In the April-June quarter of 2026, OpenAI's operating loss reached $12.3 billion. That's up 32% from $9.3 billion in the January-March quarter, while revenue growth over the same period was a mere 18%, rising from $5.7 billion to $6.7 billion. And yet the company's price tag hasn't moved.
The $852 billion valuation set in a $122 billion funding round in March 2026 was carried over unchanged into a $7 billion employee share sale completed on August 10. The same stock now carries two prices side by side. Even in the secondary market, the markup was a mere 3%—while Anthropic, trading in that same market, has climbed more than 10% above its last funding round. If losses have been what pushed the valuation higher, then this half-year, in which losses swelled to a record, is precisely when the price tag should have risen. It didn't.
From a $9.3 Billion Loss to $12.3 Billion, an $852 Billion Valuation Frozen for Four Months
The first-half 2026 breakdown reported by the Wall Street Journal worsens quarter over quarter. January-March: $5.7 billion in revenue against a $9.3 billion operating loss. April-June: $6.7 billion in revenue against a $12.3 billion operating loss. The cumulative first-half operating loss hit $21.6 billion. An internal full-year forecast reported by The Information puts the figure at $14 billion—a number adjusted outside U.S. GAAP. The GAAP-based first-half actual of $21.6 billion, which includes stock-based compensation and similar items, is 1.5 times the $14 billion forecast calculated under different standards.
A private company's valuation isn't priced daily like a public stock. It only updates the moment someone actually buys or sells shares and a per-share price is set. The $852 billion figure was fixed in the $122 billion funding round that closed in March 2026, and the same number was simply reused in August's employee share sale. This isn't a case of the market choosing to hold steady—it's a case of the previous number being carried forward, unchanged, at the next opportunity to reprice.
A common point of confusion here is the difference between a confirmed valuation and a target valuation. The figure repeatedly cited in Japanese-language reporting—roughly ¥160 trillion—is a conversion, at ¥159.5 to the dollar, of CEO Sam Altman's stated IPO target of "over $1 trillion." What's actually confirmed in a transaction today is $852 billion, or roughly ¥136 trillion at the same exchange rate. The gap between the two is 17%, and nothing in the company's actual performance has closed it. That ¥136 trillion figure alone exceeds Japan's roughly ¥122 trillion general-account budget for fiscal 2026.
Even the size of the full-year loss varies by more than fourfold depending on how you count it. The internal forecast of $14 billion is an adjusted metric the company defines itself, and outsiders have no way to verify what's been excluded. Annualize the first-half actual of $21.6 billion at the same pace, and the full year comes out closer to $40 billion. A recalculation on a GAAP basis by the forecasting site futuresearch.ai, built from Q1 and Q2 actuals, puts it at roughly $60 billion, with a range spanning $38 billion to $95 billion. Before anyone can debate whether the valuation is justified, the size of the loss itself isn't settled.
Lining up how the loss and revenue have each grown makes clear what changed over this half-year: revenue rose 18% quarter over quarter, while the loss rose 32%. OpenAI, as a private company, hasn't publicly explained why the gap between $21.6 billion and $14 billion is so large, even accounting for different accounting standards. Without that explanation, the price tag has simply sat there for six months.
Five Deals, 29-Fold: How $29 Billion Became $852 Billion
When Microsoft put in $10 billion in January 2023, OpenAI was valued at $29 billion. From there, a Thrive Capital-led employee tender pushed it to $80 billion; a $6.6 billion raise on October 2, 2024 brought it to $157 billion; a $40 billion raise on March 31, 2025 brought it to $300 billion (with SoftBank contributing $30 billion and a syndicate the remaining $10 billion in that round); an employee share sale in October 2025 brought it to $500 billion; and a $122 billion raise on March 31, 2026 brought it to $852 billion. That's roughly a 29.4-fold increase in 38 months.
Two of those five steps didn't put a single dollar into the company. The $80 billion tender that closed by February 2024 and the $500 billion valuation from October 2025 were both transactions in which existing shareholders or employees sold shares to third parties. No new shares were issued, so no capital was raised and no existing shareholder was diluted. Even so, the prices set in those deals became the starting point for the next round and remain a step on the ladder. The rise in valuation and the company's actual fundraising have looked like the same motion but have proceeded on separate tracks.
The spacing and size of the step-ups have also shifted. Going from $29 billion to $80 billion took 13 months and multiplied the valuation 2.8 times; the most recent step, from $500 billion to $852 billion, took just five months and multiplied it 1.70 times. The size of each jump is shrinking even as the intervals between transactions shorten. The more funders there are and the more frequent the deals, the smaller the increment that gets added each time.
The corporate restructuring completed on October 28, 2025 sits at a turning point partway through this ladder. The nonprofit was renamed OpenAI Foundation and retained control, while the for-profit subsidiary converted into OpenAI Group PBC, a public benefit corporation. Post-conversion ownership stakes were disclosed as 26% for OpenAI Foundation (roughly $130 billion) and 27% for Microsoft (roughly $135 billion). That $135 billion figure matches exactly 27% of $500 billion—meaning the $500 billion price set by an employee secondary transaction 26 days earlier, on October 2, became the direct reference point for calculating these stakes.
In that same conversion, Microsoft gave up part of its rights in exchange for locking in its ownership figure. Its technology usage rights to OpenAI's models and products were extended to 2032, while usage rights to research IP now run only until either AGI (artificial general intelligence) is verified as achieved or 2030, whichever comes first—and consumer hardware is excluded entirely. The declaration of AGI itself now rests with OpenAI, to be verified by an independent panel, and Microsoft's right of first refusal on compute procurement has been eliminated. In terms of capital structure, this is the point at which OpenAI moved closer to being an ordinary company that investors can simply price. It's also after this conversion that an IPO began to be discussed as a realistic exit option.
The Conditions Behind a Valuation 65 Times Actual Revenue
Divide $852 billion by 2025 full-year revenue of $13.07 billion, and you get roughly 65x. Industry compilations put AI company revenue multiples at a median of 20x to 30x, with even the top tier of infrastructure players capping out around 40x to 100x—putting OpenAI's 65x more than double the median. In 2025, OpenAI recorded a GAAP net loss of $38.5 billion and an operating loss of $20.92 billion. The net loss alone is nearly three times revenue. And this is the company carrying a price tag 65 times its revenue.
Greg Jensen, co-CIO of Bridgewater Associates, has reportedly said that a revenue multiple of 35x already prices in "a monopoly-like outcome that doesn't exist yet." If that's the read at 35x, then 65x is pricing in something further out still. What buyers are paying for isn't the value of the current business—it's their share of the payoff if the scenario in which this company ultimately monopolizes the market comes true.
The logic behind why losses don't count against the company lies in the order in which buyers examine them. Buyers first look at how fast revenue is growing, then at how large the addressable market might become. As long as the losses can be explained by advance purchases of compute and training costs, they get reinterpreted as prepayment for future production capacity, and the bottom line of the income statement stops factoring into the discount calculation. This reinterpretation carries one condition: the growth in losses must stay below the growth in revenue.
Divide the same company by a different denominator, and the picture changes. OpenAI's annualized revenue run rate, which stood at $20 billion at the end of 2025, doubled to $40 billion by July-August 2026. Divide $852 billion by $40 billion and you get 21x—squarely inside the industry's median band. Whether you land on 65x or 21x flips the conclusion entirely, and the line between bullish and bearish comes down to whether you're looking at trailing actual revenue or the most recent annualized run rate.
This particular yardstick carries a side effect from management's perspective: the only way to bring the multiple down is to grow the denominator—revenue. Cutting costs doesn't move the multiple by a single point. Since compressing losses does nothing to improve the story behind the valuation, management has little incentive to rein in spending. Part of what looks like losses pushing the valuation upward is really an artifact of which yardstick gets chosen.
How much these two multiples converge depends on 2026's actual results. First-half revenue was $5.7 billion and $6.7 billion, for a combined $12.4 billion; the $40 billion run rate implies roughly $10 billion per quarter. If the second half hits that pace, full-year revenue lands around $32 billion, bringing $852 billion down to roughly 26x—right at the upper edge of the industry median. Whether the 65x figure ever becomes a story about actual performance rather than a story about the future hinges on whether revenue can close the gap between the $12.4 billion first-half actual and the $40 billion annualized run rate.
Amazon is the precedent most often invoked here. It grew revenue roughly 9-fold from 1996 to 1997 while continuing to post losses, and didn't turn a full-year profit until 2003 (announced in January 2004). Uber, cited under the same logic, lost investor patience for its losses once its growth rate slowed. What separated the two wasn't the size of the losses. It was whether the growth rate running alongside those losses remained persuasive enough to keep investors reinterpreting them.
The first half of 2026 was the half-year in which that pairing changed direction for OpenAI. While revenue grew 18%, losses grew 32%, and the grounds for reworking 65x down to 21x have gotten thinner each quarter. The price tag isn't frozen because buyers have stayed bullish or turned bearish—it's frozen because the evidence needed to move it in either direction hasn't fully arrived yet.
Who's Actually Backing This Bet
In November 2025, Sam Altman said OpenAI had signed compute agreements worth roughly $1.4 trillion over eight years, covering 30 gigawatts, with partners including Nvidia, AMD, and Broadcom. No primary source document has been made public; the numbers come from Altman's own statements. In February 2026, CNBC exclusively reported that the spending target presented to investors had been revised to roughly $600 billion through 2030. The $1.4 trillion figure covers total contracts over eight years, while the $600 billion figure is an investment target through 2030—different time horizons, so the two aren't directly comparable. Both are forward-looking plans, not money already spent.
The party actually generating the funds to cover these payments is less OpenAI itself than its counterparties. OpenAI has not issued corporate bonds. Meanwhile, global AI-related bond issuance is projected to reach roughly $570 billion in 2026, running at four times the prior year's pace. Oracle, which has been fronting fundraising on OpenAI's behalf, had its long-term rating cut by S&P from BBB to BBB- in July 2026, with its OpenAI-related contracts specifically named as a "key credit risk." Debt that never appears on OpenAI's own balance sheet is eating into its counterparties' credit ratings.
OpenAI hasn't publicly explained why it doesn't issue its own bonds. What's clear is the outcome: it's Oracle's rating that took the hit, and as long as OpenAI itself issues no bonds, its own creditworthiness never gets tested by the market. A borrower whose repayment capacity is never scored is the beneficiary of trillion-dollar-scale capital spending.
The lending platform Nvidia announced on August 10, 2026 makes this structure even clearer. Built with six partners including Apollo, Blackstone, and KKR, the framework mobilizes more than $500 billion in third-party capital. A residual value guarantee underpins part of these deals, though the announcement didn't specify how that $500 billion breaks down by company.
A lender taking GPUs as collateral can't easily predict what price those GPUs will fetch if the borrower defaults. Nvidia has said it may provide residual-value support of up to 25% of the deal total for some transactions—and if that support is exercised, the unpredictable part shrinks, making the loans easier to originate. This can become a loop in which Nvidia partially backs the resale value of its own product in order to help generate the financing needed to buy that same product. The framework remains at the memorandum-of-understanding (MOU) stage, pending final agreements, and which deals and terms it will ultimately cover are still to be determined.
Nvidia is also propping things up from the equity side. It put in $30 billion in the $122 billion round in March 2026, making it a major backer. In that same round, Amazon put in $50 billion—more than Nvidia's contribution. Separately from the six-partner platform announced on August 10, SEC filings confirm that on August 17, 2026, Nvidia entered into a guarantee agreement covering up to $105 billion for OpenAI's lease and power payment obligations tied to SB Energy's Ohio AI facility. A guarantee originally reported at roughly $250 billion shrank to this confirmed figure once the specific contract was finalized.
Tracing the money reveals the same company wearing several hats at once. Nvidia puts in capital as an investor, absorbs the lender's risk as a guarantor, and, as the seller of the compute itself, ultimately receives the money spent on the GPUs it helped finance. Nikkei reported in February 2026, drawing a comparison to the circular investments of the dot-com era, on the money flowing back and forth between Nvidia and OpenAI. In a structure where the seller helps create the buyer's purchasing power, whether the underlying demand is real can't be confirmed until the product is actually put to use.
David Sacks, the White House's science and technology adviser, has warned that the biggest risk in this lending scheme is a glut of compute supply. His concern is that the fiber optic cable laid en masse and left unused during the dot-com era could return in the form of "dark GPUs." The reason OpenAI's continued losses haven't threatened the company itself is that a substantial share of the stakes has been placed on Microsoft, Nvidia, and Oracle instead. The flip side is that any correction to the valuation is likelier to show up first in the credit ratings and bond prices of its partners, not in OpenAI itself.
Profitable Anthropic and the "Reckless" xAI
Anthropic's valuation climbed to $183 billion in its Series F in September 2025, to $380 billion in its Series G in February 2026, and to $965 billion in its May 28 Series H (which raised $65 billion). That's a 5.3-fold increase in a little over eight months, and it now surpasses OpenAI's $852 billion. Annualized revenue rose too, from $9 billion at the end of 2025 to $47 billion in May 2026 and then to $65 billion by the end of July.
The decisive difference lies in profitability. In the April-June 2026 quarter, Anthropic reportedly posted revenue of more than $11.5 billion (up 143% from the prior quarter) and turned profitable on an adjusted operating basis—its first-ever profitable quarter. In that same quarter, OpenAI posted a $12.3 billion operating loss on $6.7 billion in revenue. It was a quarter in which Anthropic came out ahead on both revenue and profitability.
If losses themselves were what drove valuations higher, a newly profitable Anthropic should have seen its price tag stall. The opposite happened. Dividing valuation by annualized revenue: OpenAI comes out to roughly 21x ($852 billion ÷ $40 billion), and Anthropic to roughly 15x ($965 billion ÷ $65 billion), both as of July-August 2026. OpenAI still commands roughly a 1.4x premium per dollar of revenue, even though Anthropic leads on both current revenue and profitability.
What investors are pricing is growth rate and expected future market position. Anthropic's profitability let investors reprice it as a company that can sustain its growth without depending on outside capital. OpenAI's 21x still bakes in the fundraising risk that comes with an assumption of continued losses.
xAI represents an extreme version of the same logic taken to its limit. It raised $20 billion in a Series E on January 6, 2026 at a $230 billion valuation, then merged with SpaceX 27 days later, on February 2, primarily through a stock swap. The combined post-merger valuation reached $1.25 trillion, and the $250 billion deal value assigned to the acquired xAI side was called the largest M&A transaction in history. Up until the merger, xAI's finances were invisible from the outside.
The numbers finally surfaced in SpaceX's S-1. The disclosed figures for January-March 2026 showed revenue of $818 million against an operating loss of $2.47 billion—losing roughly three times what it took in. Capital expenditure in that same quarter reached $7.7 billion. Harrison Rolfes, a senior research analyst at PitchBook, called the finances "reckless-looking compared to a traditional SaaS company."
Looking purely at multiples, the top two companies aren't even the most extreme. Annualizing xAI's January-March 2026 revenue of $818 million gives roughly $3.27 billion, putting its Series E valuation of $230 billion at roughly 70x that figure. Perplexity, with roughly $450 million in ARR as of March 2026, was reportedly valued at $23 billion around the same time, or roughly 51x. These are rough estimates from mismatched time periods, but the highest revenue multiples appear to belong to the smaller players, not the largest ones.
While these companies remain private, these numbers only surface through disclosure accidents. xAI's actual condition came to light as a byproduct of SpaceX's IPO preparations; OpenAI's 2025 results leaked out as internal documents. Revenue of $13.07 billion, an operating loss of $20.92 billion, a net loss of $38.5 billion—these are not figures the company chose to share with anyone beyond its investors. The only material available to check the pricing sits in the hands of the people doing the pricing.
SoftBank's ¥15.8 Trillion Stake
SoftBank Group is reported to hold 11.66% of OpenAI as of April 2026. At the $852 billion valuation, that comes to roughly $99.3 billion—or about ¥15.8 trillion at ¥159.5 to the dollar (as of August 28, 2026). The company put in $30 billion in the $40 billion round in March 2025 and committed another $30 billion to the March 2026 round; as of August 29, it had paid in $20 billion (in April and July), with the remaining $10 billion due in October. A 10% drop in valuation would erase roughly ¥1.6 trillion of that holding's value in yen terms.
This stake carries a specific accounting consequence for a publicly listed shareholder. SoftBank holds its OpenAI shares as fair-value-through-profit-or-loss (FVTPL) assets, meaning it must reassess their value every quarter and book the change as profit or loss, even absent any new transaction. So while OpenAI's own valuation stays fixed until the next deal, the value SoftBank reports each quarter can move on its own, driven by market conditions or internal model revisions. Through the earnings of this Tokyo Stock Exchange-listed company, Japanese pension funds and individual investors are exposed to that fluctuating valuation regardless of whether any actual transaction has occurred.
Japanese operating companies, meanwhile, have lined up on the other side of the ledger. SBI Holdings became the first Japanese financial group to sign a strategic partnership with Anthropic, on June 2, 2026. NEC announced on June 11 a tie-up connecting Anthropic with eight financial companies, including Sumitomo Mitsui Financial Group and Daiwa Securities. Hitachi and Fujitsu also announced their own partnerships in that same April-to-June window—six major Japanese companies gravitating toward Anthropic in that stretch.
That wave of partnerships, concentrated in April-June, straddles May 20, the date Bloomberg reported Anthropic was heading toward its first profitable operating quarter. What Japan's major companies chose was not an investment in a company still climbing a valuation ladder, but an operational partnership with a company moving toward profitability. Returns from an equity stake don't materialize until the valuation is revised, while the payoff from a partnership can be captured within a company's own operations.
A split has emerged: capital flowing to OpenAI, operations flowing to Anthropic. SoftBank's stake can swing by trillions of yen on the strength of a single new funding round, while the companies partnered with Anthropic collect returns in the form of products and operational results—untethered from any valuation swing. Japan's AI money is moving in two separate directions: one side betting on the price tag, the other side using the product to earn its return.
The Secondary Market Started Correcting the Price First
Against the $852 billion set in the funding round, OpenAI shares on the Forge Global benchmark traded at roughly $880 billion as of August 2026—just 3% above the round price. Over the same period, Anthropic shares traded in private markets at around $1.1 trillion, roughly 14% above its $965 billion Series H valuation set in May. The dynamic in which OpenAI led comfortably from late 2025 through early 2026 flipped over the course of this summer, with Anthropic overtaking it.
The reason for this gap lies in how the two markets set prices differently. A new funding round's valuation is negotiated between the company and a lead investor, and depending on the terms attached to preferred shares, the resulting number can come out higher than the underlying reality. Secondary trades, by contrast, are one-off transactions between existing shareholders and buyers, typically involving shares closer to common stock that don't benefit from those same preferential terms. That's why the size of the markup in secondary trading reflects market participants' view of the next round's valuation before that round happens. A markup as thin as 3% signals a prevailing view that the next number won't rise much past $852 billion.
Who's selling also helps read this gap. August's $7 billion allotment was a company-arranged employee sale that priced at the same $852 billion as the funding round. That within weeks of that sale, the secondary market only added 3% on top, shows the market still isn't confident the next round will land meaningfully above $852 billion.
Even the expected timing of an exit isn't aligned internally. OpenAI confidentially filed for an IPO with the U.S. Securities and Exchange Commission (SEC) in June 2026, targeting a valuation above $1 trillion. Altman has publicly stood behind that figure, while CFO Sarah Friar has reportedly told staff internally that 2027 is a more realistic timeline for going public. The person citing the target price and the person estimating the timeline are working from assumptions a year apart.
An IPO is both a fundraising mechanism and a mechanism for verifying a valuation. Because the filing was confidential, the submitted documents aren't public yet, but if the IPO does happen, numbers that until now could only be inferred from leaked documents and counterparties' credit ratings will appear in audited form. WeWork, which SoftBank valued at $47 billion in 2019, saw its valuation collapse the moment its S-1 that same year disclosed a $1.9 billion loss, and the company filed for bankruptcy in 2023. Disclosure has a way of settling, all at once, a price tag that had been building up while the company stayed private.
What built the 29-fold rise over three years and two months wasn't the losses themselves. As long as revenue growth outpaced loss growth, investors kept reinterpreting those losses as prepayment for future production capacity. The persuasiveness of the growth rate underpinning that reinterpretation reversed for the first time in the first half of 2026. The fact that the price tag stayed frozen for 132 days, staying nearly flat even in the secondary market, while Anthropic seized the valuation lead, reflects the market pricing in the collapse of that reinterpretation's premise.
Three numbers remain to be confirmed. Whether the next funding round prices above $852 billion. Whether an IPO, arriving in 2027 as Friar suggests, holds a valuation near the $1 trillion mark. And whether the full-year operating loss extends the trajectory of the first half's $21.6 billion, or whether a quarter emerges in which revenue growth once again outpaces loss growth. If that third number reverses first, the 65x multiple will stop being a bet on the future and start being a statement about actual performance—and OpenAI will finally emerge from a state of being priced without adequate disclosure.