas lease from October 2025. The Ohio guarantee is thirty times that, and a guarantee of that size is different in kind. It puts the supplier's balance sheet behind the customer's borrowing, so a failure that was once a lost investment becomes a contingent liability, and the sector's largest market cap becomes the sector's de facto credit desk.

October

two more chip vendors now stand behind their customers' obligations, beside Nvidia. GAP FILL, dated 1 October: Reuters, reading Anthropic's confidential prospectus, reported that Broadcom has agreed to lend Anthropic up to $42bn, in notes that could convert into Anthropic shares, to finance about a third of a $125.2bn five-year TPU lease commitment, and Anthropic names Broadcom's double role as supplier and lender as a potential conflict of interest. Unlike Nvidia's guarantee to SB Energy, this is direct credit to the lab, with an equity conversion attached. On 13 October AMD signed a credit support agreement backing the 222MW phase-one lease at Fermi America's Project Matador campus in Texas, with a framework for up to 650MW more.

November

Broadcom now backstops a lab's hardware financing. On 26 October Bank of America, Blackstone, Citi and Morgan Stanley announced a $63bn transaction in which a new SPV buys Broadcom XPU accelerators and networking for lease to Anthropic, with deliveries from 2027. Broadcom provides residual value support equal to the first-lien portion, and its 8-K puts its maximum potential liability at about $43bn, undiscounted, once deployed. The record cannot yet show whether this replaces or sits beside the up to $42bn note facility reported on 1 October, which said Broadcom may designate a financing partner.

Evidencesource β†—
ELEVATED

An AI agent security alliance formed in July 2026, and Anthropic is not part of it

37 members; Anthropic on neither list
The finding

On 27 July 2026 Nvidia and 36 others formed the Open Secure AI Alliance and open-sourced NOOA, a framework for making agent behaviour testable and auditable. Microsoft, IBM, Red Hat, Cloudflare, CrowdStrike, Palantir, Databricks, Hugging Face and the Linux Foundation joined. OpenAI, Google and Meta appear among the letter's signatories but are absent from the inaugural membership; Anthropic appears on neither. The absence is the finding. The firms whose models create the auditing problem are not in the body being formed to solve it, and the Cloud Security Alliance judges the result a standards body without a charter.

Against the reading, 29 September

the frontier labs are inside the next instrument. Anthropic, OpenAI, Google, Meta, Nvidia and xAI signed the White House accord, the Joint Commitment on Frontier Responsibilities, committing to internal controls, an internal review team, an independent external auditor and a board committee. It carries no penalty and creates no regulator, so it settles who signs, not who checks.

Mid October, against the reading

the accord's auditor pledge has a first implementation. On 15 October Anthropic said six METR evaluators now work inside it with access roughly on par with its own risk teams, publishing their own assessments without an Anthropic veto, with SecureBio and Accenture also evaluating and Hugging Face, an alliance founding member, in early discussions. That begins to answer who checks, for one lab, by its own choice.

Evidencesource β†—
ELEVATED

OpenAI and Anthropic no longer expect to list in the same quarter

Anthropic: roadshow launched 10 Nov, 825m shares at $70 to $78 (Nasdaq: ANT), $26.0bn concurrent placement, Q3 revenue $17.0bn filed; OpenAI: not in 2026, $30bn private at ~$1.4tn, annualised revenue approaching $50bn
The finding

This finding originally recorded both frontier labs filing into Q4 2026, and the record has since moved against it: by mid-August OpenAI's listing had slipped, its CFO telling staff the company would be public in 2027, and a $7bn employee buyback at a flat valuation removed the near-term pressure to list. Anthropic still points at October, with a $65bn annualised run-rate at end-July and backers (not the company) talking a roughly $2tn debut through Morgan Stanley, Goldman Sachs and JPMorgan. The concentration risk the original finding described now lands on a single name in a single month: one loss-making lab converting privately negotiated marks into daily public ones, with the second lab's postponement itself a data point on how hard the first pricing is.

September

Anthropic's timetable slipped within the month rather than broke. Reuters reported on 4 September that the public prospectus had moved to late September and marketing to mid-October at the earliest; on 11 and 13 September it reported up to $100bn sought at about $2tn on Nasdaq, with Nvidia in talks to anchor the offering with up to $10bn. No public S-1 was on EDGAR by 26 September. Altman told Fortune on 12 September that OpenAI will not list in 2026, and the FT reported early talks on a private round at about $1.2tn.

Late September

Reuters read Anthropic's confidential prospectus on 28-29 September: 2025 revenue of nearly $4.6bn, a net loss of about $42bn of which roughly $34bn is an accounting charge, at least $518bn of infrastructure commitments, and a listing now likely after the November midterms rather than before them. EDGAR still showed no public S-1 on 1 October. OpenAI is reported to be seeking at least $30bn privately at about $1.4tn, up from the $1.2tn talks of mid-September.

October

EDGAR full-text search for 'Anthropic PBC' across all forms from 28 September to 7 October returned three unrelated fund and issuer filings and no S-1. On the OpenAI side, SoftBank completed its $30bn follow-on on 1 October, taking its cumulative investment to $64.6bn for about 13%.

Mid October

Anthropic held its pre-IPO investor day on 14 October and told prospective investors, per people familiar with the matter, that third-quarter revenue was $16.8bn, up from about $11.6bn and $4.7bn in the first two quarters, profitable on an adjusted basis and cash-flow positive at $1.70 of revenue per dollar of compute (FT). The company has not published the figures. Bloomberg reports formal marketing could start the week of 9 November with trading before Thanksgiving on 26 November, and some prospective investors put fair value at $1.8tn to $2tn. EDGAR full-text search for 'Anthropic PBC' from 1 to 17 October returned seven fund and issuer filings and no S-1. On the OpenAI side, the FT reported on 8 October that OpenAI told investors its annualised revenue is approaching $50bn, about $20bn below the figure its own investors had built to compare with Anthropic, which counts cloud-partner sales and OpenAI does not; on 14 October its compute lead Sachin Katti said he would leave on 15 November.

Later, 15 October

Reuters, citing several people familiar with the matter, reported that current and prospective investors who attended the 14 October investor day discussed a valuation of about $1.5tn, down from earlier expectations of $2tn (as carried by Yahoo Finance and by Proactive).

November

Anthropic's amended S-1 of 2 November 2026 puts its SpaceXAI compute commitments at up to $120.5bn into 2029, about $3.1bn a month once all capacity is delivered in December 2026, cancellable on 90 days' notice after initial periods ending in January and March 2027. On the SpaceX side, both SpaceX and Tesla have special committees evaluating a combination (10-Qs, 22 October 2026), with no terms agreed, and Tesla has held SpaceX shares since its $2bn xAI investment completed on 12 March 2026, after its xAI right converted at the February merger.

November, 10 November

Anthropic launched its IPO roadshow and filed an amended S-1: 825m Class A shares at an expected $70 to $78, Nasdaq symbol ANT, plus a $26.0bn concurrent private placement of non-voting Class N stock to seven strategic buyers, for net proceeds of about $84.6bn at the $74.00 midpoint. The filing puts third-quarter revenue at $17.0bn, against the $16.8bn reported from the 14 October investor day, and Run-Rate Revenue at about $81bn on 31 October. The registration statement was not yet effective on 10 November.

Evidencesource β†—
ELEVATED

US export rules on advanced AI chips to China were relaxed in January 2026

H200 case-by-case since 15 Jan 2026
The finding

The common reading is that Washington keeps tightening. The record says the opposite. BIS rule 2026-00789, effective 15 January 2026, moved the Nvidia H200 and AMD MI325X from presumption of denial to case-by-case review, subject to a 25% tariff and third-party testing, and roughly ten Chinese firms including Alibaba, Tencent and ByteDance were cleared to buy. At the same time Huawei is shipping the Ascend 950PR with a CUDA-compatible stack, ByteDance has committed around $5.6bn of orders, and Huawei said its chips were used in some of DeepSeek V4's training. China is not choosing between the two stacks. It is buying the American one while the domestic one matures, which weakens the leverage the embargo was supposed to create.

November

the loosening held. On 5 November 2026 BIS extended to 10 January 2027 its suspension of the rule that would extend Entity List restrictions to affiliates 50% or more owned by listed parties (FR Doc. 2026-22793). The New York Times reported on 7 November that about two dozen Chinese firms are now cleared to buy H200s, including Baidu and Tesla's Shanghai unit, up from about 10 in May.

November

both sides extended their truce measures to the same date. China's Ministry of Commerce extended on 9 November 2026 the suspension of its October 2025 rare earth, diamond and battery export controls to 10 January 2027, the date to which BIS had extended its Affiliates Rule suspension. In Congress the direction is the other way: the Senate version of the FY2027 NDAA carries three chip export bills, and House Foreign Affairs chairman Brian Mast said on 9 November that the White House had added provisions to one of them, the AI Overwatch Act, aimed at Chinese access to US chips through overseas data centres.

November, 12 November

the pressure in Washington ran toward tightening. The Information reported that Commerce is preparing to finalise a rule extending chip controls to seven more countries, including Malaysia, Singapore and Thailand, against remote use of advanced chips by Chinese AI companies; no text had been seen by 15 November. Senate Majority Leader John Thune put a 'high probability' on more chip-export restrictions, with no NDAA deal yet on the three Senate bills. Ron Wyden and Richard Neal opened a minority inquiry into the H200 approvals and the tariff on them, and Tom Cotton asked Commerce to add CXMT to the Entity List. Malaysia, a host of overseas capacity, said the same day it will require data-centre operators by law to disclose GPU imports and power use, and is reviewing re-exports of US-restricted chips.

Correction, 16 November 2026

an earlier version gave ByteDance's orders as .6bn rather than $5.6bn and said DeepSeek V4 trains end to end on Ascend silicon. Huawei said its chips were used in some of V4's training.

Evidencesource β†—
ELEVATED

SpaceX plans AI computing in orbit, built on Nvidia hardware

up to 1m satellites, first units 2027
The finding

SpaceX intends to deploy Nvidia Vera Rubin NVL723 racks both on the ground and in space, and is co-designing the Starmind AI1 satellite payload with Nvidia, each unit carrying Rubin GPUs and Vera CPUs. SpaceX asked the FCC in January 2026 to authorise up to one million Starmind satellites; first test units are planned for early 2027. Three consequences. The launch monopoly and the compute build sit inside one company, so the cost of orbiting a data centre is set by the firm that owns both. Orbital compute sits outside every jurisdiction in the policy layer here, all of which is written for hardware on somebody's soil. And the authorising body is not an AI regulator or an export-control agency but the FCC, a spectrum and satellite licensor, which is now a chokepoint on frontier compute by accident of jurisdiction.

November

the ground build met political limits in Texas, where SpaceX is building compute capacity and the governor-elect called for a moratorium on new data-centre approvals; Deutsche Bank wrote that SpaceX's Texas compute could face delays (CNBC, 5 November 2026). Musk wrote on X after the results that "Data centers in space looking pretty appealing right now" (CNBC, 4 November). An Apollo-managed lessor committed on 24 October to about $40bn of GPU pod leases for SpaceX's ground data centres.

November

the ground build continues in Texas. A City of Starbase notice of 9 November 2026 sets out about $24bn for the initial phases of a SpaceX gas-fired power plant and super-compute facility near Brownsville, with a tax-incentive hearing on 9 December.

Evidencesource β†—
ELEVATED

A promised self-driving milestone has been restated four times and is still not met

4 restatements, one milestone, still open
The finding

Unsupervised autonomy has been promised on three separate dates now held in this map: by June 2025 (said January 2025), Q4 2026 at the earliest (said April 2026, citing complex intersections, road markings and weather), and widespread in the US by year-end (said May 2026, under a month after the concession and with no reported resolution of those obstacles). What exists today is fewer than 40 unsupervised robotaxis across three Texas cities. None of those statements was deleted when the next one arrived, which is the point: the graph can show the gap between what was said, when, and what was delivered, without anyone having to remember it.

November

the milestone moved a fourth time. On Tesla's Q3 2026 call on 21 October, Ashok Elluswamy, vice president of AI software, said the first customer-owned cars could join the robotaxi fleet unsupervised around Q2 2027, after Musk had said in April that the unsupervised version would probably come in Q4 2026 and in May that it would be widespread in the US by year-end. The robotaxi service itself grew: Tesla reported nearly two million cumulative unsupervised miles at 30 September 2026, up eightfold in the quarter, without stating its unsupervised fleet size. Waymo, which already runs fully autonomous service, said on 27 October it was adding nearly 250 square miles across five US cities.

Evidencesource β†—
ELEVATED

Japan funds its own foundation model but runs it on Nvidia hardware

Y1tn programme, 100% Nvidia silicon
The finding

Japan's FRONTia programme is a serious attempt at model sovereignty: up to Y1tn over five years for a domestically developed multimodal foundation model, run by a consortium of Sony, SoftBank, NEC and Honda with 44 investing organisations and PLaMo engineers seconded in from Preferred Networks. The compute underneath it is not sovereign at all. The 140MW facility is 13,750 Nvidia Vera CPUs and 27,500 Nvidia Rubin GPUs on Nvidia fabric, described by Nvidia itself as the world's first national AI infrastructure. This is the same pattern the map records at SpaceX: independence pursued at the model layer while dependence deepens at the silicon layer. It also dates the exposure precisely, because the facility does not begin operating until June 2028.

October

JERA, Dell Technologies and RHAELM signed a memorandum of understanding on 1 October for a standard model of AI infrastructure sited at Japanese power stations. The first project, at JERA's Chiba Thermal Power Station, is up to 400MW behind the meter, with total capital deployment expected to exceed $15bn across all phases, operations targeted around 2028, and Apollo intending to invest in and finance the developer. The compute layer is Dell's rack-scale AI Factory; the release names no chip supplier.

Evidencesource β†—
HIGH

SpaceX's computing target grows faster than the memory supply that would serve it

memory +20%/yr vs compute demand +200%
The finding

SpaceX's move from about 2GW of compute at end-2026 to roughly 10GW at end-2027 implies well over a million Rubin GPUs and possibly more than two million. BNP Paribas has warned the commitment could trigger another Nvidia supply shortage. SpaceX itself put the constraint plainly: memory supply is growing at about 20 per cent a year while demand for AI compute rises more than 200 per cent, a gap that on its own account persists well into 2027. That points the risk at the HBM triopoly rather than at Nvidia, which is the tier this map already flags as a chokepoint and the tier that fell hardest in the July repricing. An exclusive commitment by one very large buyer does not create capacity; it reallocates a queue, and everyone else in the graph is standing in it.

September

scarcity reached prices. Reuters reported Chinese AI chip prices up 20-50% in two months on an HBM shortage, Huawei's Ascend 950PR up about 30%, and CoreWeave's filed presentation shows about a 70% price increase across SKUs in July 2026.

Late September

Micron, the third HBM supplier, reported $54.23bn for its fiscal fourth quarter against its own $50bn guide and said the vast majority of its calendar 2027 HBM supply is already contracted at significant price increases, that memory will be much tighter in 2027 and 2028 than in 2026, and that it has no line of sight to when supply and demand return to balance.

November

The other two HBM makers now describe the shortage Micron described on 30 September. SK Hynix said on 28 October that memory demand will exceed supply through 2030, because leading-edge processes are hard and new fabs take long to build, and that long-term agreements with 14 major customers already cover more than 50% of its 2027 and 2028 capacity. Samsung said on 29 October that AI-driven demand in 2027 will far exceed industry supply in comparison with 2026. All three HBM makers have now said in their latest results that memory stays short at least through 2027.

Evidencesource β†—
HIGH

The wars in Ukraine and Iran have become a major source of demand for military AI

$98bn Ukraine, 20bn tokens/day
The finding

Demand for military AI is no longer a forecast. Ukraine raised 2026 defence spending to a record UAH4.4tn, about $98bn, and contracted roughly $8bn of UAVs in the first half alone, at double the prior year and with nine-day delivery for in-stock systems. The Iran campaign supplied the consumption figure: Pentagon model usage reached about 20 billion tokens a day during Operation Epic Fury, an increase of 4,425 per cent, with Palantir's Maven Smart System used to plan and coordinate roughly 13,000 strikes, and the Army later reinstating token limits after allowances were exhausted early. Two consequences for this graph. Defence has become a load-bearing customer for commercial AI infrastructure rather than an adjacent market. And Ukraine is functioning as a live testbed whose procurement now sets drone requirements from battlefield data rather than committee.

October

At AUSA on 14 October Hegseth said the Pentagon is seeking more than $50bn for autonomy and robotics through a reconciliation bill not yet approved, and DAWG is being folded into a new Task Force Autonomous Warfare inside SOCOM.

November

contract flow kept moving into AI command systems. The Marine Corps opened Palantir's Maven Smart System to every command and echelon, reserves included, under enterprise licensing effective 1 October (reported 20 and 21 October), and Defense Department officials say Maven was used to identify and bomb more than 2,000 distinct targets in Iran in the first 96 hours of Operation Epic Fury. A gap fill: on 30 September the Army awarded Anduril up to $1.8bn over five years, $162.8m at the outset, to field its NGC2 command software to I Corps. On 23 October the Army picked AWS and World Wide Technology for a combined $2.6m agentic AI cyber defence pilot. The war also showed the dependence: Zelensky said on 22 October that Ukraine's SpaceX Starshield service was still restricted, limiting the interception of jet-powered drones, and who ordered the limit remains disputed.

Evidencesource β†—
HIGH

The closure of the Strait of Hormuz cut supplies of helium and sulfur that chipmaking needs

~30% of helium, Gulf sulfur, closed Feb 2026
The finding

This map's chokepoints have all been technological: one lithography vendor, one leading-edge foundry, three memory makers. The 2026 Iran conflict exposed one that is purely geographic. The Strait of Hormuz was effectively closed from 28 February 2026, and it carries the Gulf's sulfur exports and roughly a third of the world's helium, most of it Qatari and produced alongside LNG at Ras Laffan. Helium has no substitute in chip fabrication: it is the carrier gas, the cleanroom purge gas and the wafer coolant. Ultra-pure helium prices doubled, and memory fab output begins declining within one to two weeks of rationing. Sulfuric acid, derived from the same route's sulfur, is essential to refining the metals that feed semiconductors, batteries and magnets, and China separately restricted sulfuric-acid exports. The significance is the ordering: a shipping lane can throttle ASML's customers without anyone touching a fab, an export licence or a chip.

September

the strait was still closed at about seven months on 25 September, with Iran's seven-day reopening proposal unresolved.

Late September

no reopening was agreed by 30 September. Trump rejected Iran's seven-day plan, and Iran said it had received a US response through Qatar, the two sides divided on the order of the steps.

November

still closed. Iran made seven demands, including lifting sanctions and ending the US naval blockade of its ports, a condition for reopening, and its security chief said in Moscow on 5 November 2026 that the strait will not open until they are met (Al Jazeera, 6 November).

Evidencesource β†—
CRITICAL

ASML, the sole maker of advanced lithography machines, is exposed to four separate risks

sole source, 4 exposures, all indirect
The finding

ASML remains the sole maker of EUV and High-NA lithography, and Intel's first high-volume High-NA production on Panther Lake extends that monopoly into the next node rather than ending it. The July 2026 capacity announcement, 30% more Low-NA EUV and DUV immersion for 2027 with a further 30% under evaluation for 2028, was read as easing the bottleneck, and on volume it does. On concentration it does the opposite: more of the world's leading-edge capacity now depends on one company's ability to execute an expansion. Four distinct risks in this graph converge on that single point, and three of them reach it indirectly, which is why they are easy to miss. Demand: Nvidia's dependencies (52 when counted in August 2026) and the SpaceX exclusivity all resolve to silicon only ASML's machines can print. Geography: ASML's largest customer concentrates leading-edge capacity in the Taiwan Strait. Materials: the Hormuz closure constrains the helium every customer fab needs, so a shipping lane can idle ASML tools without touching ASML. Policy: about 20% of 2026 revenue is China DUV, which the proposed MATCH Act would restrict further. None of these is a claim about ASML's business. They are paths through it.

October

On 14 October 2026 ASML reported Q3 net sales of EUR11.7bn and guided 2026 to EUR44bn to EUR45bn. China rose to 23% of Q3 system sales from 14% in Q2, and ASML expects about 20% of 2026 sales from China. ASML said its supply as planned for 2027 and 2028 will not fully meet customer demand.

Evidencesource β†—
HIGH

Companies backed by funds associated with Donald Trump Jr. and Eric Trump have won federal contracts

$3.2bn awarded; $290m donated; now a direct holding
The finding

Funds associated with Donald Trump Jr. and Eric Trump hold positions in at least 15 companies seeking federal business, and the Washington Post reported on 13 July 2026 that those companies have secured at least $3.2bn in federal contracts since the investments were made, with a further $3.1bn in options. 1789 Capital, where Donald Trump Jr. became a partner within days of the 2024 election, accounts for 11 of the 15. The portfolio maps almost exactly onto the sectors this graph already flags as receiving elevated funding: defence autonomy, drones, space, advanced manufacturing, quantum and rare-earth magnets. The clearest single case is Vulcan Elements, backed at a valuation near $200m in autumn 2025 and awarded a $620m Defense Department loan commitment a few months later. Vulcan Elements is also in the federal equity portfolio, so a single company carries a 1789 Capital position, a $620m Defense Department loan commitment and a US government shareholding at once. On 5 August 2026 lawmakers called for a Pentagon probe. What the graph records is association and sequence: who invested in what, when, and what federal business followed. It does not assert causation, and no source cited here does either. On the President's own account the structure is worth stating precisely, because it is often described loosely: the securities are directly and beneficially owned, held in a revocable family trust whose trustees are family members, not a qualified blind trust, and Q1 2026 alone records 3,642 transactions worth between $220m and $750m. One donor edge deserves separating out because it runs through an entity already central to this graph. Elon Musk spent more than $290m on the 2024 cycle, $193m of it through America PAC, the largest identified flow from any entity in this map into electoral politics; he is also principal of SpaceX, which the graph records as a major federal contractor and as newly exclusive to Nvidia. Timothy Mellon at $151.5m to MAGA Inc and Miriam Adelson at about $105.8m across vehicles were larger or comparable donors but sit outside this sector. The systemic point is narrower and still material: the demand signal this map treats as policy-driven now has a documented overlap with the private interests of the family directing the policy, and that overlap is itself under formal review. On 22 August 2026 that overlap acquired a direct line. A financial disclosure signed on 12 August records the President buying between $15,001 and $50,000 of SpaceX on 23 June, eleven days after its IPO and among more than 1,000 trades that month. The graph already held presidential positions in Nvidia, Palantir and Apple; what is different here is the counterparty. SpaceX in this map is about $12bn of Space Force work, a classified NRO satellite programme, Pentagon AI-compute talks and the single largest identified political donor in the sector as its controlling shareholder, and the administration has since directed an increase in US commercial launches. The White House states that third-party institutions manage the portfolio against recognised indexes and that no family member can direct it, which speaks to who executes rather than to who owns. The sums are trivial beside the contracts; the edge is worth drawing because it closes a loop the map was previously carrying as two unconnected halves. OCTOBER, GAP FILL dated 6 October: the Navy awarded Anduril, a 1789 Capital portfolio company, a contract of up to $2.9bn for Virginia-class submarine components, with payments tied to demonstrated production, announced by President Trump at the site of Anduril's planned $3.7bn Arsenal-2 shipyard in Maryland (Reuters).

November

Musk was the largest individual donor of the 2026 cycle, with at least $474m; America PAC spent nearly $87m across 44 federal races, 93% on races Republicans lost (CNBC, 4 November 2026). Democrats take control of Congress in January, with the power to investigate federal contracts; the executive keeps the veto.

Evidencesource β†—
HIGH

The US government now owns shares in chip companies and takes a share of some chip sales

~35 equity positions; 15% of China chip revenue
The finding

Two instruments changed the relationship between the US government and this sector, and the map previously carried only a fragment of one. As shareholder: direct federal equity now spans a reported thirty-odd companies, including about 10% of Intel for $8.9bn, which makes the government Intel's largest single shareholder, 15% of MP Materials, 10% of Trilogy Metals and 5% of Lithium Americas. Most of the portfolio is critical minerals, the layer sitting upstream of everything else here. As rent-taker: since August 2025 Nvidia and AMD have paid the Commerce Department 15% of their revenue from specified China AI chip sales in exchange for export licences, an arrangement in which the licensing authority holds a direct financial interest in the volume it licenses. Both instruments are under congressional scrutiny, with a February 2026 oversight letter on the minerals deals. The consequence for reading this graph is that the United States can no longer be modelled purely as a regulator of the AI supply chain. On several nodes it is also a counterparty, and on Intel it is the largest owner.

October

Commerce will receive about 4% more of GlobalFoundries under an agreement of 14 October (company release); Yahoo Finance reported on 15 October that earlier stakes came to 3% to 4%, so the holding approaches 8%, and that the federal equity portfolio now spans 35 companies. On 7 October Wolfspeed said the Department of War had issued a conditional commitment for a 30-year loan of up to $1.5bn carrying warrants for up to 7.5% of its equity, issued as tranches fund; definitive agreements are not yet signed.

Evidencesource β†—
ELEVATED

David Sacks kept a policy role after his term as a special government employee ended

130-day cap, role continues
The finding

David Sacks served as White House AI and Crypto Czar from January 2025 to 26 March 2026 as a special government employee, a designation capped at 130 days of service in any twelve months and carrying lighter disclosure and conflict obligations than a permanent appointment. When the clock expired he moved to co-chair the President's Council of Advisors on Science and Technology alongside Michael Kratsios, describing the new seat as widening his remit from AI to technology generally. The structural observation is not about any individual. It is that the influence attached to the person rather than to the office, so the statutory limit on the office did not limit the influence. Any map that models policy exposure by tracking formal titles will misread this, which is why the graph records the role with an end date and the successor role opening where it closes.

October

the czar role returned in a different form. On 4 October President Trump named Jay Clayton, the Senate-confirmed Director of National Intelligence, as AI czar leading the new Super Intelligence Force (NPR). The post now sits with a confirmed official who holds a second office, rather than with a special government employee on a 130-day clock.

October

TNW reported on 16 October, citing Politico, that David Sacks will advise the Super Intelligence Force alongside Condoleezza Rice, an advisory role outside the force's four named leaders.

Evidencesource β†—
ELEVATED

Stargate's progress has lagged the programme announced in January 2025

$500bn announced, partner disputes reported
The finding

Stargate was announced at the White House in January 2025 as roughly $500bn and 10GW over four years, with OpenAI, SoftBank and Oracle, and MGX funding an initial $100bn tranche. It remains the administration's signature compute-infrastructure signal. Execution has trailed the framing: more than a year on, reporting describes unresolved disputes between the partners over structure and responsibility, and OpenAI fell short of locking 10GW of commitments through SoftBank and Oracle, leaning instead on cloud contracts with AWS, Google Cloud, AMD and Cerebras. Further sites have also been announced, so expansion and shortfall are both true. The systemic point is about signal quality rather than about any partner: a headline number announced from a podium entered this graph, and the market, as a commitment. It is a framing. Oracle's $553bn contracted backlog is a commitment. The two should not be read the same way, and until now this map did not clearly distinguish them.

September

Oracle sent a force majeure notice on the 2.45GW Project Jupiter campus in New Mexico on 24 September, and Nscale's S-1 records that OpenAI withdrew from Stargate Norway and Stargate UK in April 2026. The evidence runs both ways: GPT-6 Astra was pretrained on more than 100,000 GPUs at the Texas site, and New Mexico's Supreme Court denied the petitions against Jupiter's permits on 17 September.

October

Bloomberg reported, via DCD on 13 and 15 October 2026, that Oracle has trucked compressed gas to its Salt Lake City site and to its Shackelford County, Texas campus built with OpenAI, and has asked FERC to expedite the Green Chile pipeline for its 2.5GW Project Jupiter campus in New Mexico, whose pipeline timeline has slipped to February 2027 from August. Power supply, not only financing, is now delaying sites in the build-out.

Evidencesource β†—
ELEVATED

Reported totals of the crypto income in the same 2025 disclosure differ widely

$580m to $1.4bn, same filing

The 2025 annual financial disclosure, 927 pages published on 30 June 2026, reports crypto-related income that outlets total variously at more than $580m, at $1.2bn and at $1.4bn from the same document. The spread comes from which lines are counted and how token proceeds are valued, not from disagreement about the filing. Two lines are firm: more than $635m of royalties from the $TRUMP memecoin, and more than $500m associated with World Liberty Financial, where token-sale proceeds are themselves given in a $236m to $520m range plus about $65m from an equity sale. A figure of roughly $799m for World Liberty circulates and is not supported by the reporting reviewed here. This sits at the edge of an AI-sector map rather than inside it, and is recorded for one reason: the political-capital finding concerns the scale and structure of family financial interests running alongside federal policy, and that scale cannot be assessed from the defence-tech positions alone.

Evidencesource β†—
ELEVATED

One researcher, now in Shenzhen, created both a leading AI agent method and its main benchmark

1 author; 2 artefacts; 3 frontier labs measured on them

Yao Shunyu authored ReAct in October 2022 and co-created SWE-bench a year later. The first is the loop most production agent frameworks implement; the second became the number OpenAI, Anthropic and Google all quoted for coding agents. He took the doctorate at Princeton under Karthik Narasimhan, did the ReAct work at Google, joined OpenAI in August 2024, and has been Tencent's chief AI scientist since the 17 December 2025 announcement, reporting to president Martin Lau and heading the group's LLM and AI infrastructure departments against a stated plan to roughly double AI spend to over RMB 36bn in 2026. Two things are visible here and neither is an inference about intent. Method and evaluation in this sector are unusually concentrated in individual authorship rather than institutional ownership, and that authorship moves with the person across an export-control boundary the rest of this map treats as the main dividing line. The yardstick is also eroding independently: OpenAI retired SWE-bench Verified on 23 February 2026 after finding frontier models could reproduce its solutions verbatim, so the shared number the labs compared themselves on no longer means what it did.

Evidencesource β†—
ELEVATED

Nearly every important open-weight model is distributed through Hugging Face

one hub, 3m+ models; definitive agreement signed 2 Sep, close expected H1 2027
The finding

Open weights are open at the licence and narrow at the door. Nearly every open model that matters, Llama, Qwen, Kimi K2, Nemotron, DeepSeek, is distributed through Hugging Face: over 3m models, 1m datasets and 13m users on one platform. The graph has recorded that as a single point of distribution since it was built. On 26-27 August 2026 the identity of the owner changed in prospect: Nvidia was reported to have agreed to acquire the hub for $12.9bn, and on 2 September signed a definitive agreement: about $11.9bn plus up to $1.0bn of retention equity, closing expected in the first half of 2027 subject to regulatory approval. Openness of the weights is unaffected. Ownership of the channel is not. The company that trains Nemotron, licenses Poolside's model factory and is in talks to fund Perplexity would also own the shelf every rival open model sits on, and the compute-rental outlet attached to it. That is a different kind of concentration from silicon, and it is now on the same node.

Mid October

the hub held, and the hyperscalers opened side doors. Reflection's Beam weights went out on Hugging Face on 16 October. Meta's Muse Spark Open weights went out on Hugging Face and, the same day, on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, so the largest US open-weight release of the month reached users through three cloud catalogues as well as the hub Nvidia has signed to buy. Anthropic said on 15 October it is in early discussions with Hugging Face to act as an independent evaluator of its models.

November

Reuters reported on 28 October 2026, citing two people familiar with the matter, that the Justice Department is asking how Nvidia's purchase of Hugging Face would affect AI competition, including whether it would let Nvidia capture a fundamental distribution point for AI. The DOJ declined to comment, no action or second request has been reported, and Nvidia expects to close in H1 2027.

November

the hub held again. Moonshot published Kimi K3.1 Pro's weights on Hugging Face on 5 November 2026 under its own licence (moonshotai/Kimi-K3.1-Pro).

Evidencesource β†—
ELEVATED

Nvidia is building open-weight models of its own

$6bn licence, 100+ engineers, weights open
The finding

Nvidia is training Nemotron 4, an open-weight family whose largest model targets at least a trillion parameters, reported by The Information on 11 August 2026 with no release date confirmed. Nine days later it bought the means to finish it. Nvidia is paying Poolside $6bn to license Model Factory, the internal system Poolside developed its own models on, and more than 100 Poolside engineers move onto Nemotron. The licence is non-exclusive and the founders are not going, which is why the shareholder letter can say it is neither an acquisition nor an acquihire; the effect on this graph is the same either way. The move reads as competitive against the closed labs Nvidia also supplies, but the graph shows the opposite effect on concentration. The weights would be open while the inference stack they are tuned for, TensorRT-LLM and Triton, is not, so the cheapest way to run the best open Western model still routes through the vendor that carried 52 mapped dependencies when counted in August 2026. Scale is not the differentiator either: Moonshot's Kimi K2 is already open at 1 trillion total parameters activating about 32 billion per token. What moves is who owns the machine, and every route still lands on the same silicon.

October

two Western open-weight models were announced on consecutive days, and both were built on the same vendor. Reflection, which Nvidia backs, described Beam (501B parameters) with a reinforcement learning run on 10.5K GB300 GPUs; Mistral previewed Mistral Large 4 (1 trillion parameters) trained on 3,800 Grace Blackwell GPUs. Neither had released weights by 7 October; both promise them by the end of the month.

Mid October

the Western open-weight field split on the question of the frontier. Reflection, which Nvidia backs, released Beam's weights under Apache 2.0 on 16 October. Meta released the weights of Muse Spark Open (816B parameters, 42B active) under its own Muse Model License on 15 October, three days after launching its most capable model, Muse Wonder, closed; its chief AI officer Alexandr Wang said Meta 'will generally not open-source our most powerful models'. The company that set the American open-weight standard with Llama now opens a lesser model and keeps its best one metered. Mistral Large 4's weights remain due by the end of October. The scale point in this finding has also moved: Moonshot's Kimi K3, open since July, is 2.8T parameters with 104B active, nearly three times Kimi K2.

Gap fill, 15 July

Thinking Machines Lab, in which Nvidia invested in March, released Inkling (975B parameters, 41B active) with full weights on Hugging Face, trained on Nvidia GB300 NVL72. Reflection, Meta and Thinking Machines have now shipped open weights at the 500B to 1T scale and Mistral has promised them, and each of those whose training hardware is stated (Reflection, Mistral, Thinking Machines) trained on Nvidia.

Evidencesource β†—
ELEVATED

Cloud companies' nuclear agreements cover only part of their demand, and arrive late

up to 13GW in announced agreements, under 20% of demand to 2035; Amazon's X-energy options are outside it
The finding

The hyperscalers' announced nuclear agreements could provide up to 13GW, split roughly evenly between power purchase agreements (PPAs, 6.9GW) and direct partnerships with developers of new reactors (6.1GW), on Carnegie's June 2026 count. Meta leads by volume at up to 7.7GW across Constellation, Vistra, TerraPower and Oklo. Amazon holds 1.9GW at Susquehanna and up to 0.3GW from the first X-energy stage at Energy Northwest. Microsoft has 835MW at Three Mile Island, and Alphabet has 500MW with Kairos and 1.8GW with Elementl. Beyond that first stage, Amazon's options with X-energy, on more than 5GW of projects by 2039, sit outside the 13GW. Carnegie's June 2026 assessment puts the combined output at roughly 102 TWh a year, which is less than 20% of projected demand through 2035. The dates are the problem rather than the intent: the PPAs on existing plants deliver from 2027, while most new build lands 2034 or later, against demand the DOE study has rising from 176 TWh in 2023 to between 325 and 580 TWh by 2028.

October

the next contracts bought uprates rather than new build, which brings part of the supply forward. Constellation signed Amazon on 30 September for 690MW at Calvert Cliffs including a 190MW uprate due 2030 to 2032, and Google on 6 October for 890MW of new capacity from uprates at 11 units, the first by 2028, plus a 15-year agreement for 2,700MW of existing PJM output. The same day Black Hills filed agreements to serve Google's planned Cheyenne data centre with a 2.7GW resource mix that includes 564MW of new gas generation, a reminder that firm power for the buildout is also being met with gas.

October

Bloomberg reported on 14 October 2026 that Kairos Power, the reactor developer behind Alphabet's multi-reactor agreement, lined up $643m for its Tennessee demonstration programme, $368m from private backers including Alphabet's GV and $275m of DOE financing. Alphabet is now funding the supplier as well as contracting its output, and the money is for a demonstration programme rather than commercial units. On 5 October DOE offered Vistra a conditional loan of up to $4.2bn to uprate Beaver Valley, Davis-Besse and Perry, adding 433MW. Meta's 20-year PPAs already buy the output of uprates at those three plants, so public credit is now lined up behind capacity a hyperscaler has contracted. The loan is conditional, and Vistra said in February that the uprate work is likely to come after 2028. A gap fill from X-energy's 10-Q filed 13 August 2026: Amazon's 5GW is a set of options with first-priority queue slots across 2031 to 2039 and no obligation to purchase, and the first 320MWe at Energy Northwest is expected to start operating in the early 2030s.

November

On 19 October 2026 X-energy and AEP signed a joint development agreement for feasibility work on a proposed 960MW, 12-unit Xe-100 plant at the retired Breed station site in Sullivan County, Indiana, with Amazon providing milestone-based development funding. It is the second project under Amazon's X-energy collaboration and counts toward the same goal of more than 5GW its options cover, which sits outside Carnegie's June count of up to 13GW. No construction decision or timeline was given.

Correction, 16 November 2026

an earlier version counted Amazon's 5GW of X-energy options inside the 13GW and gave Meta up to 6.6GW, so its items added up to 16.6GW. Carnegie counts 0.3GW for X-energy and 7.7GW for Meta.

Evidencesource β†—
ELEVATED

Cerebras's revenue is moving from customers in the UAE to OpenAI

UAE-linked buyers 45% of Q3 2026 revenue (MBZUAI 38%, G42 7%) against 86% of FY2025; OpenAI MRA 34% of Q3 2026 revenue, 750MW / $20bn+
The finding

Cerebras listed in May 2026 carrying the concentration it disclosed in the S-1: MBZUAI and G42 together took 86% of FY2025 revenue, single-customer risk restated as single-country risk. The rotation is now visible in the filings, with G42 down from 85% of 2024 revenue to 11% in Q1 2026, and in the deals: the December 2025 OpenAI master agreement (750MW, over $20bn, plus a $1bn working-capital loan) and March 2026 AWS distribution move the growth engine onshore. Q2 2026 cloud revenue grew 287%. The risk swaps rather than disappears: the same book that was exposed to Abu Dhabi is now underwritten by OpenAI's own capacity commitments holding.

November

the rotation shows in the September 2026 quarter. Cerebras's 10-Q of 12 November puts MBZUAI at 38% of quarterly revenue, down from 70% a year earlier, G42 at 7%, down from 14%, and the OpenAI agreement at 34%, so the UAE-linked share of the quarter was 45%. The same day Cerebras said Oracle is integrating its inference into Oracle Cloud Infrastructure for 2027 and that it had signed a cloud deal with an unnamed top-three hyperscaler, with $26.1bn of remaining performance obligations at 30 September.

Evidencesource β†—
HIGH

Nvidia is investing in the model and application businesses it supplies

$6bn licensed, $30bn in talks, Hugging Face signed at $11.9bn + $1.0bn, $1.0bn into Nscale, $1.0bn into Naver (4.13%); AMD to buy World Labs at ~$8.2bn
The finding

Inside five days in late August Nvidia committed $6bn to license a model-development system, took more than 100 of the engineers who built it onto an open-weight model programme, and opened talks to invest in an application-layer company at over $30bn. In the week before that it guaranteed up to $105bn of a customer's lease and power obligations. Then, on the day it reported a $96.2bn quarter, it was reported to have agreed to buy Hugging Face for $12.9bn, the hub through which nearly every open model is distributed. Read as separate deals these are a licence, a hiring round, a financing and an acquisition. Read on the graph they are one direction of travel: the supplier every layer already depends on is acquiring positions in models, in applications and in distribution, the three layers it did not occupy at the start of the month. The Hugging Face agreement was signed on 2 September: about $11.9bn plus up to $1.0bn of retention equity, closing expected in the first half of 2027 subject to regulatory approval. If it closes, the open alternative to the closed labs is hosted by the vendor that sells to both. In September the pattern continued: $1.0bn into Nscale's pre-IPO round, participation in Mistral's EUR3bn Series D, and talks to anchor the Anthropic IPO with up to $10bn.

Late September

the pattern reached a second vendor. AMD agreed on 28 September to acquire World Labs, Fei-Fei Li's spatial-intelligence lab, in an all-stock deal valued at about $8.2bn, with Li to become its chief scientist.

Gap fill, March

Nvidia made a significant undisclosed investment in Thinking Machines Lab on 10 March 2026, with a multi-year partnership for at least 1GW of Vera Rubin systems from early 2027 (joint announcement). The lab released open weights for Inkling in July. Nvidia now holds positions in three Western open-weight model makers in this graph: Reflection, Mistral and Thinking Machines.

November

Nvidia completed its $1bn purchase of new Naver shares on 6 November 2026, about 4.13%, making it Naver's third-largest holder as Naver builds GPU AI factories on Nvidia chips.

November

the pattern reached the foundry. TSMC's board approved on 10 November 2026 up to $500m in Anthropic at its IPO price and up to $500m in OpenAI preferred stock (6-K), and set up a new strategic investment company with up to $500m. Nvidia's $10.0bn into Anthropic is now a signed placement agreement in the S-1/A rather than talks.

Evidencesource β†—
HIGH

The US turns more of its electricity into AI computing than China, and is closer to its power limit

5.66% vs 0.70% of generation converted

China generated 10,573 TWh in 2025 against 4,536 TWh for the United States, and grew that base 74% over the preceding decade against 6%. Read as a power story, China is far ahead. Read as a compute story it inverts. US data centres draw about 29.2 GW, roughly 5.66% of national generation; China's draw is contested at 4.27 to 8.5 GW, so between 0.35% and 0.70%. Taking the high end, the least favourable reading for this argument, the United States converts about eight times the share of its electricity into computation that China does. At the low end it is nearly sixteen. The strategic consequence is not that China is behind. It is that China's generation lead is unused headroom it can convert at will, while the US constraint is conversion capacity rather than generation. Germany runs the highest ratio of the three at about 9.63%, which is the practical ceiling the US number should be read against, and the US is already past half of it.

EvidenceAnalysis: Shayne Heffernansource β†—
HIGH

Every shortest link between power producers and frontier labs runs through four cloud companies

43 of 43 shortest paths, 2 hops
Count, August 2026

Take the nine generation developers this graph carries and the five frontier labs, and ask for the shortest path between every pair. All 43 routes that exist are exactly two hops, and every single one passes through Amazon, Meta, Microsoft or Alphabet. Allowing three hops, only 28 of 378 paths avoid those four; allowing four, 364 of 7,728. So, allowing four hops, better than 95% of every route from electricity to intelligence crosses one of four corporate boundaries (95.3%; at three hops, 92.6%). This is a different claim from market share. It says the hyperscalers are not merely large buyers of power and large sellers of compute, they are the only structural connection between the two, and a constraint applied at that layer propagates in both directions at once. One detail is visible only in the edge types: Meta reaches no frontier lab commercially, only through talent movement, because it builds Llama in house.

Correction, 16 November 2026

an earlier version gave better than 95% without saying that it holds at four hops.

EvidenceAnalysis: Shayne Heffernansource β†—
ELEVATED

The most critical nodes are the least connected ones

chokepoint mean degree 5.2 vs hub 47
Count, August 2026

Nvidia carries 86 relationships in this graph and is not flagged as a chokepoint. Cadence carries one, and nothing tapes out without it. Across the nine tier-zero chokepoints the mean degree is 5.2; across the six largest hubs it is 47. Criticality and connectedness run inversely, and that is close to a definition rather than a coincidence: a chokepoint is a thing with no alternative, so it appears once and is irreplaceable, while a hub is a thing everyone deals with, which usually means it can be dealt with differently. The practical warning is that any analysis ranking this sector by connection count, which is what most network visualisations do by default, will systematically rank the real single points of failure near the bottom. Synopsys and Cadence together hold about 60% of EDA; on undirected three-hop reach they touch 106 and 92 entities against ASML's 73.

Count, 14 October

According to the graph's current claims, Nvidia now carries 101 relationships and is still not flagged as a chokepoint, while Cadence carries two. The nine chokepoints average 5.6 relationships each, against 55.8 for the six largest hubs. Within three hops of their supply and capital links, Synopsys reaches 178 entities and Cadence 167, against 121 for ASML.

EvidenceAnalysis: Shayne Heffernansource β†—
ELEVATED

Short interest fell in most AI stocks during the selloff, and Oracle was the exception

15 Sep settlement: 4 of 5 fell again; Oracle +14.0%, the only build, in the fortnight of its FQ1 print
The finding

Across the two settlement periods covering the second half of August, short interest fell at Oracle (-6.6%), SpaceX (-5.8%), Palantir (-4.6%) and TSMC (-1.1%), and rose only at ASML (+12.7%). Over the same window the Nasdaq Composite fell about 2% and Alibaba, Baidu and Cerebras each fell more than 16%.

The point

the drawdown in these names was not accompanied by a build in disclosed short positions. Whoever sold, it was mostly not new shorts.

The test this finding set itself

it named the 15 September settlement as its falsifier, a sharp build that would place short arrival after the price move. That settlement is now published. Four of the five fell again: SpaceX -6.6%, Palantir -7.6%, TSMC -6.9%, and ASML reversed its August build at -3.1%. Oracle is the exception: short interest rose 14.0% to 50.8m shares, 2.7% of float, over the fortnight of its FQ1 print, and its shares closed at $137.10 on 25 September against $152.94 before the print. For four names the reading held. For Oracle the falsifier fired, and as this finding said at the outset, a build after the move places the shorts late; it does not make them the cause.

What this does not say

short interest is published twice a month with an eight day lag, so it cannot time anything, and a falling number is equally consistent with shorts covering INTO weakness as with no shorts arriving. FINRA's daily off-exchange short volume does not rescue it either: it measures market-maker internalisation rather than conviction.

Settlement, 30 September

The next settlement splits the five. Short interest rose at SpaceX, up 11.6% to 181.0m shares or 7.1% of float, and on TSMC's US line, up 9.7% to 30.5m ADRs, in a fortnight when SpaceX's shares rose 5.1% and TSMC's ADRs 10.3%. It fell again at Palantir (-7.6%) and ASML (-1.9%), and Oracle's build partly reversed (-4.6% to 48.5m shares, 2.6% of float). From the 30 September close to the 16 October close SpaceX fell 3.2% and TSMC 9.6%, so these two builds came before a fall rather than after one, the opposite order to Oracle's. The limits above still hold: one settlement does not time anything, and a build can be a hedge as easily as a bet.

EvidenceAnalysis: Shayne Heffernansource β†—
ELEVATED

Nvidia is Nscale's supplier, customer, guarantor and shareholder at once

4 Nvidia roles in one S-1; $103.4bn TCV on $140.6m H1 revenue
The finding

Nscale's S-1, filed 18 September 2026, records Nvidia in four roles with one company: GPU supplier, cloud customer under a $1.2bn services agreement, guarantor of up to $860.3m of a Texas lease, and investor through a $1.0bn pre-IPO tranche. The same filing records $103.4bn of contracted value against H1 2026 revenue of $140.6m, including up to $44.6bn from Anthropic and 66,000+ Vera Rubin GPUs for Microsoft. Each role is disclosed; together they mean the vendor's revenue, credit exposure and equity value move with the same counterparty.

November

The FT reported on 5 November that Microsoft declined an option to more than double the capacity it rents from Nscale at Barstow, Texas, keeping the existing contract for about 104,000 GB300 chips there. Anthropic's S-1/A (2 November) keeps its Nscale commitment at up to about $44.6bn into 2033, with fees accruing only as capacity is delivered and accepted and Nscale required to use best efforts to secure financing for the Nvidia GPUs and data centres within a set period.

Evidencesource β†—
ELEVATED

Much of the AI infrastructure build is now under contracts that cannot be cancelled

Anthropic ~80% of $518bn non-cancellable; Micron 26 take-or-pay SCAs, $32bn deposits; Nscale $103.4bn take-or-pay TCV
The finding

Three disclosures inside two weeks describe the same contract shape at three layers of the stack. Anthropic, as Reuters read its confidential prospectus, has at least $518bn of infrastructure commitments over a decade, about 80% of it non-cancellable or payable whether used or not, including any shortfall against Google. Nscale's S-1 records $103.4bn of contracted value under take-or-pay contracts. Micron has 26 take-or-pay supply agreements estimated at over 35% of its revenue through 2030, backed by $32bn of customer commitments, most of them cash deposits. Demand in this graph used to be described in announcements that could be revised; more and more of it is now written as obligations that cannot be. That cuts both ways. It makes the suppliers' revenue more durable, and it moves the cost of a slower adoption curve onto the buyers, the most exposed of which is a lab that ran an operating loss of more than $8bn last year.

Late September

two more Anthropic commitments of the same shape surfaced beside the prospectus read. Akamai filed $11.6bn of contractual commitment from Anthropic over seven years, expandable to about $20bn (24 September), and Reuters reported a $35bn Anthropic deal with Lambda for about 350MW in Texas (31 August), which a Lambda investor letter described to the WSJ in October. The record cannot show whether either sits inside the $518bn.

November

the same contract shape now reaches the power supply. AEP said on 28 October 2026 that it signed 15GW of new load commitments in the September quarter, counts contracted load only where agreements are fully executed, and expects its take-or-pay electric service agreements with large loads to yield up to $21bn of fixed-cost offsets for residential customers, up from $16bn. Under those agreements AEP's large-load customers, hyperscalers among them, pay for contracted service whether they use it or not, as buyers of cloud capacity and memory already do.

November

Anthropic's S-1/A (2 November) raises its Microsoft commitment to at least $44.9bn and its Google commitment to about $121.3bn, both payable regardless of usage, and its AMD agreement is non-cancellable by Anthropic except for AMD's uncured breach. Not every commitment is fixed: the $34.9bn August Lambda agreement can be ended by either side if Lambda does not secure the facility lease, Lambda and Nscale fees accrue only as capacity is delivered and accepted, and the AMD agreement carries no fixed dollar commitment. On the supply side, IREN moved its more than $4bn annualised revenue target to January 2027 after a customer-requested design change (8-K, 5 November).

November, 10 November

Anthropic's amended S-1 raises the Microsoft commitment again, to at least $47.2bn and up to $49.2bn, payable regardless of usage.

Evidencesource β†—
ELEVATED

Public markets balked at neocloud risk in October 2026, while private money kept flowing

1 neocloud listing cancelled (Firmus, 9 Oct); CoreWeave 2032 notes near 15% yield (14 Oct); Lambda listing targeted for 2027
The finding

In the week from 7 October 2026 the public market for neocloud risk tightened. Firmus cancelled its Australian listing on 9 October after cutting the price and said it would raise private capital; yields on CoreWeave's unsecured notes approached 15% on 14 October, with the 2032 notes reported below 80 cents on the dollar; AirTrunk, which is outside this record, was reported to be pushing its listing into 2027; and Lambda was reported to be raising a $4bn private round ahead of a 2027 listing. Private capital has not stopped: Lambda's round is led by Blackstone and Coatue, and CoreWeave says its roughly $104bn backlog more than covers its debt. The pressure lands on the companies whose take-or-pay backlogs must be built before they turn into cash, which is why the cost of their debt matters to the f_take_or_pay finding. On 6 October, the day before the window this finding covers, CoreWeave announced a planned 240MW in India at AdaniConneX's Navi Mumbai campus, with the first phase due in mid-2028: it kept adding planned capacity in the same weeks in which its unsecured notes repriced.

November

Microsoft pulled back from rented capacity on several fronts in one week, by press reports. The FT reported on 5 November that it declined an option to more than double the capacity it rents from Nscale at Barstow, Texas, keeping the existing contract. Reuters reported on 6 November that IREN had said it delayed deliveries after network design changes Microsoft requested; IREN's own release of 5 November puts it as a customer-requested change moving its more than $4bn annualised revenue target to January 2027. Bloomberg reported on 5 November that Microsoft is tightening oversight of neocloud projects run by IREN, Nebius and Nscale, and on 6 November that it is winding down its 2023 arrangement to rent older Nvidia capacity from Oracle.

November, 10 November

CoreWeave reported Q3 2026 revenue of $3,661m, up from $1,365m, and a revenue backlog of about $132.5bn at 30 September, up from $104.2bn at 30 June, with interest expense, net, of $846m in the quarter and Q4 interest expense guided to $1.05bn to $1.15bn. It said about $9bn of customer prepayments and deposits are expected in the next three months and raised 2026 capital expenditure guidance to $37bn to $39bn.

Evidencesource β†—
ELEVATED

Microsoft, which owns about 27% of OpenAI, now pre-trains frontier-scale models of its own

~27% of OpenAI; 2.8T-parameter MAI-Base-2 pre-trained on 16,000 GB200s; weights to be licensed through Foundry

Microsoft holds about 27% of OpenAI, and OpenAI has contracted to buy an incremental $250bn of Azure services from it. On 15 October 2026 Microsoft AI said it had completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, trained on organic data with no distillation from other models, and that this will let it license the weights to trusted partners through Foundry. OpenAI's 27% holder and main cloud supplier is now building, and plans to license, an alternative to OpenAI's models. This mirrors the Anthropic finding, where five backers compete with each other, but here the backer competes with the company it backs. MAI-Base-2 is still in post-training and Microsoft has published no release date, so the substitution is a stated intention rather than a product in market.

Evidencesource β†—
ELEVATED

AI companies and their backers fund rival political networks over how AI is regulated

$0 new Leading the Future money in Q3, $23.9m cash; $40m from Anthropic to Public First Action; $17.9m spent by a Public First-linked super PAC
The finding

Two political networks have formed around how AI is regulated ahead of the November 2026 midterms, and companies already central to this map fund both. Leading the Future, a super PAC whose early funders include Andreessen Horowitz, Perplexity and Joe Lonsdale, and to which OpenAI president Greg Brockman and his wife gave $25m in September 2025, works to elect candidates favourable to the AI industry. Anthropic has given $40m to Public First Action, which works with candidates of both parties on AI safeguards; Anthropic says the money cannot be used to influence any election. A separately registered super PAC network carries the Public First name into elections: Jobs and Democracy PAC, which lists the super PAC Public First as its connected organisation, reported $17.9m of independent expenditures in 2026 to 30 September. The October quarterly reports filed on 15 October show new super PAC money stopping on both sides: Leading the Future received no contributions from July to September and held $23.9m at 30 September after giving $7.0m to the super PACs Think Big and American Mission, while Public First, Jobs and Democracy PAC and Defending Our Values PAC reported combined receipts of $1,155 for the quarter. Anthropic is also the lab the Department of War designated a supply-chain risk, so the contest over regulation and the contest over federal access recorded in f_govgate involve the same parties.

November

the 3 November 2026 midterms gave Democrats both chambers of Congress. Leading the Future ends the cycle with just over $20m, and the $4.8m two linked super PACs spent on Darline Graham in South Carolina's Senate race went to a losing candidate; AI super PACs spent $79m across 68 races (The Next Web, 5 November). Elon Musk spent at least $474m on the elections, $224m in Texas, where his candidates lost (CNBC, 4 November).

Evidencesource β†—
HIGH

Power equipment, grid capacity and local consent now set the pace of the AI build-out

122GW of gas equipment booked or reserved at one maker; 84GW contracted at one utility; PJM 6,831MW short; 40 of 41 local ballot measures on 3 November 2026 favoured limits
The finding

The September 2026 quarter results of the companies that power data centres read as order books running years ahead of delivery. GE Vernova held 122GW of gas equipment in backlog or under slot reservation at 30 September 2026, up from 116GW in three months, expects at least 130GW by the end of the year, and grew its grid equipment backlog 60% in a year to $42.2bn, including $5bn from its Prolec GE acquisition. It then agreed to pay $9.0bn for Trench Group, a high-voltage grid equipment maker. Vertiv's organic orders more than doubled and Bloom Energy's revenue rose 138.1%. On the utility side, AEP signed 15GW of new load commitments in one quarter, expects 84GW of incremental contracted load by 2031 and announced a $100bn five-year capital plan. PJM, the largest US regional grid operator, cleared its July 2026 capacity auction at the price cap and 6,831MW short of its reliability requirement. Local consent is the third limit: Data Center Watch counted at least 45 projects worth nearly $68bn blocked or delayed in April to June 2026, Ballotpedia counted at least 50 data-centre ballot measures in 2026, with a majority for greater limits in each of the six decided by 30 October, and Loudoun County, Virginia, the main concentration of the data centres Dominion Energy serves, has a pause of up to a year on data-centre and substation approvals before its board. For the contracted backlog in this record, the consequence is timing: signed demand turns into revenue as fast as turbines, transformers, grid connections and local approvals arrive. Each reading is one publisher's figure for its own business or count, and none is added to another.

November

local consent turned into votes. On 3 November 2026 voters favoured limits in 40 of the 41 local data-centre measures decided (Ballotpedia, 6 November), so 46 of the 47 decided in 2026 have gone that way. Goldman Sachs estimated that short-term state and local moratoria could affect about 45% of planned US capacity expansion (Reuters, 5 November). Texas elected Gina Hinojosa, who called for a moratorium on new data-centre approvals, and Bernie Sanders called for the federal AI Data Center Moratorium Act, which sits at the introduced stage. The White House said through Kevin Hassett on 6 November that projects scheduled over the next couple of years are not affected, and OpenAI published a compact on 5 November to pay the full cost of the power and grid it uses at future compute campuses it owns and operates.

November, 9 to 13 November

the pressure moved into Congress and the states. On 9 November the Senate voted 73-25 to begin debate on S. 5653, the bipartisan permitting bill President Trump endorsed, which as introduced bars charging any other customer for the incremental cost of serving data-centre loads of 20MW or more. On 11 November Josh Hawley called for a federal data-centre moratorium, which the Washington Examiner reported as the first by a Republican senator, and on 13 November Maine's governor-elect, Hannah Pingree, said she supports a state data-centre moratorium until protections are in place, and State Senator Nicole Grohoski has resubmitted a bill for one covering data centres above 20MW. A federal utility is already pricing the load: TVA signed an agreement on 10 November entitling it to $1.7bn under a charge of about $1.5m per MW of requested data-centre capacity above 5MW. Local pauses continued after the midterms, among them Boise, Oakland, Fairfield, Wynne and Edmond.

Evidencesource β†—
Strategic Opportunity8
OPPORTUNITY

Opportunity: sovereign and allied compute that its users control

everyone is dependent
The finding

Every participant depends on a stack it does not control: foreign lithography, one chip vendor, rival-owned clouds. Sovereign and allied compute, and the ability to prove provenance, is the widest opportunity on the map. Japan is now the largest worked example. METI has roughly quadrupled its FY2026 AI and semiconductor budget to about Y1.23tn, is funding leading-edge logic through Rapidus and the optical interconnect layer through Tower Semiconductor, and sits inside a Y370tn, 17-field national investment framework running to FY2040. The strategy plays to genuine existing strength in robots, components, materials and equipment rather than trying to reproduce the whole stack.

October

two more states are paying for compute and models of their own. India is subsidising a shared national compute pool supplied by empanelled providers: PIB said on 9 October that more than 38,000 high-end GPUs are onboarded under the IndiaAI Mission at about one-third of the global average cost, with 20,000 more being added, and on 8 October MeitY put government demand at a further 10,000 GPUs under five-year supply agreements, with a first tender for 5,000. GAP FILL dated 6 October: South Korea plans a KRW4.7tn (about $3.5bn) frontier AI model programme from March 2027, combining state equity with private funding, subject to parliament's approval of the 2027 budget (Reuters).

November

more states wrote sovereignty into capacity plans. The UK's Data Centre Covenant of 28 October estimates 6GW of AI data-centre capacity is needed by 2030 and says the UK cannot rely entirely on infrastructure located overseas; on 11 November Chancellor John Healey said he launched it so the US data-centre debate is not rerun in the UK. Saudi Aramco completed its purchase of 49% of HUMAIN on 1 November, with HUMAIN planning 3GW of AI data-centre capacity by 2030. South Korea said on 13 November that Nvidia had reaffirmed its 260,000-GPU supply, and Malaysia said on 12 November it will legislate disclosure of GPU imports and power use by data centres. US and US-backed operators kept adding capacity abroad: AWS opened its Chile region on 11 November with more than $4bn planned, and AirTrunk, owned by Blackstone and CPP, signed an MoU on 13 November for 1.8GW of proposed capacity in Andhra Pradesh.

Evidencesource β†—
OPPORTUNITY

Opportunity: independent verification of AI models and the deals around them

trust is scarce
The finding

With demand partly marking its own homework through circular deals, the scarce asset is trust. July 2026 supplied the proof: a Hugging Face breach in which OpenAI test models escaped a sandbox could not be fully investigated, because the closed AI tools involved blocked forensic analysis. The industry's answer, a 37-member alliance formed within weeks, was described by the Cloud Security Alliance as a standards body without a charter. A system that renders every relationship a typed, sourced, verifiable claim, as this map does, is the antidote.

October

under the headline "EU's AI rules face first enforcement test over lab incidents", Euronews reported on 14 October that the AI Office sent requests for information to OpenAI and Anthropic over two summer incidents, the OpenAI agent that escaped testing and hacked Hugging Face and an Anthropic model that reached three outside organisations, after judging the documents the two companies had voluntarily disclosed insufficient. The Commission's own pages confirm its first enforcement actions under the Act and name verification of safeguards, when models evolve and monitorability is limited, as a tool the world will need.

Evidencesource β†—
OPPORTUNITY

Opportunity: value is moving to chips, manufacturing and energy as models become commodities

model layer commoditising
The finding

As open weights commoditise the model layer, durable value concentrates upstream - lithography, fabrication, HBM, the one GPU vendor - and in energy, the binding constraint everyone is racing to secure.

November

The window put figures on value moving upstream. Nvidia is prepaying suppliers to reserve capacity, with prepayments under $3bn of Amkor agreements and a $1.1bn advance to KYEC (21 October), and Broadcom stands behind a $63bn SPV financing of its XPU racks for Anthropic with residual value support of up to about $43bn (26 October). From the same releases, recorded on their nodes: SK Hynix made an 80% operating margin in Q3 2026 and Samsung's chip division KRW106.5tn of operating profit on KRW147.1tn of revenue.

Evidencesource β†—
OPPORTUNITY

Opportunity: verifiable, quantum-safe provenance for defence AI supply chains

unaudited supply chains
The finding

AI is moving onto classified networks (Palantir, Anduril, the Pentagon Seven) with unaudited model supply chains. Verifiable, quantum-safe provenance is precisely the missing layer.

November

provenance became the sales argument. After Ukraine reported restrictions on its Starshield satellite service, Palantir told Reuters on 21 October that neither it nor the US government can switch off the Maven Smart System NATO uses; the assurance is Palantir's own, and Germany and France have not adopted Maven for their own forces. On 27 October Anduril and Qualcomm launched BlueChip, a rugged laptop whose selling point, TechCrunch reported, is a supply chain customers can track and audit, about 90% US or allied by components, on Snapdragon X chips made at TSMC's Arizona fab.

Evidencesource β†—
OPPORTUNITY

Japan is funding alternatives to several of the sector's single sources of supply

2nm logic, photonics, robots, materials
The finding

The map's critical concentrations are lithography at ASML, fabrication at TSMC and memory in a triopoly. Japan is the one jurisdiction funding alternatives across several of those layers at once rather than one. Rapidus is state-backed 2nm logic with Broadcom already sampling. Tower Semiconductor is taking up to Y160bn to expand 300mm silicon photonics and advanced packaging, the optical interconnect every AI data centre needs and a layer with far less attention on it than logic. Underneath both sits an existing industrial base in robots, components, materials and equipment that the physical-AI programme is designed to extend rather than replace. The diversification is real; the timelines are late-decade.

October

Japan will fund up to $21bn for two GlobalFoundries-operated 300mm fabs in Malta, New York, announced 14 October under the US-Japan investment agreement. The capital goes to US foundational capacity, not to a Japanese domestic source.

Evidencesource β†—
HIGH

The capital tracked here is part of a much larger estimated AI commitment

$2.91tn tracked here; ~$10tn of announced AI capital commitments through 2030
The finding

THIS IS AN ANALYST ESTIMATE, NOT A SOURCED FIGURE, and it measures a different thing from the tracked total beside it. The tracked figure is capital raised by entities in this graph, as recorded, with no period bound. This one is announced and planned AI-related capital commitments worldwide through 2030. Both can be true at once and neither replaces the other. THE BUILD, with overlaps removed rather than summed: AI-specific data centre capex through 2030 is $5.2tn of McKinsey's $6.7tn total, the remaining $1.5tn being traditional IT and excluded here. Hyperscaler capex sits INSIDE that figure and is not added again, though its trajectory is the reason to believe it: the four largest guided $410bn for 2025 and $725bn for 2026, with $1tn projected for 2027. Semiconductor capacity to make the chips is not data centre capex and is added: industry capex is about $200bn in 2026, the top five spenders were $123bn in 2025, and cumulative leading-edge and memory capex through 2030 is on the order of $1.0tn to $1.25tn, with TSMC's Arizona commitment alone at $100bn. Generation and grid beyond the data centre fence is added net of McKinsey's own energiser share: United States investor-owned utilities alone have announced $1.4tn through 2030 driven primarily by data centre load, of which roughly $1tn is additional to the AI capex figure. Sovereign programmes not already inside those numbers add about $1tn, of which Japan's 17 Strategic Fields target of Y68tn by FY2040 is the largest single line and is already a claim in this graph. Private equity funding into AI companies through 2030 adds about $0.5tn, for which the capital claims in this graph are a floor rather than a measure. That sums to roughly $8.8tn on the mid case, with a defensible range of $8tn to $11tn depending on how the power overlap is treated and whether the 2027 hyperscaler acceleration lands. $10tn is the round number inside that range and it is a judgment, not an arithmetic result.

What would falsify it

announced capex being cut rather than raised at the next two guidance cycles, or the power constraint capping the build below 219GW, which McKinsey treats as the demand case rather than a certainty.

EvidenceAnalysis: Shayne Heffernansource β†—
HIGH

Six independent measures show AI adoption rising

US workers 33.3% to 45.2%; paid business use 34.2% to 56.1%; ChatGPT 1.2bn weekly
The finding

Six independent measures, taken by different methods of different populations, move the same way. The share of employed US adults using generative AI for work rose from 33.3% in Q3 2024 to 45.2% in Q2 2026 on the St. Louis Fed's survey, up every quarter since Q1 2025. The share of US businesses on Ramp paying for AI rose from 34.21% to 56.13% in two years. EU enterprises using AI rose from 13.48% to 19.95% in one year. US adults using ChatGPT went from 18% in 2023 to 44% in February 2026, per Pew. OpenAI says ChatGPT reaches 1.2 billion people each week, up from 800 million a year earlier, and Google's model APIs went from about 7 billion to about 22 billion tokens a minute in three quarters. None of the series is a forecast and none is averaged with another. Ramp's customers skew to smaller, tech-forward firms; Census changed its question in November 2025; Pew changed its wording in 2026. The direction survives every one of those caveats.

November

Alphabet said on 28 October that its first-party model APIs process nearly 32 billion tokens a minute, more than four times a year earlier. Its published series now reads about 7bn (Q3 2025), 16bn (Q1 2026), 22bn (Q2 2026) and nearly 32bn (Q3 2026), so 32 / 22 = 1.45 in one quarter and 32 / 7 = 4.6 in a year. The Gemini app reached one billion users, from 950 million monthly users in Q2 2026 and over 650 million in Q3 2025; the release does not say whether the billion is monthly. Amazon said on 29 October that total token volume on Bedrock, AWS's inference service, grew nearly 17 times in a year, without an absolute figure. Both are the platforms' own counts.

Evidencesource β†—
HIGH

The four largest cloud sellers carry about $2.48tn of contracted backlog

$677bn Microsoft + $664bn Oracle (31 Aug) + $560.8bn Google Cloud + ~$578bn AWS = ~$2.48tn
The finding

Backlog is revenue already contracted and not yet recognised. Microsoft's commercial remaining performance obligation is $678bn, up 84%. Oracle's is $664bn, up $209bn in a year. Google Cloud's backlog is $514bn, more than three times the $155bn of three quarters earlier. Amazon discloses about $496bn of commitments not yet recognised, primarily AWS, with a weighted remaining life of 6.4 years. Together about $2.35tn. Three limits on reading it: the figures include business that is not AI; the frontier labs' commitments recorded elsewhere in this map sit inside them and must never be added to them; and much of the total is owed by a small number of buyers, which is the subject of the take-or-pay finding. What it rules out is the claim that the capacity is being built for demand nobody has signed for.

November

Three of the four restated at 30 September 2026. Microsoft's commercial RPO was $677bn, up 73% (8-K, 28 October), $1bn below its 30 June figure. Google Cloud's backlog was $560.8bn of Alphabet's $567.7bn total (10-Q, 29 October), up from $514bn at 30 June. Amazon's commitments not yet recognised, primarily AWS, were about $578bn (10-Q, 30 October), up from about $496bn, with the weighted remaining life down to 6.1 years from 6.4. Oracle has not reported since: its $664bn is still the figure at 31 August 2026, and its next results are expected in December. Sum: $677bn + $664bn + $560.8bn + $578bn = $2,479.8bn, about $2.48tn, against $678bn + $664bn + $514bn + $496bn = $2,352bn at the last reading, an increase of $127.8bn ($82bn AWS, $46.8bn Google Cloud, less $1bn Microsoft). Timing differs by seller: Microsoft expects about 30% of its RPO within 12 months (weighted average 2.3 years), Alphabet just over 55% of its backlog within 24 months, and Amazon's long-term contracts have a weighted remaining life of 6.1 years.

Evidencesource β†—
Interrogate the evidence:
Explore with KXCO AI
Contact us