OpenAI and Anthropic Aren't Short of Money. They're Short of Time.

The two most valuable private companies in the world have spent the past few weeks missing their own deadlines. Anthropic closed a $65 billion round in May, at a valuation ahead of OpenAI's, whose record $122 billion round had closed at an $852 billion valuation ^[1][2]. This month, Anthropic pushed its stock-market debut to the eve of the November 3 midterms, with the investor pitch now expected to start in mid-October at the earliest ^[3]. OpenAI had already slid its own debut to 2027 by late August ^[4], and last week it was fielding unsolicited proposals valuing it at $1.2 trillion, which it says it is not in formal discussions about ^[5]. Sam Altman has been blunt about the schedule.

now would be an "ill-advised" moment to go public — Sam Altman

Companies holding $187 billion of fresh capital between them do not usually miss their own deadlines. The pressure is not on the balance sheet. It is in the contract book, on the payment schedule inside it. The labs buy their computing capacity under take-or-pay contracts: they promise to pay for the capacity whether or not they end up needing it, the way a gym member pays for a year of sessions whether or not he ever shows up. Across the industry those promises run to trillions of dollars, and almost none of the money has moved yet, because the data centers are still being built; the commitments sit on the books as backlog. One credit analysis published in August compares the arrangement to subprime lending, calls the present stretch a teaser period of deferred costs, and puts the reset in 2027 and 2028, when backlog converts into billing. In its arithmetic, OpenAI's commitments approach $1.2 trillion, a figure that may exceed any plausible revenue path and leave the equity dependent on continuous refinancing ^[6]. Between now and then, nearly everything is negotiable except the date. Growing into that before the bills arrive would be easier if either side of the income statement would cooperate. Neither will. On the revenue side, prices are falling in a way that does not reverse: in May, DeepSeek turned a limited-time discount into a permanent 75 percent price cut on its V4-Pro, a model that costs 12 to 19 times less than the flagships from OpenAI or Anthropic for the same work ^[7]. One analyst explained why no one else gets to wait this out.

It is an efficiency gain being passed through. — Sanchit Vir Gogia

OpenAI matched the direction on August 1, cutting its Luna model by 80 percent, to 20 cents per million tokens of input, and named competition as the reason ^[8].

I think we’ll have a lot of ways we can help people get more value for less spend. — Sam Altman

On the input side, nothing is falling. Micron, which sells the memory these data centers run on, has signed multiyear Strategic Customer Agreements, meaning volume commitments with pricing floors, and expects memory to stay scarce through at least 2027, with 2027 tighter than 2026 ^[9]. The credit market, meanwhile, has picked its side: US tech firms have borrowed more than $600 billion globally for the buildout, and the builders of AI demand are now priced as riskier credits than the suppliers of its scarce inputs. TSMC's bonds trade at lower yields than Meta's, and SK Hynix's notes yield just nine basis points, nine hundredths of a percentage point, more than Amazon's ^[10]. April: the pivot. OpenAI shelved consumer projects like its video tool Sora, hired its first chief revenue officer, and set out to lift business customers from 40 percent of revenue to 50, while both labs ran at a loss and pointed themselves at public markets ^[11]. At month's end came the miss: internal revenue and user targets fell short, enterprise and coding share slipped to Google and Anthropic, and the selloff dragged Oracle, CoreWeave, SoftBank and Nvidia down with it. Sarah Friar, OpenAI's chief financial officer, attached the stakes to a number: without faster growth, she warned, the company might struggle to fund the roughly $600 billion of compute commitments it had signed, and she questioned whether it was ready to go public at all ^[12]. May: with the record $122 billion round closed at an $852 billion valuation, Friar had a phrase for what fundraising had become at a company that had set that record ^[2].

We’re going up a vertical wall of demand. — Sarah Friar

June: the paperwork. Both labs confidentially filed S-1s, the registration documents that open a public offering, with the stated purpose of funding infrastructure, OpenAI's describing a $600 billion buildout through 2030. The backdrop was a price war metered in tokens, the units of text that AI pricing is sold in; corporate customers rationing their AI budgets over cost, some defecting to cheaper open-source models; and OpenAI weighing sharp cuts of its own after Anthropic's valuation had passed it ^[13]. August: the price moved first, then the date. The Luna cut took effect on the first of the month; on the 19th, the listing slid to 2027. Friar supplied the description of the new date ^[4].

The IPO is not a finish line, it is a milestone, another fundraise. — Sarah Friar

The month carried a human line as well: revenue chief Denise Dresser, Brad Lightcap and product business chief Fidji Simo, the executive hired months earlier to run the April pivot, all went out the door, at a company whose revenue was running at a $40 billion yearly pace, still behind the $65 billion pace at Anthropic ^[4]. September: one week, three moves. On the 13th, Friar pitched recursive self-improvement, large models training the smaller ones, to a Goldman Sachs investor conference, framed explicitly as a way to cut training costs significantly ^[14]. On the 16th, OpenAI turned its chat products into advertising space: conversational ads and Sponsored Agents, a business already running at about a billion dollars a year ^[15]. The launch carried its own framing.

GPT-5.6 Luna is the closest we've come to intelligence too cheap to meter. — Michele Catasta

Also on the 16th, the $1.2 trillion proposals surfaced ^[5]. A cheaper way to train, a new line of revenue, a bigger number on the door, all in one week. Anthropic is the control case. Revenue has run from roughly $9 billion a year at the end of 2025 to more than $65 billion by late July ^[3]. The May round arrived tied to years of capacity, five gigawatts at Amazon, five more from Google and Broadcom, and GPU access through SpaceX's Colossus ^[1]. A $15 billion credit line is being finalized, up to $40 billion of Alphabet backing sits behind the float, and up to $5 billion of AMD equity rides on a two-gigawatt deployment that starts in early 2027 ^[3]. The August credit analysis projects Anthropic's contract book covered by revenue by 2027, while doubting OpenAI's can be ^[6]. Here is the lab with the growth, the capacity, the backer and, by the projection, a payable bill. It delayed its listing anyway, to the edge of the midterms, with a $2 trillion target riding on the revenue line ^[3]. A funding crunch explains one missed deadline. A shared calendar explains two. And then there is the fact that cuts against the whole reading, and it deserves its full weight: the money has not stopped. The two record rounds closed ^[2][1]. Global investors absorbed more than $600 billion of the builders' bonds ^[10]. Micron locked its customers into the floors that defend its margins through 2030 ^[9]. Access to capital is not what the labs lost this year; the market repriced the risk and kept the window open. What did not reprice is the schedule: the backlog becomes the bill in 2027 and 2028, whatever the mood ^[6]. The clearest summary of the six months, though, comes from one speaker. In September, with a $1.2 trillion valuation floating around her company, Sarah Friar insisted the finances were sound ^[5].

We still have an incredible balance sheet. — Sarah Friar

In May, with the record round closed, she had described what that same balance sheet required next ^[2].

We’re going up a vertical wall of demand. — Sarah Friar