At 09:14 UTC, the only number that mattered crossed the tape: 1%.
Eric Schmidt holds roughly 1% of Anthropic, per Crypto Briefing. That is the entire verifiable payload. One data point. No filing, no valuation anchor, no timestamp, no share-class definition, no confirmation of whether the position sits on his personal book or inside a family office or an SPV.
A percentage without a denominator is not a fact. It is a range.
Run it both ways. At Anthropic's April 2022 Series B post-money of roughly $4.1 billion, 1% is $41 million. At the ~$350 billion mark reportedly attached to the Microsoft/Nvidia talks in November 2025, 1% is $3.5 billion. Same digit. Eighty-five times apart. Anyone quoting "Schmidt's multibillion Anthropic position" without naming the mark is quoting the ceiling of a range they cannot defend.
That gap is the story. Not the stake.
Context: a procurement schedule with a share certificate stapled to it
The ladder, public-report basis:
- Series A, May 2021 โ $124M raised, ~$550M post.
- Series B, April 2022 โ $580M, ~$4.1B post. FTX/Alameda led.
- Series C, May 2023 โ $450M, ~$5.0B post.
- Series D, September 2024 โ $750M, ~$18.4B post.
- Series E, March 2025 โ $3.5B, ~$61.5B post.
- Series F, September 2025 โ $13B, ~$183B post. ICONIQ led.
- Microsoft/Nvidia, November 2025 โ reported up to $5B + $10B, mark near $350B.
Two things break before any modeling does. First, velocity: $61.5B to $183B in six months, then $183B to ~$350B in roughly two. Second, and more consequential, the identity of the marginal buyer changed. Series A and B were financial capital โ Tallinn, Moskowitz, Schmidt, FTX. The last two marks are strategic capital: cloud vendors and a silicon supplier who also sell Anthropic the compute that Anthropic resells as tokens.
That is not a valuation curve. It is a supply agreement with an equity kicker bolted on.
Core: the dilution math, the multiple, and the loop
What 1% actually cost. Compounding new-capital over post-money across every round, Series A holders retain roughly 60โ65% into the current cap table. To still hold 1% today, Schmidt's Series A nominal had to be near 1.5โ1.7% โ an outlay around $8Mโ$10M, or $3Mโ$5M if he pro-rata'd only partially. That implies 100โ300x on cost at the top mark. Confidence C: the share-class basis is unpublished, and if the 1% was bought in a later round or taken down in secondary, the multiple collapses by an order of magnitude.
The multiple. At the $183B Series F mark against a reported ~$5Bโ$7B annualized run-rate, you get 26โ35x revenue. Defensible for a company compounding at that slope. At $350B against the same denominator, you get ~50x โ and that only clears if the 2026 revenue line lands in the $20Bโ$26B range implied by the mark. That is a strong assumption wearing a base case's clothes.
The loop. Here is where the audit background does the talking. In 2017 I spent four months inside the Hard Hat Protocol contracts and caught an integer overflow in the staking logic before mainnet. The lesson that survived: the number on the dashboard is downstream of the code that computes it, and the code is downstream of whoever funded it. In 2020 I reverse-engineered Uniswap V2's rebalancing to build a volatility exploit simulator, and learned the second version: a price set by a participant who also supplies the liquidity is a quote, not a price.
Apply both to this tape. Anthropic reportedly commits roughly $30 billion of Azure compute. Microsoft invests. Nvidia invests and sells the silicon. Amazon invests and sells Trainium. Google invests and sells TPU. Investor, supplier, and โ through Bedrock and Vertex resale โ channel, are the same counterparty, three times over. In DeFi we have a name for an asset whose price is set by the entity that also provides its liquidity and pays its yield: reflexive collateral. It works until the incentive to defend the mark disappears. Then it works in the other direction, faster.
The mark is being set by strategic consideration, not revenue verification. That is the most important sentence in this piece, and it is absent from the source.
Revenue quality. Enterprise and API first, consumer subscription second. The wedge is coding and agent workloads, where Claude has held a top-tier position on SWE-bench-class evaluations and supports Sonnet-level pricing at roughly $3/$15 per million tokens. The problem is not the slope, it is the margin structure. Large accounts arriving through AWS Bedrock and Google Vertex pay a channel take, on top of inference cost, on top of a compute commitment. Three cuts before gross profit. Concentration is undisclosed: if a handful of coding tools and a few large enterprises carry a disproportionate share, a single self-hosted substitute or a price war transmits straight to the revenue line.
Competitive position. Frontier capability gaps have compressed to quarters, not years. Open weights โ Llama derivatives, DeepSeek, Qwen โ keep compressing price at the mid and low end. Anthropic's moat is not parameter count. It is enterprise trust in regulated verticals plus multi-cloud compute assurance, and neither of those is a permanent asset. The 1% is effectively a call option on converting developer mindshare into platform lock-in before the agent cycle matures.
1% of economics, 0% of control. Anthropic is a Delaware public benefit corporation with a Long-Term Benefit Trust that progressively accrues the right to elect a board majority. Schmidt's 1% is economic only. It carries no meaningful governance. The popular read โ 1% stake, 1% say โ is wrong by construction.
The exit is not the mark. ROFRs, lockups, and the standing two-digit discount on frontier-AI paper at Forge and Hiive mean $3.5B on paper is not $3.5B in hand. I built an NFT arbitrage bot in 2021 that held a 200ms latency edge across OpenSea and LooksRare and cleared โฌ50,000 in six weeks. Floors are illusions until the bot sees the spread. The mark is what a strategic buyer says an asset is worth at a size nobody can actually trade.
Contrarian: the safety capital is capturing the acceleration
The unreported angle is not financial. Anthropic's early backers arrived with explicit existential-risk framing โ Tallinn, Moskowitz, Schmidt. Capital allocated out of concern about capability acceleration is now the best-performing capital in the AI complex. Against a Forbes-basis net worth of $30Bโ$35B, a single non-controlling AI position at 5โ10% of personal wealth is an abnormally concentrated exposure for an ultra-high-net-worth book, and it creates a constituency with a financial interest in not slowing down.
In 2022 I spent two weeks dissecting Anchor's yield model and published two days before the collapse. The mechanism was circular yield โ a protocol paying for its own TVL. The Anthropic structure is not fraudulent; real revenue exists. But the topology rhymes. Honest framing: same shape, better collateral.
For crypto readers the actionable version is narrower and worse. In 2024 I ran a dashboard tracking IBIT wallet flows off chain explorers โ institutional accumulation was legible in near real time. Private AI marks are legible quarterly at best, and only when someone leaks them. AI-adjacent tokens re-rate off those marks with a lag of days. "Decentralized compute" and "agent" narratives get a bid they did not earn. Those tokens have no cap table, no Azure commitment, no revenue line. They have a chart that correlates to a headline. Speed is the only metric that survives the crash โ and in a bear tape, the reflex bid on narrative proxies is the first thing to go bidless.
Takeaway
Watch four prints, not the headline: secondary transactions on Forge and Hiive, which will price the real discount; whether the Microsoft/Nvidia round closes and dilutes the 1% below 1%; Anthropic's disclosed revenue concentration from cloud-channel resale; and the $30B Azure commitment as a fixed liability sitting against a variable revenue line. If the revenue line misses and the commitment does not, the ~$350B mark becomes a quote with no bid behind it. Then the only question that matters is who is still standing when the spread finally prints.