Anatomy of a $122B Round: How OpenAI Bought Itself a Decade
The largest private fundraising round in the history of capitalism closed in March 2026 at north of $122 billion. The buyer is OpenAI, the seller is a syndicate of investors led by the company's existing partners, and the product is a four-to-six-year runway to build out the compute capacity that the company's CFO, Sarah Friar, believes will define the next decade of the technology industry. Speaking on the allin podcast on 4 June, Friar described the round in language borrowed from financial engineering rather than product launches: "maximum optionality."
That framing matters. The conventional read on a private round this size — roughly four times the 2019 Saudi Aramco IPO of approximately $30 billion — is that it is a bridge to a public offering. The market is primed for OpenAI to file an S-1 any day, especially after rival Anthropic was reported to have done so confidentially. Friar's response, on camera and on the record, was to push the conventional read off the table. An IPO, she said, is "a milestone. It is not a destination."
What follows is the anatomy of how the round was structured, what the company is actually buying with the money, and why the public offering is being treated as a fundraising event rather than a graduation ceremony.
A round that redraws the scale
The $122 billion number is, on Friar's telling, not the kind of capital raise the venture model was built to process. She invoked the Aramco IPO — the previous record holder at roughly $30 billion — only to point out that the comparison runs the wrong way around. A private round four times larger than the largest IPO in history implies that whatever is being funded cannot be financed through public markets at the speed, or with the patience, that the underlying buildout requires.
The thesis is that the company is buying time, not buying an exit. The compute commitments OpenAI is making — one-gigawatt campuses in Michigan and Texas, and a build-to-suit arrangement with SoftBank Energy in Texas — take three to five years to come online. Sam Altman, on the same day as the podcast aired, was in Saline, Michigan cutting the ribbon on a 1GW data center that will not produce useful compute until late 2027 or early 2028. The cash is not a war chest. It is a down payment on a build cycle that begins now and is settled in tokens, not dollars, in 2030.
This is the part of the analysis where the conventional IPO narrative frays. If the round were merely a bridge to a public offering, the company would have filed by now and let the S-1 process price the stock. It has not. The signal is that the company is happy to keep paying the private-market premium for the option to stay private through the build cycle.
The IPO as milepost, not summit
Asked directly whether she would let Anthropic file first, Friar did not blink. "The market is a weighing machine, not a popularity machine. No one remembers who went first, Google or Yahoo, Lyft or Uber." The line is from a Benjamin Graham playbook, and it is being deployed for a specific reason: to defuse the narrative that filing sequence confers strategic advantage.
Friar's argument is that the public offering is a financing tool, not a validation event. "An IPO is a milestone. It is not a destination. Do not run your company as if that's some sort of destination. It's just another way to fund raise." Read against the Anthropic confidential filing, the position reads as a deliberate signal: the company is not in a race to a tape-cutting ceremony, it is in a race to the first quarter in which its compute capacity is no longer the binding constraint on revenue.
That is a different game. It is also a longer game. The companies that have lost the IPO-as-milestone framing — those that have optimised for the public debut and then had to manage the disappointment that follows — are well known. The companies that have framed the public market as one more capital pool have tended to do better. The question is whether the public-market investor base is patient enough to underwrite a buildout whose payoff is 2030 to 2032.
The gigawatt ledger
The economics that underpin the round are precise enough to be tested. Friar returned to a thesis she first articulated roughly 18 months ago: one gigawatt of AI compute produces approximately $10 billion a year in revenue. The ratio has held up, she said, even as the underlying cost of each gigawatt has risen with power and memory prices.
The counter-entry is the build cost. Standing up one gigawatt — land, power, shells, and chips — runs roughly $50 billion all-in. The Michigan data center, the Texas build-to-suit, and the multi-CSP contracts with Oracle, CoreWeave, Microsoft, GCP, AWS, and a layer of neoclouds all stack against that denominator. "We're going up that kind of vertical wall of demand right now and there's just not enough tokens available… in 2026 we still won't have enough compute," she said.
The math then becomes simple. If OpenAI wants to be a $200 billion-a-year revenue business in 2030, it needs roughly 20 gigawatts of compute online. At $50 billion a gigawatt, that is $1 trillion of capex — funded through a mix of partner capex (Microsoft, Oracle, CoreWeave, GCP), direct capex (the Texas SoftBank build-to-suit), and a series of multi-CSP contracts that move the bill from capex to opex on OpenAI's books. The $122 billion private round is the equity layer of a much larger capex stack.
"The fact that OpenAI has more compute on a per gigawatt is getting more expensive [power, memory]… but the intelligence we get on the other side… is more than making up for that," she said. The framing is that compute gets more expensive and more valuable at the same time, and the company that locks in the most gigawatts first captures the spread.
The Rubik's cube of compute
Two years ago, OpenAI ran on one cloud service provider — Azure — and one chip, Nvidia's. The risk surface was a single rectangle. In June 2026, the company runs a multi-CSP, multi-chip stack that Friar called a "Rubik's cube" of optionality. Nvidia remains a priority partner, with the next training run running on the Vera Rubin generation, but AMD and Cerebras are in the pipeline and a custom Broadcom chip is being built. On the cloud side, the company has contracts with Oracle, CoreWeave, Microsoft, GCP, AWS, and a layer of neoclouds she would not name.
The shift does two things at once. It moves capex off OpenAI's balance sheet and onto the balance sheets of partners — a lever that turns a $50 billion capital requirement per gigawatt into a multi-year operating lease. And it gives the company the ability to walk if any one partner raises prices, fails to deliver, or attempts to lock the company in. "My job as a CFO is to create optionality for not just this company but just this era that we're living in," she said.
The exception that proves the rule is the SoftBank Energy build-to-suit in Texas. That is direct infrastructure investment, not a partnership arrangement, and it reflects a calculation that some of the capacity OpenAI needs by 2030 will not be built by any CSP at the speed the company requires. The Texas project requires more capex than a CSP arrangement. It is being done anyway. The Michigan site carries its own public ledger: 2,500 union jobs, roughly $1 billion in Michigan taxes, and a $45 million commitment to Codex education credits, alongside a pledge not to raise local electricity rates.
The convergence play and the Anthropic counter
The structural argument Friar made against the narrative that Anthropic has "blown past" OpenAI in enterprise or developer share is that the company runs a single AI infrastructure layer serving multiple interfaces — ChatGPT consumer, Codex, enterprise API. "We're not trying to be a consumer company or an enterprise company — we're very much both. Our mission at OpenAI is AGI for the benefit of humanity, not for the benefit of humanity who can pay or for the benefit of humanity who live in an enterprise." Revenue is split roughly 50/50 between the two. The implication is that a competitor winning one segment is buying share inside a product category, not against the infrastructure underneath.
It is a defensible position, and it is not the only one. The counter-narrative — that Anthropic's enterprise wins, particularly in agentic coding and developer tooling, are durable because developer mindshare is sticky — has real evidence behind it. The five million Codex users reported on the podcast, up from near zero in January, suggests OpenAI is fighting back in the same category. Whether the unified-infrastructure thesis beats the focused-vertical thesis is a question 2027 will answer, not 2026.
The consumer side is being built for the same reason: defensive depth. ChatGPT has roughly 900 million weekly active users and at least 11% of search market share — a number Friar argued is far higher in effect because a ChatGPT conversation of 50 questions counts as one engagement, while a Google refresh also counts as one. The advertising opportunity she sketched — "if you know Google and Meta had a baby, it would be ChatGPT" — is being built on top of memory, context, and high intent, and will follow Google principles, with the best result un-sponsored and an ad-free tier always available. A new consumer device, an earpiece-like form factor developed in partnership with Jony Ive, is expected to be unveiled at the end of 2026 or early 2027. She has used a prototype and called it "natural" and "lovable."
The pricing math is the part that gets less attention. GPT-5.5 launched with a 2x price increase and a 20-30% net cost reduction per token — a 97% cost depreciation from GPT-4 to GPT-5.4 over roughly two years. Daily question volume by tier tells the same story: free users ask seven questions a day, the first paid tier roughly fifteen, the $20 Plus tier about three times free usage, and the Pro tier about eleven times. "If I was optimizing only for today, I would give every token to the API — order of magnitude more than to the consumer. However… we're playing our own game." The game is to be at scale in 2030. A year ago the company was pitching an agentic product to investors at $2,000 a month, a price Friar now calls "laughable in hindsight" given what developers are willing to pay.
The $122 billion is the cost of staying at the table until then. Whether the table exists in the same form when the tokens finally clear is a question no fundraise can answer.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://www.youtube.com/watch?v=TjrShuj_Zsg