OpenAI and Anthropic, direct competitors selling the same category of product to the same customers, both filed draft IPO paperwork with the SEC within weeks of each other in mid-2026, and both are now working toward a public listing in the same window. Two rivals racing each other to the same public markets on the same clock is not, on its own, remarkable. What makes it worth a second look is what each company's own disclosures say about why.
| OpenAI | Anthropic | |
|---|---|---|
| Draft S-1 filed | Confidentially, June 2026 | 1 June 2026 |
| Last major private round | $122 billion, closed 31 March 2026 | $65 billion Series H, closed 7 May 2026 |
| Resulting valuation | $852 billion | $965 billion (more than double its $380 billion valuation in February) |
| 2025 revenue (actual, not run rate) | $13.1 billion | Around $9 billion |
| Annualised run rate, most recent reported | Over $40 billion (August 2026), up from around $20 billion at end of 2025 | $65 billion (end of July 2026), up from $47 billion in May |
| Target IPO window | Fourth quarter 2026; reporting suggests this may slip into 2027 | As early as October 2026; some backers reportedly expect a float valuation near $2 trillion |
| Infrastructure commitment | Around $600 billion in projected total compute spend through 2030 | Comparable multi-year compute commitments, not disclosed at the same precision |
None of these figures are audited public financials; they are drawn from press reporting on private disclosures, sourced individually below. Run rate and actual revenue are different measures, run rate annualises current sales pace, actual revenue is what was booked, and the gap between OpenAI's $13.1 billion 2025 revenue and its $20 billion year-end run rate is a useful reminder to read every "revenue" figure in this space for which of the two it actually means.
Even the more flattering read, Anthropic's run rate now exceeding OpenAI's by a wide margin, doesn't change the underlying question. It says one horse is running faster than the other, not that either has crossed the line.
Both companies have been extraordinarily well funded privately, increasingly by the same small circle of counterparties, and two numbers from that circle are worth being precise about, because getting them wrong is exactly the kind of error that lets an argument like this one get waved away.
The first is $600 billion. Reuters reported in February 2026 that OpenAI was targeting roughly that figure in total compute spend through 2030, down from an earlier $1.4 trillion, 30-gigawatt figure Sam Altman had previously cited. Nvidia's Jensen Huang referenced the same $600 billion figure again in an August blog post tied to a separate announcement, framing it as roughly what OpenAI's chip spending could be worth to Nvidia over the same period. It is a spending and revenue projection, not a financing instrument, and it has not shrunk, collapsed, or been renegotiated down to anything else.
The second is $105 billion, and it is a different animal entirely. On 17 August, Nvidia disclosed in an SEC filing that it would guarantee up to $105 billion in residual value tied to OpenAI's 20-year lease with SB Energy for a data centre campus in Pike County, Ohio, an initial 4.25-gigawatt phase of a roughly 8-gigawatt site. This is not $105 billion of cash changing hands. It is a conditional backstop, triggered only if OpenAI defaults or becomes insolvent, and Nvidia separately put $1.5 billion of direct investment into SB Energy itself. SB Energy's own S-1 filing is unusually blunt about why the guarantee exists at all: OpenAI is not an investment-grade tenant. Put plainly, Nvidia is underwriting the credit risk of its own biggest customer's landlord, so that the landlord can raise the debt to build the building OpenAI wants to rent. Jensen Huang pushed back publicly on the obvious reading, "Is this circular financing? No. OpenAI will pay the lease," he said, which is a claim about intent, not a rebuttal of the structure.
Those are two separate, real numbers describing two separate, real things, and the accurate version is arguably more useful to the sceptic than the inaccurate one: this is not a headline that quietly shrank under scrutiny, it is a chipmaker acting as a credit backstop for its largest customer's real estate, disclosed in the same month that customer was racing toward a public listing.
The idea that an IPO is the moment these companies first reach past Silicon Valley to ordinary savers is not quite right, OpenAI already did that six months earlier. Buried inside the $122 billion round that closed on 31 March 2026 was a tranche opened for the first time to individual investors, routed through JPMorgan, Morgan Stanley and Goldman Sachs. OpenAI had budgeted for roughly $1 billion of retail demand. It got more than $3 billion, a threefold oversubscription that one of the participating banks reportedly could not process without its systems buckling under the load. OpenAI's CFO has since confirmed the company intends to reserve shares for retail investors in the IPO itself.
That is worth pausing on. The backdoor to ordinary capital did not wait for a prospectus, a roadshow, or a listing date. It opened the moment the private syndicate's appetite showed signs of a ceiling, which is the clearest documented instance yet of the pattern this piece is arguing for, not a prediction about what an IPO might do, but a trend already under way.
The venture and strategic investors who built the circular financing web went in with detailed knowledge of the structure, the counterparties, and the exposure. An IPO, and the retail tranche that preceded it, reaches a different audience in kind, not just in size. It is not that FTSE or S&P trackers are obligated to buy in on day one, S&P 500 inclusion requires a trailing profitability record neither company currently has, and neither is even eligible for a UK index built around UK-listed companies. The more accurate mechanism is quieter and just as automatic: global and total-market index funds, the default building block of most UK and Manx pension scheme equity allocations, add constituents by market capitalisation and free float, with no earnings screen at all. Once either company lists at the scale being discussed, that money follows mechanically, without an active fund manager ever making an individual judgement call about OpenAI's or Anthropic's burn rate.
None of this means the underlying technology or the companies building it are worthless; genuine enterprise demand for frontier models is real and growing, and infrastructure-heavy industries have used supplier financing to bootstrap real markets before. What it means is narrower: two companies racing each other to public markets in the same window, having already tested retail appetite privately six months ahead of any prospectus, immediately after their own largest backers reached the edge of what a private syndicate would underwrite, is a pattern worth naming for what it looks like, rather than accepting the framing each company will put on its own filing.
The circular financing web did not solve the economics of frontier compute; it merely delayed the reckoning until pension funds and retail brokerages could be positioned beneath the drop.
Is the AI South Sea Bubble about to burst?
Editor's note: Concrete companion piece to the queued "Can AI Afford Itself" macro-bubble investigation, and to "The Harness Does the Talking". Framing term drawn from "The Theatre Pulldown".
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
