Every number below is an estimate produced by models and triaged by an editor. None is an audited statistic. Where the panel agreed within a narrow band the figure is given as a median with the range; where it did not, the range is the finding.

1. The structural ranking

Median rank across twelve models. Totals are the median of each model’s de-duplicated figure, in millions of people, with the panel’s min and max. Ranks are the robust signal; totals are not additive across entities because every model used its own scale.

ByteDance, added in a second round (section 4), would enter this table at 5th with ~2,450M, behind Microsoft and ahead of Apple, OpenAI and Amazon.

Two things the table cannot show. First, the top three saturate at the internet population, so the metric stops discriminating: it cannot tell Nvidia from Google from Meta, only that each touches nearly everyone online. What separates them is substitutability, which is section 3. Second, ranks 11 to 13 are one cluster; the panel’s median rank was 12 for all three, and the order given is by median total.

2. Reach estimates, the derivations the panel converged on

• Nvidia. Direct 0 to 150M (GeForce, DGX Cloud). Indirect and dependency together 4.3 to 5.5B: internet users 5.5 to 5.8B times the 80 to 85% of AI compute on Nvidia silicon, minus offline and CPU-only traffic. Unique reach beyond what other entities already claim: 500 to 800M (gaming, HPC, robotics). Least sure: the Nvidia share in inference as TPU, Trainium and Ascend grow.
• Google. Direct 400M to 2B depending on whether AI Overviews count as knowing use. Indirect 3.3B: Search, YouTube, Android and Play de-duplicated. Dependency 0.5 to 2B via Vertex, Workspace seats, Android OEM features. Least sure: the share of Search sessions that now surface a Gemini synthesis.
• Meta. Direct ~1B Meta AI MAU (last disclosed ~1B, 2025; 1.2B is plausible, unverified). Indirect 3.2B: family daily people under ranking, ads and safety models. Dependency 100 to 900M for third-party Llama deployments, the widest disagreement on any single cell.
• Microsoft. Direct 150 to 500M Copilot users across Windows, Office, GitHub. Indirect 1.3 to 1.5B: Windows device base with embedded AI. Dependency 300M to 1.3B via Azure OpenAI serving downstream apps. Least sure: paid M365 Copilot seat activation.
• Apple. Direct 400 to 900M on Apple Intelligence-capable hardware. Indirect 1.1 to 1.45B humans on the active device base, all under CoreML. Dependency 0 to 700M, mostly the ChatGPT handoff.
• OpenAI. Direct 350 to 900M (ChatGPT WAU 400M Feb 2025, ~700M Aug 2025, extrapolated). Indirect 200 to 600M via API apps with 30% overlap removed. Dependency 600 to 900M via Copilot and Siri handoffs. Least sure: Siri to ChatGPT opt-in frequency.
• Amazon. Direct 100 to 400M (Rufus, Alexa+). Indirect ~1B retail consumers under pricing, recommendation and logistics ML. Dependency 500M to 1.8B end users of Bedrock and SageMaker customers, the cell with the least ground truth.
• Tencent. Direct 80 to 700M (Yuanbao, WeChat AI search). Indirect: WeChat 1.38B MAU times 50 to 80% exposure to Hunyuan ranking, search and Mini Program recommendations. Dependency 100 to 450M via Tencent Cloud and games.
• Alibaba. Direct 50 to 600M (Tongyi/Quark, DingTalk). Indirect 800M to 1B commerce consumers. Dependency 200 to 650M for Qwen fine-tunes on Aliyun or self-hosted. Least sure: Qwen’s international developer share.
• Xiaomi. Direct 80 to 500M (XiaoAI prompts). Indirect 500 to 700M HyperOS MAU. Dependency 50 to 100M partner IoT. SU7 fleet adds nothing measurable yet.
• DeepSeek. Direct 20 to 120M app users, peaked early 2025 and settled. Indirect 100 to 650M via official API and Chinese platform embedding (Tencent Yuanbao served DeepSeek R1 for roughly half its 2025 answers). Dependency 200 to 400M via distills re-hosted in the West.
• Anthropic. Direct 30 to 75M. Indirect 100 to 400M via Claude Code, Cursor, Notion, Slack, GitHub Copilot integrations. Dependency 250 to 500M via Bedrock and Vertex enterprise workflows. Least sure: end users inside corporate SaaS pipelines, which no lab discloses.
• SpaceX + Tesla + xAI. Direct 10 to 200M (Grok on X, FSD owners). Indirect 30 to 600M: X feed ranking is the only large number and half the panel excluded it as legacy ML. Dependency 0 to 300M, mostly Starlink routing, which several models refused to count as AI.
• Zhipu. Direct 20 to 100M ChatGLM. Indirect 50 to 250M SOE and telecom deployments. Dependency 0 to 300M open-weight GLM derivatives. Every model marked this row LOW.
• Kimi. Direct 15 to 45M. Indirect 30 to 100M enterprise long-context API. Dependency 10 to 100M. Every model marked this row LOW.

3. The nested doll

The dependency graph, with the panel’s estimate of how much of each inner doll’s reach is really the outer doll’s audience relabelled.

ASML (EUV, monopoly) > TSMC (leading-edge fab, near monopoly) > HBM: SK Hynix, Samsung, Micron
        |
     NVIDIA (80 to 85% of AI compute)
        |-- Microsoft Azure --> OpenAI --> Apple Intelligence handoff
        |        |                 |
        |        +-- Copilot       +-- API apps (Duolingo, Shopify, Stripe, thousands more)
        |-- AWS Bedrock --> Anthropic, Mistral, Cohere --> enterprise SaaS end users
        |-- Google Cloud / Vertex --> Anthropic (second home), Gemini ecosystem
        |-- Meta --> Llama weights --> global developers, including Chinese fine-tunes
        |-- Tencent, Alibaba, DeepSeek (H800, H20, stockpiles, grey channel; partial escape via Ascend, Cambricon)

• Below Nvidia there is no audience, only construction. TSMC’s Nvidia-mediated reach is 100% Nvidia relabelled by definition; ASML’s is 100% TSMC relabelled. They belong on the matrix as chokepoints, not as entities with people. GLM and GPT-6 both made this point unprompted.
• Nvidia is 70% other people’s users. Of Nvidia’s ~4.5B, roughly 70% is Google, Meta, Microsoft, OpenAI and the Chinese clouds counted once more (GLM). Nemotron, Nvidia’s own model, went further: it gave Nvidia zero direct and zero indirect reach and put the entire 4.5B in the dependency column.
• OpenAI runs on borrowed distribution. 40 to 70% of OpenAI’s effective reach arrives through Microsoft and Apple surfaces (Gemini 70%, Muse Glimmer 59%, Qwen 48%). GPT-6 Astra Pro dissented at 16%, arguing that adding partner installed bases does not prove distinct exposure. The reverse edge is also large: 40% of Microsoft’s AI reach is OpenAI IP (Glimmer), 65% of Azure AI specifically (Gemini).
• Anthropic is a cloud tenant first. 50 to 85% of Anthropic’s reach is AWS- or Google-mediated. Seen from Amazon’s side it is ~2% of Amazon’s total, which is why Anthropic looks small on a headcount matrix and large on an enterprise-revenue one.
• Apple outsources the open-domain tail. About 25% of Apple Intelligence’s functional reach is OpenAI compute behind a consent modal (Gemini, Glimmer agree).
• The Chinese cluster sits inside the Western doll. 60 to 65% of Tencent’s and Alibaba’s AI reach is Nvidia-relabelled (Glimmer). DeepSeek trained on H800s. SMIC runs ASML DUV. And the doll now nests both ways: Tencent Yuanbao served DeepSeek, and DeepSeek and Qwen run inside Western enterprise pipelines that swapped API keys to cut inference cost.

Structural reliance is not irreplaceability. Replacing an upstream supplier costs years and billions, but that does not make every downstream user an additional person for the supplier. The headcount matrix measures exposure; a second matrix, substitution cost, would invert the top: ASML above TSMC above Nvidia above everyone, and it is the matrix a policymaker should read.

4. The geopolitical divide

Mistral Medium 3.5 reported 12,000M Western and 6,000M Chinese, more than the human population; its cluster figures were discarded and only its ranking kept.

Three cross-dependencies that make the split leaky, named by every model that reached section 4:

1. Silicon. Chinese models train and serve on Nvidia (H20, pre-ban stockpiles, grey channels) and TSMC-fabbed parts; SMIC’s 7nm runs on ASML DUV; both clusters draw HBM from Korea. The “sovereign Chinese stack” is physically inside the Western supply doll, and the Western stack is physically inside a Taiwanese fab and a Dutch lithography line.
2. Open weights, both directions. Llama is fine-tuned by Chinese firms; Qwen, DeepSeek and GLM are downloaded and deployed by Western developers and hyperscalers to cut inference cost. Model IP crosses the divide even when apps do not.
3. App and hardware distribution. Xiaomi sells ~70% of its phones outside China, carrying Google’s AI to India, Southeast Asia and Latin America on Chinese screens; Apple serves ~500M Chinese users; TikTok, Temu and Shein carry Chinese ranking models west. The divide is political, not structural.

On the headcount metric the Western cluster influences roughly three times as many distinct humans as the six-entity Chinese cluster, and the gap is mostly the non-China internet that Google, Meta and Nvidia already saturate. The Chinese cluster’s deepest influence is over its 1.1B domestic users, where it is near total and the Western cluster is thin.

Adding ByteDance

ByteDance was not among the fifteen entities. A second round put it to the same panel as a seventh Chinese entity. Eleven of twelve houses delivered.

ByteDance alone reaches a median 2,450M humans, range 2,100 to 2,900: direct 2,150M (TikTok 1,300 to 1,600M overseas, Douyin ~800M domestic, with CapCut, Toutiao, Doubao and Feishu mostly inside those two), indirect 300M and dependency 150M through Doubao and Seed models served on Volcano Engine and phone OEMs, the figure every model marked least sure. In the fifteen-entity table it ranks 5th, range 3 to 8, directly behind Microsoft and ahead of Apple, OpenAI and Amazon.

The cluster nearly doubles and all the growth is overseas. Domestic reach was already capped at China’s netizen base, so Douyin adds 0 to 50M new people at home. TikTok adds roughly 1.2B people abroad, but they live inside Android, iOS and Meta’s apps, so the overlap with the US cluster rises by about a billion and China’s unique reach barely moves.

Under the silicon reading, nine of eleven models put ByteDance’s inference at 50 to 75 percent Nvidia, which leaves China’s unique reach at ~400M: the domestic users served on Ascend and Cambricon. GPT-6 Astra Pro argued zero, since mixed serving puts any user on Nvidia silicon some of the time. Qwen argued 900M, crediting the Ascend ramp. Gemini’s summary stood for the panel: ByteDance doubles China’s consumer distribution and adds nothing to its sovereign autonomy.

With ByteDance in, the US cluster is ~4,900M, the Chinese cluster ~2,650M, and the union is still the internet at ~5,800M. China’s headline is 54 percent of the US figure; its unique footprint is 8 to 20 percent of it, depending on whether the chips count.

5. Verdicts on the two prior drafts

Most overrated in the prior drafts, by panel vote: Meta AI as the largest consumer AI (5 models), then Nvidia’s 6B framing (2), OpenAI’s partner attribution (1), Tencent’s Hunyuan-specific reach (1). Most underrated: Anthropic (5 models), Nvidia’s dependency role (2), DeepSeek’s leverage (1), Tencent’s company-wide footprint (1).

6. The purchasing-power matrix

A headcount matrix weights a teenager on TikTok the same as a fund manager. This round weights by money, on nineteen entities: the sixteen above plus TSMC, ASML and Huawei. Twelve of thirteen houses delivered, Thinking Machines’ Inkling joining the panel. Five measures per entity, all panel medians with ranges:

• M1, purchasing power touched: the annual household consumption of the humans the entity’s AI reaches, in trillion USD. World household consumption is 65 to 70 trillion; the top income decile, about 800 million people, spends roughly half of it.
• M2, share of the world’s top income decile reached monthly.
• M3, the entity’s own AI-attributable revenue run-rate, billion USD.
• M4, valuation, billion USD.
• M5, money mediated: third-party commerce, ad auctions, cloud AI spend or chip sales that flow through the entity’s AI systems, billion USD a year.

The M1 ranges for Nvidia, TSMC and ASML run from near zero to over fifty trillion because the panel split on whether a chokepoint touches purchasing power at all: four models gave the fabs and the lithographer zero, seven gave them Nvidia’s full footprint. The medians sit where the majority did. Everything in this table is a model estimate; the valuations and Nvidia’s revenue are the only cells close to disclosed figures.

Who reaches the rich. Nvidia, Google, Microsoft, Meta, Amazon and Apple each reach a majority of the world’s top income decile. Apple reaches 52 to 76 percent, SpaceX, Tesla and X 8 to 40 percent, with most of the second inside the first. The claim that Apple plus Tesla and X reach 80 percent of the rich holds in the United States, around 90 percent on Gemini’s and Hunyuan’s derivations, and in Western Europe. It fails globally, landing at 62 to 77 percent, because China’s rich carry Huawei and drive BYD, X is blocked there, and India’s elite are under-represented on both. Qwen’s arithmetic: Apple 56 percent, Tesla 1 percent, X 7 percent, union 62 to 64. Reaching 80 percent of the rich requires Google or Microsoft, not Tesla.

What money changes. Three moves, agreed by ten or more models. TSMC and ASML enter from nowhere to the top third: zero audience, but every advanced AI dollar passes through a Taiwanese fab and a Dutch lithography line. ByteDance drops from 4th on humans to 7th or 8th: 2.45 billion people, young and low-spending, worth a seventh of an Apple user each on M1. Xiaomi, DeepSeek, Zhipu and Kimi fall to the floor for the same reason, large headcounts in low-consumption geographies. Apple rises three places on purchasing power and one on the combined rank, because its 1.5 billion users skew rich but it mediates little third-party money. Alibaba and Amazon hold rank through the money they mediate rather than the people they reach: Alibaba’s AI-recommended GMV, 700 billion at the median, is the largest M5 on the table. OpenAI and Anthropic keep their humans rank and gain nothing on money: enterprise API pricing lifts revenue per user, but neither owns a platform that other people’s money flows through.

Musk, on this metric. SpaceX, Tesla and xAI sit 14th on humans and 15th on the combined rank, with a valuation of 1.24 trillion that is the fourth-highest on the table. The panel’s reading is that the valuation prices the 2030 option, Starlink as a pipe and Optimus as a workforce, and that neither is AI reach today. The purchasing-power weighting does lift it above Xiaomi and DeepSeek, which the headcount ranking did not.

Method

Twelve frontier models from twelve AI houses answered one identical prompt through OpenRouter on September 19th 2026: GPT-6 Astra Pro (OpenAI), Gemini 3.8 Flash (Google), DeepSeek V4 Pro 0813, Qwen 3.8 Max 0902 (Alibaba), Hunyuan HY4 preview (Tencent), Nemotron 3 Ultra (Nvidia), GLM 5.3 (Z.ai), MiniMax M3, Kimi K3 (Moonshot), Muse Glimmer 30B (Meta), Mistral Medium 3.5, MiMo V2.5 Pro (Xiaomi). No Anthropic model sat on the panel. Each model was asked for a ranked table with three reach layers, a derivation per figure, a confidence grade, and a forced verdict on eight claims from the two prior drafts. Rankings were aggregated by median rank; totals by median with range. Xiaomi’s answer was cut before the verdicts, so the verdict counts in section 5 are out of eleven. A second round put ByteDance and the seven-entity Chinese cluster to the same twelve houses; eleven delivered and Mistral’s runaway output was discarded. A third round scored nineteen entities on purchasing power, top-decile share, AI revenue, valuation and money mediated, with Thinking Machines’ Inkling added as a thirteenth house; twelve delivered and Mistral looped again. Total spend across the three rounds 4.34 USD. The full record of what was kept, rejected and left open is in the companion review record.

The calls above are Florin’s; the prompt, the research runs, the aggregation and the drafting are his Claude agent’s.

Review record of the matrix

Every finding from the twelve-house review of the Structural Reliance Matrix, kept or rejected, with the reason.

First round, twelve houses

Asked for one answer from each of OpenAI, Google, Tencent, DeepSeek, Alibaba, Nvidia, Z.ai, MiniMax, Moonshot, Meta, Mistral and Xiaomi, on one identical prompt: rank fifteen entities by direct, indirect and dependency reach, derive every figure, grade confidence, and return a forced CONFIRMED / WRONG / UNVERIFIABLE verdict on eight claims lifted from two earlier drafts. Slugs were verified live against the OpenRouter model list before spending. Anthropic was left off the panel on purpose: the synthesis was going to be drafted by a Claude model, and a house cannot review its own entry.

Delivered on the first pass: GPT-6 Astra Pro, Gemini 3.8 Flash, Nemotron 3 Ultra, MiniMax M3, Kimi K3, Mistral Medium 3.5.

Failed on the first pass:

• Meta Muse Spark 1.3 returned HTTP 403. The error body names the cause: the model is gated behind an 18+ attestation on the OpenRouter account. Substituted Meta Muse Glimmer 30B, which is ungated. One click at openrouter.ai/settings/preferences unlocks Spark for a re-run.
• Hunyuan HY4, DeepSeek V4 Pro, Qwen 3.8 Max, GLM 5.3 and Xiaomi MiMo V2.5 Pro all returned an empty body at 14,000 output tokens: every token went to reasoning and none to the answer. Re-run with reasoning effort capped at medium and a 30,000-token budget. Four recovered on the second attempt. Xiaomi recovered on the third, after 1,003 seconds and 22,000 reasoning tokens, with the table and derivations complete and the answer cut before the verdicts. Twelve houses asked, twelve delivered.
• MiniMax M3 and Kimi K3 filled their budget inside the falsification section and never reached the JSON block. Their rankings were parsed from the section 1 table instead.
• Hunyuan HY4 still hit the length limit after 25,000 reasoning tokens, with the table and cluster totals complete and the verdicts cut at D6.

Cost of the round including re-runs: 1.97 USD. GPT-6 Astra Pro alone was 1.15 USD of that, on 41,600 tokens.

Panel behaviour

Self-reported knowledge dates ran from June 2024 (GPT-6 Astra Pro, Nemotron 3 Ultra, Hunyuan HY4) to March 2026 (Gemini 3.8 Flash). DeepSeek and Mistral reported 2026-09, which is today’s date and almost certainly the prompt echoed back; their 2026 figures were weighted as extrapolations, not observations.

GPT-6 Astra Pro marked all fifteen entities LOW confidence, put OpenAI at 250M and 10th, and refused to convert Microsoft or Apple installed bases into OpenAI people. It was the only model to reject D4. The harshest reviewer of OpenAI on the panel was OpenAI’s.

Meta Muse Glimmer confirmed D2, Meta AI’s 1.2B and “largest consumer AI”, and ranked Meta 2nd, the highest anyone placed it. Two other models also confirmed D2, so it is not alone, but it is the house voice on its own claim.

DeepSeek V4 Pro ranked DeepSeek 10th, above Xiaomi and xAI, and called D7 WRONG in its own favour. Qwen 3.8 Max did the same for DeepSeek, not for Alibaba, which it ranked 9th at 880M, below the panel median.

Hunyuan HY4 gave Tencent the lowest total on the panel, 1,100M and 7th. Kimi K3 ranked Kimi last. GLM 5.3 ranked Zhipu 14th. Xiaomi MiMo ranked Xiaomi 10th at 500M, the panel median rank and the lowest Xiaomi total on the panel, and stripped its own IoT device counts from the people figure. Nemotron ranked Nvidia first but assigned it zero direct and zero indirect reach, all 4.5B as dependency, and was the only model to put Microsoft outside the top five, at 9th.

Mistral Medium 3.5 reported a Western cluster of 12,000M humans and a Chinese cluster of 6,000M. Its cluster totals were discarded. Its verdict line named Anthropic “most overrated” while its own table gave Anthropic 750M and 12th, the second-highest Anthropic figure on the panel; the verdict was recorded as internally inconsistent and not counted.

Confirmed and kept, from the panel

• Nvidia first. Ten of eleven put it at rank 1; MiniMax put Google first. Kept at 1, with the label changed: its people are a dependency layer, not an indirect audience.
• Google above Meta. Nine of eleven. Kept.
• Microsoft 4th, Apple 5th, OpenAI 6th, Amazon 7th, Tencent 8th, Alibaba 9th, Xiaomi 10th. Median ranks with no more than two models dissenting on each. Kept.
• Anthropic, DeepSeek and xAI as one tie at 11 to 13. All three had median rank 12. Ordered by median total: DeepSeek 520M, Anthropic 420M, xAI 350M. The order inside the tie is an editorial call, not a panel finding.
• Zhipu 14th, Kimi 15th. Every model.
• D3, D5, D6, D8 rejected as stated by eight or more models each. D4 confirmed by ten. D8 unanimous.
• The nested-doll shares: OpenAI 40 to 70% borrowed distribution; Anthropic 50 to 85% cloud-mediated; Apple ~25% OpenAI behind the consent modal; Nvidia ~70% other entities’ users relabelled. Kept as ranges because no two models agreed on a point figure.
• TSMC and ASML added below Nvidia as chokepoints with no audience of their own. GLM and GPT-6 raised it unprompted; the prompt had only mentioned them in passing.

Ranking changes on panel consensus

• OpenAI moved from 6th to 6th, unchanged, but its total was cut from the draft’s “1B+ via Microsoft and Apple” to a 1.0 to 1.8B range with the borrowed share stated. The claim survived; the framing that all partner users are OpenAI users did not.
• Tencent moved from 8th to 8th, unchanged, with the “entire Chinese population” figure replaced by 50 to 80% of WeChat MAU.
• Anthropic moved from 14th to 12th. Every model rejected 14th; none put it above 11th.
• DeepSeek moved from 12th to 11th on median total. Five models kept it at 12, four moved it up. The cost-disruption argument was recorded as a different dimension, not as headcount.
• xAI/Tesla/SpaceX moved from 11th to 13th. Only one model kept it at 11; the spread ran to 15th.

Rejected, with reason

• Nvidia at 6B. Exceeds the internet population on every model’s baseline (5.5 to 5.8B). Replaced with 4.3 to 5.5B.
• Nvidia’s “unique” reach as its headline. GLM’s 500 to 800M is right on its own terms and was recorded in the nested doll, but the matrix measures exposure, not counterfactual removal. Kept 4.5B as the headline with the relabelling stated.
• Starlink as AI reach. Three models counted it, four refused. Excluded: connectivity is not AI exposure.
• IoT device counts as people. Every model stripped them for Xiaomi. Excluded.
• Meta AI as largest consumer AI. Google’s Search and Android surface reaches 2 to 3B on nine models’ derivations. Rejected on “largest”, kept on the ~1B figure.
• DeepSeek at 7th on a leverage-weighted metric. GLM’s proposal. Rejected for this matrix because no other entity was scored on leverage; recorded as the right next matrix to build.
• Mistral’s cluster totals. Discarded, above human population.
• Gemini’s D3 CONFIRMED. The only model to confirm it, against ten; its own table gave Tencent 1,350M, not the population.

Second round, ByteDance and the China cluster

Florin asked what China’s de-duplicated total would be once ByteDance, missing from the fifteen, was added. The same twelve houses got a 435-word prompt asking for ByteDance’s three layers, the seven-entity Chinese cluster total, the overlap with the US cluster, and China’s unique reach under two readings: apps and clouds only, and counting Nvidia and TSMC silicon dependency. Reasoning was capped from the start this time, and no model returned an empty body.

Delivered: eleven. Mistral Medium 3.5 hit the length limit at 98,000 characters of repeating text after 25,000 reasoning tokens and was discarded. Round cost 0.91 USD, of which GPT-6 Astra Pro was 0.47. Both rounds together 2.88 USD.

Panel behaviour: no house had a stake in ByteDance, and it showed. The eleven ByteDance totals sat inside 2,100 to 2,900M, the tightest agreement on any entity in either round. The disagreement moved to the silicon reading. GPT-6 Astra Pro put China’s unique-with-silicon reach at zero, arguing that mixed serving puts any user on Nvidia silicon some of the time, and objected that TSMC is Taiwanese and calling it US-cluster broadens the definition. Qwen put it at 900M, crediting the Ascend ramp, and was the only model above 600M. MiniMax gave the US overlap as 800M against a panel median of 1,750M, the one outlier on that cell. Knowledge dates: three models reported 2026-09, today’s date, again read as the prompt echoed back.

Confirmed and kept: ByteDance at 2,450M, rank 5 of 16, median of eleven. Chinese cluster 2,650M with ByteDance. Overlap with the US cluster 1,750M. Unique beyond US 950M on apps, 400M with silicon. ByteDance inference 60% Nvidia, range 45 to 75, from nine models inside that band.

Rejected, with reason: Mistral’s output, degenerate. MiniMax’s 800M overlap, an outlier against ten models at 1,250M or above; its cluster total was kept. GPT-6’s TSMC objection was recorded rather than adopted: the matrix treats TSMC as a chokepoint under Nvidia, not as a US entity, and the silicon reading is explicitly the one where the chokepoint counts.

Third round, weighting by money

Florin’s objection was that the headcount ranking underestimates Musk and Apple, because their users hold most of the world’s purchasing power. Rather than adjust the ranking by hand, the panel was asked to weight it. Nineteen entities, the sixteen plus TSMC, ASML and Huawei, five money measures each, and a combined humans-and-money rank. Thinking Machines’ Inkling joined as a thirteenth house to match the AI 100 reviewer roster. Twelve delivered. Mistral Medium 3.5 looped to 67,000 characters and was discarded for the second round running. Round cost 1.46 USD, GPT-6 Astra Pro 0.81 of it. All three rounds 4.34 USD.

Panel behaviour: the split was not by house this time but by philosophy. Four models, Nemotron, Inkling, Glimmer and MiniMax, gave TSMC and ASML zero purchasing power on the grounds that a fab has no consumers; seven gave them Nvidia’s entire footprint on the grounds that every AI dollar passes through them. Medians landed with the seven. GPT-6 Astra Pro rejected the 80 percent claim on arithmetic: a summed 82.5 percent is not a union, and a union of 80 needs overlap under 2.5 points. Nvidia’s Nemotron gave Nvidia a 96 percent top-decile share, the highest self-score in any round; Qwen gave Nvidia the same figure, so it was kept. Three models again reported today’s date as their knowledge date.

Confirmed and kept: Nvidia and Google 1 and 2 on every measure. Apple 4th on purchasing power, from 7th on humans. ByteDance down from 4th to 7th or 8th. TSMC 8th and ASML 12th combined, from no rank at all. Alibaba’s 700 billion of AI-recommended commerce as the largest money-mediated figure on the table. Apple 52 to 76 percent of the top decile, Tesla and X 8 to 40 with heavy overlap, union 62 to 77 globally and about 90 in the United States.

Ranking changes on panel consensus: the combined rank used for the Companies tab draft is Nvidia, Google, Meta, Microsoft, Apple, Amazon, ByteDance, TSMC, OpenAI, Tencent, Alibaba, ASML, Huawei, Anthropic, SpaceX+Tesla+xAI, Xiaomi, DeepSeek, Zhipu, Kimi. Meta and Microsoft tied at median 5 and were ordered by humans. OpenAI, Tencent and Alibaba tied at median 10 and were ordered by mean rank, then by the purchasing-power rank.

Rejected, with reason: the 80 percent claim as a global figure, on eleven of twelve derivations; kept as a US and Western Europe figure. Mistral’s output. The four zero scores for chokepoints, recorded as a philosophy rather than adopted: the matrix already treats TSMC and ASML as construction under Nvidia, and reading them as touching money is the reading the substitution-cost matrix needs. Musk moving up on “potential”: the panel priced the 1.24 trillion valuation as the 2030 option on Starlink and Optimus and refused to count it as reach today.

Still open

• The money definitions. M1 and M5 rest on the panel’s income-mix assumptions and on GMV and ad-flow estimates no company discloses by AI share. Florin has to decide whether the Companies tab shows M1 alone, M1 with M5, or the combined rank, before anything is published.
• Huawei. First scored this round, 13th combined, on twelve estimates with no Huawei model on the panel and HarmonyOS’s device base unverified.
• ByteDance’s Nvidia inference share. Eleven estimates from 45 to 75 percent, none grounded in a disclosure, and the whole silicon reading turns on it. If Ascend serves most of Douyin by late 2026, China’s unique reach doubles; if Nvidia does, it halves.
• Llama downstream reach. Panel range 100 to 900M for third-party deployments, the widest disagreement on any single cell. No model had a census and none claimed one.
• Bedrock end users. 500M to 1.8B. This is Amazon’s dependency column and Anthropic’s biggest layer, and it rests on no disclosure from either company.
• Xiaomi’s verdicts. MiMo V2.5 Pro delivered its ranking on the third attempt but was cut before section 5, so the D1 to D8 tallies are out of eleven houses, not twelve.
• Meta’s frontier voice. Muse Spark 1.3 is one account setting away; the panel heard Meta through Glimmer 30B.
• Model knowledge dates. Three of eleven models reason from a mid-2024 base and project forward. Their 2026 figures are scenario estimates, and they said so.
• The substitution-cost matrix. Every model that reached the nested doll said the headcount metric saturates at the top three and cannot separate Nvidia, Google and Meta. Substitutability would. Not built here.

The calls above are Florin’s; the prompt, the research runs, the aggregation and the drafting are his Claude agent’s.

Fourth round, the top 20 under the stance-and-significance rule

Florin’s decision after the money round: the people list is not a reach ranking. It declares the co-evolutionary stance, ranks on significance with symmetric treatment of potential and of harm, keeps heads of state on the magnitude of their instruments, and states the two conflicts (the editor as an Amazon seller, the drafting agent as Anthropic’s model). Twenty placements went to thirteen houses with the rule in the prompt and one instruction: judge on this function, not yours.

Twelve delivered; DeepSeek returned an API error. Round cost 0.70 USD, all four rounds 5.04 USD. No placement reached the eight-dissent threshold. Weakest: Ternus and Kavukcuoglu at five AGREE and four DISAGREE each, both on tenure; the Amodeis at six AGREE with four CANNOT-TELL, mostly on 2026 facts past the models’ cutoffs.

Panel behaviour: the houses did what was asked and judged on the declared rule. The most-named finding, six of twelve, was an asymmetry the drafting agent had introduced: the ByteDance entry carried an attention-economy harm clause with no source while Meta’s carried Reuters. Two houses caught the Musk caption calling FSD and Starlink “in progress” when both are realized. Muse Glimmer’s NO verdict rested on its own January cutoff and was discounted. Mistral proposed LeCun for 16, its first usable output in three rounds.

Kept: the rule, and nineteen of twenty placements pending Florin’s choice for the seat at 16. Fixes adopted: source or drop the ByteDance harm clause; scope Musk’s in-progress tag to Optimus and Grok; state Ternus’s tenure in the entry. Recorded, not adopted: Apple’s App Store as an attention footprint, since the rule scores AI conduct.

Full vote table in `Part1_reorder_proposal_2026-09-19.md`.