Power Concentration in AI — Live Indicators
“Latest Observation” means the most recent comparable observation located for each indicator. Reporting frequencies differ, so observation dates vary by row and are stated in each cell.
The argument does not depend on every indicator increasing each quarter. This page tracks the structural conditions identified in the essay — capital mobilization, governance control, infrastructure concentration and dependency, and the resulting physical and political footprint — to see whether they are strengthening, weakening, or persisting over time. Indicators are updated when a new comparable observation becomes available; because the underlying data are reported on different schedules, comparisons may be quarterly, annual, event-driven, or based on a new research release.
| Indicator | Latest Observation | Previous Observation | Direction | Source |
|---|---|---|---|---|
| Capital Mobilization | ||||
| Combined reported 2026 capex guidance, company definitions — Amazon, Alphabet, Microsoft & Meta 1 |
$720–745B post-Q2 guidance |
$695–725B post-Q1 guidance |
↑ | Company guidance: Amazon · Alphabet · Microsoft · Meta |
| Nvidia Data Center revenue |
$89.0B Q2 FY2027 |
$75.2B Q1 FY2027 |
↑ | Nvidia 10-Q |
| Governance Control | ||||
| Zuckerberg voting power / economic interest — Meta |
60.8% / ~13.5% Apr. 1, 2026 |
61.0% / ~13.6% Apr. 1, 2025 |
→ | Meta proxies: 2026 · 2025 |
| Page & Brin combined voting power — Alphabet |
52.7% Apr. 2026 |
52.3% Apr. 2025 |
↑ | Alphabet proxies: 2026 · 2025 |
| OpenAI Foundation equity interest and governance control |
26% equity at recapitalization close appoints and can replace all OpenAI Group directors; checked Sep. 9, 2026 |
No prior observation | — | OpenAI — Our structure |
| Occupied Anthropic board seats held by Long-Term Benefit Trust appointees 2 |
Not determinable from public disclosures 6 seats occupied; checked Sep. 9, 2026 |
Majority of 7 Apr. 14, 2026 |
— | Anthropic: governance page · Apr. 2026 appointment announcement |
| Infrastructure Concentration & Dependency | ||||
| Nvidia estimated share of AI training accelerators |
>90% Mar. 2026 analyst estimate |
No prior observation | — | Reuters / Summit Insights |
| TSMC share of global foundry revenue 3 |
72.5% Q2 2026 |
72.3% Q1 2026, revised comparative |
↑ | TrendForce |
| AWS + Microsoft + Google share of cloud infrastructure services |
63% Q2 2026 |
63% Q1 2026 |
→ | Synergy: Q2 2026 · Q1 2026 |
| Hyperscalers' share of Nvidia Data Center revenue 4 |
54.7% Q2 FY2027 |
57.2% Q1 FY2027, current classification |
↓ | Nvidia 10-Q |
| Footprint & Political Accountability | ||||
| Global data-center electricity consumption |
485 TWh 2025; ~1.5% of global electricity |
415 TWh 2024; ~1.5% |
↑ | IEA |
| Data centers' share of Virginia Power electricity sales |
28% 2025 |
26% 2024 |
↑ | Dominion Energy / SEC |
| Virginia localities with at least one data-center-related NDA |
25 of 31 — 81% 2025 FOIA study |
No prior observation | — | University of Mary Washington |
| Publicly tracked AI-linked independent expenditures 5 |
$53.9M through Jul. 25, 2026 |
No prior observation | — |
Elect Humans
public-filings compilation |
↑ indicator increased → materially unchanged ↓ indicator decreased — no comparable prior observation
Important: the arrow describes movement in the indicator itself. It does not automatically mean that “concentration increased” or “concentration decreased.” Interpretation depends on what the indicator measures.
Source and interpretation notes
1 Capital expenditure. The $720–745B range is the sum of the most recent 2026 capital-expenditure guidance from the four companies. They do not define capital expenditure identically, so this is an aggregate of reported guidance rather than a like-for-like measure.
Meta economic interest. Zuckerberg's economic-interest percentage is calculated from proxy-reported beneficial holdings and shares outstanding. Meta itself describes his economic interest as approximately 13%.
2 Anthropic. Anthropic's governance page listed six directors when checked on September 9, 2026: Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell and Vas Narasimhan. Anthropic has announced Long-Term Benefit Trust appointments for some of these seats but not all, and it did not announce Jay Kreps's departure from the board. How many occupied seats are held by Trust appointees is therefore not determinable from public disclosures. In April 2026 Anthropic described Trust appointees as holding a majority of a seven-member board.
Nvidia accelerator share. The >90% figure is an analyst estimate rather than a company-reported market statistic. No prior estimate with a sufficiently clear and comparable denominator is used to manufacture a trend.
3 TSMC foundry share. TrendForce's September 2026 release puts TSMC at 72.5% of global foundry revenue for Q2 2026, on revenue of about $40.2B, up 12.1% sequentially. Its June release put Q1 2026 at 72.0%; TrendForce's current dataset, as reported by CNA / Focus Taiwan and the Taipei Times, puts that quarter at 72.3%. The sequential growth rate implies TSMC's Q1 revenue is unchanged at about $35.9B, so the revision falls on TrendForce's estimate of total foundry revenue rather than on TSMC. This row uses the revised comparative, making the quarter-on-quarter increase 0.2 percentage points.
Market-share measures. TSMC and cloud-provider shares rely on independent market-research organizations because individual company filings cannot by themselves establish industry-wide market shares. Synergy reports the top-three cloud share at 63% in Q2 2026, unchanged from Q1, while the cloud infrastructure market grew 43% year over year.
4 Nvidia hyperscaler dependency. Nvidia reports Hyperscale revenue as a submarket of Data Center, covering the public clouds and the world's largest consumer internet companies. In Q2 FY2027 it reclassified a company from AI Clouds, Industrial, & Enterprise (ACIE) to Hyperscale due to a change in that company's business model and recast the prior period. Q1 FY2027 hyperscale revenue was first published as $37.9B of $75.2B in Data Center revenue, or 50.3%; on the current classification it is $43.1B of the same $75.2B, or 57.2%. The current-basis observations available are Q1 FY2026 at 56.9% (derived from six-month totals), Q2 FY2026 at 58.8%, Q1 FY2027 at 57.2% and Q2 FY2027 at 54.7%. Q3 and Q4 FY2026 have not been republished on this basis, so the quarterly series is discontinuous. The Q2 year-over-year comparison, 58.8% to 54.7%, is like-for-like. The reported Hyperscale category is also a floor on hyperscaler-linked end demand: Nvidia attributes part of ACIE growth to hyperscalers utilizing AI clouds.
Electricity. The IEA figure covers all data centers, not AI facilities alone. Its current central projection is roughly 950 TWh of global data-center electricity consumption by 2030. That projection provides context but does not determine the direction arrow.
Virginia NDAs. The 25-of-31 finding means researchers found at least one data-center-related NDA in 25 Virginia localities with an existing, approved, or proposed project. It does not mean that 25 projects were secretly approved or that the approvals themselves occurred “under” NDAs.
5 Political expenditures. The $53.9M figure is a secondary compilation of publicly disclosed independent expenditures across AI-linked federal and state political networks. It is not equivalent to money raised and should be treated as a publicly observable floor, not a comprehensive total of all political activity.
Method
This page is designed as a living empirical companion to Power Concentration in AI, not as a composite index or a scorecard in which every arrow is expected to point upward.
Latest Observation means the newest publicly available observation that can be stated on a reasonably comparable basis.
Previous Observation means the preceding genuinely comparable observation. It is not automatically the previous calendar quarter. Some indicators update quarterly, some annually, some when corporate structures change, and others only when new research becomes available.
An older number is never carried forward and presented as a new observation merely because time has passed. No new observation does not mean “unchanged.” A → is used only when a newer comparable observation actually indicates little or no material movement.
When a provider revises a previously published comparative, the tracker uses the provider's revised figure. If the superseded figure has previously appeared on this page, the revision is disclosed in the notes rather than silently rewriting the record.
Projected values may provide context but never determine a direction arrow.
Primary company, regulatory, and government sources are preferred where they can establish the relevant fact. Independent research organizations are used where a statistic requires market-wide measurement or original research that individual companies cannot provide. Secondary compilations are explicitly identified as such.
The purpose of the tracker is not to confirm the essay's argument. It is to record evidence that can move in either direction — including evidence of persistence, diversification, declining dependency, or weakened concentration.
Read the accompanying essay: Power Concentration in AI . Indicators are reviewed as new comparable observations become available.
Read the essay: Power Concentration in AI