The Capital Loop

Part 1 showed the adoption curve: AI reaching toward most of the human race through cheap phones, free tiers, and satellites. This piece follows the other curve, the one financing that arrival. I expected a normal capital stack. What the filings show instead is a circle: the same handful of balance sheets funding each other, buying from each other, and booking each other's spending as revenue. None of it hidden. All of it disclosed. Almost none of it read as one picture.

No advice in this piece. No verdicts. The ledger, and what it shows.

The concentration, measured

In Q1 2026, global venture funding hit a record $297B, and 4 companies, OpenAI, Anthropic, xAI and Waymo, took $188B of it. That's 63.3%. OpenAI and Anthropic alone took 43% of H1's $510B. Underneath the record totals, seed deal count fell 30% year over year while seed dollars rose 31%: bigger checks, fewer founders.

Those are the headline numbers. The structure underneath them is stranger.

The loop, drawn with real wires

The money doesn't flow from investors to companies to customers. It moves in circles, and this isn't my framing: Bloomberg's own deal-map headline reads "How Microsoft, OpenAI and Nvidia Keep Paying Each Other." The wires, per public filings and reporting:

Nvidia and its buyers. Nvidia put $30B into OpenAI's March 2026 round, with a $100B letter of intent behind it (still not definitive as of late 2025, and labeled that way everywhere it appears here). It also holds roughly 13% of CoreWeave after a $2B January investment, and CoreWeave committed $6.3B in GPU purchases. UBS estimates the OpenAI arrangement alone represents about 13% of Nvidia's 2026 revenue. Meanwhile Nvidia's latest 10-Q shows 3 customers accounting for 44% of its first-half data-center revenue, the largest at 16%, down from 23% a year earlier.

Microsoft and OpenAI. $13B invested since 2019. After the October 2025 restructure, Microsoft holds 27% of OpenAI's for-profit arm, a stake worth about $230B at the March 2026 round price. Flowing back: OpenAI committed $250B in incremental Azure compute purchases. $13B out, $250B of committed spend returning.

Amazon, Google and Anthropic. Amazon has committed roughly $33B across 3 rounds (with $20B contingent on milestones) and Google roughly $40B (with $30B contingent). Flowing back: Anthropic committed $100B+ to AWS Trainium over 10 years and runs on more than 1M Trainium2 chips today, plus Google TPU capacity. Both clouds fund the lab. The lab buys both clouds.

And the round that closes the circle: OpenAI's $122B raise at an $852B valuation in March 2026 included $50B from Amazon ($35B contingent on an IPO or an AGI milestone) and $30B from Nvidia. The company's 2 biggest classes of supplier are now among its largest shareholders.

The commitment wall

OpenAI's committed compute spend runs to roughly $1.4T through the early 2030s. Itemized where reported: Broadcom $350B, Oracle $300B, Microsoft $250B, Nvidia $100B (letter of intent), AMD $90B, AWS $38B, CoreWeave $22B.

Against that: OpenAI's entire audited 2025 revenue was $13.1B, with a $20.9B operating loss. To be fair to the bull case, the revenue is moving fast: $2B a month in March 2026, roughly $3.3B a month (about $40B annualized) by August, with 900M+ weekly ChatGPT users approaching a billion. That growth is real. The commitments still assume it compounds for a decade.

Who pays for the build

Hyperscaler capex: $388B in 2025, projected above $600B for 2026, a jump of more than 50%, with roughly 75% targeting AI. Amazon around $200B, Alphabet around $180B, Meta around $125B, Microsoft around $115B.

Increasingly, debt pays for it. The hyperscalers raised $108B in debt in 2025, with roughly $1.5T of issuance projected by Goldman and Morgan Stanley analyses, because aggregate capex now runs above internal cash flow. Alphabet's free cash flow went negative in Q2 2026, the first time since 2004. An August 2026 WSJ analysis, a single analysis and labeled as such, counted about $3T in off-balance-sheet AI obligations across 9 major firms: $1.2T in leases not yet commenced (4x in a year) plus $1.9T in purchase commitments.

Zoom out to the index: Goldman calculates that Nvidia and Micron, 2 tickers, account for a third of all S&P 500 earnings growth in 2026. Nvidia became the first company to touch a $5.5T market cap. Global AI investment crosses $1T this year.

What the top of the market costs

• Anthropic: $965B valuation (May 2026, after a $65B Series H), against a reported $47B ARR. Roughly 21x.
• OpenAI: $852B (March 2026), against ~$40B annualized by August. Roughly 21x on current run rate, ~35x at close.
• xAI: valued at $250B inside its February all-stock merger into SpaceX, the largest merger in history at $1.25T combined, on 2025 revenue of $3.2B. Roughly 78x.
• CoreWeave (public): a $104B revenue backlog as of June 30 against $5.1B of 2025 revenue, with $25B+ of new commitments added early in Q3.

Minted, squeezed, absorbed

Now the stat that looks off-theme until you wire it in: 250 new unicorns by Aug 15, 2026, already past 2025's full-year total of roughly 190. Read alone, that's a thriving competitive field. Read against 3 adjacent facts, it's something else:

1. Minted: per Crunchbase, virtually every early-stage unicorn of the last few quarters is AI-focused, and much of the capital minting them comes from the same concentrated pool, including the giants' own venture arms.
2. Squeezed: the entry rung beneath them narrowed 30% in a year. Fewer founders funded than any recent year.
3. Absorbed: the cycle's biggest reported startup exits, Inflection, Character.AI, Windsurf, Adept, Scale, Groq, Hugging Face, each went to one of the largest platforms. Big tech spent $20B+ on reverse acquihires between 2024 and 2026, deals structured so no merger review fires; senators have called them de facto mergers. And in the 9 months through September: xAI folded into SpaceX at $250B, Nvidia took Groq's assets for about $20B, and Nvidia agreed to buy Hugging Face, the hub hosting 3M open models and 18M developers, for $12.9B.

Call that a healthy pipeline or call it a farm system. The numbers are identical either way. The reading is yours.

Where the two curves meet

Part 1 says the next 4 billion users arrive through cheap phones, free tiers, and satellites. This piece says the free tiers, the models, the clouds, and increasingly the open-model hub itself sit on a handful of balance sheets that fund each other and buy from each other.

Those 2 facts sit in tension, and I'm going to let them sit. The same open-weight models that made intelligence 60x cheaper are the reason free tiers exist at all. The same concentrated capital pouring $600B a year into data centers is what makes coverage and capacity real. Whether this structure ends as the most productive buildout since electrification or as the vendor-financing chapter of a future case study is exactly the argument the data doesn't settle yet. What the data does settle is what the structure IS. That's what this piece and the film put on the table.

And if this structure feels new, it isn't. In part 3, The Longest Warning, I trace the whole thing back: 183 years of documented AI, the government programs that ran decades before the public heard of them, and the cautions that came from the builders themselves in every generation. The loop in this piece has a 70-year-old birth certificate, and that story deserves its own telling. What to DO inside all of it, as a builder, an investor, or a country, comes after that.

Sources: Bloomberg (OpenAI round; circular-deals graphic) · CNBC (Nvidia/Hugging Face; SpaceX/xAI; CoreWeave earnings) · TechCrunch · Crunchbase (quarterly funding, unicorn board) · Goldman Sachs research (AI investment; Nvidia/Micron earnings contribution) · WSJ (off-balance-sheet analysis, Aug 2026) · UBS (Nvidia revenue-share estimate) · CoreWeave investor relations · SEC filings (Nvidia 10-Q and 8-K; SpaceX S-1) · Amazon, Anthropic and OpenAI company announcements.

The full series essay is on Stay Level.