The Trillion-Dollar
Loss Leader

On June 15, 2026, the independent journalist Ed Zitron published what he described as OpenAI's audited financial statements for 2024 and 2025. The Financial Times verified them independently the same week, and three further outlets, Fortune, Ars Technica, and Benzinga, confirmed the core figures within hours. For a company that has never published audited accounts, the disclosure is the single most consequential piece of financial information to surface before its prospectus, and it arrived eight days after OpenAI confidentially filed a draft registration statement with the SEC — and one week after its closest rival, Anthropic, filed its own.

The numbers describe a business expanding at a pace with no real precedent and losing money at a pace to match. Revenue rose from $3.7 billion in 2024 to $13.07 billion in 2025. Over the same year, total costs and expenses reached $34 billion, the loss from operations widened to $20.92 billion, and the net loss attributable to the company came to $38.53 billion, roughly 7.5 times the prior year. The operating figure, the one that describes the actual business rather than the accounting around it, works out to about $1.60 of spending for every dollar of revenue.

This report makes four arguments.

• The headline loss is mostly accounting; the operating loss is what describes the business. Of the $38.53 billion, roughly $41.55 billion is a one-time, non-cash charge from the company's conversion to a public benefit corporation. True cash burn was closer to $8 billion. But stripping the accounting away does not rescue the picture: the $20.92 billion operating loss is the cleanest read on the underlying economics, and it grew faster than revenue would have had the business been profitable.
• The cost base scales faster than the top line. Research and development alone, $19.18 billion, exceeded total revenue. Sales and marketing rose nearly fivefold to $5.73 billion, the wrong direction for a company whose bull case rests on demand pulling itself in.
• The dependency on a single vendor is structural. Of the $34 billion in 2025 costs, $17.2 billion flowed to Microsoft. The most-watched private company routes roughly half of its cost base through one supplier, which recaptures much of the spend as its own revenue.
• The valuation asks the market to underwrite all of this at roughly 76 times trailing revenue. A $1 trillion listing would rank among the fifteen largest companies in the S&P 500 on day one, on a business that has never turned an operating profit and does not project positive cash flow before the end of the decade.
• And the benchmark has moved. Anthropic, OpenAI's nearest rival, is now valued higher (~$965 billion), growing faster (it projected $10.9 billion of revenue in a single quarter against OpenAI's $13.07 billion for all of 2025), priced more cheaply against revenue, and leads in enterprise. It even projected a quarterly operating profit, though on a contested, non-GAAP basis tied to a temporary compute discount. It filed to go public a week before OpenAI. Section 12 treats this comparison in full, because it is the one a roadshow will force.

None of this is a verdict on the technology, and none of it forecloses the bull case. Inference cost per token has fallen by roughly an order of magnitude a year since 2021; if that continues, the unit economics of serving the product eventually invert. The Microsoft revenue-share cap removes a real drag on future cash flows. Revenue growth of this magnitude has historically forgiven a great deal. What the audited documents make difficult to argue is that the path to profitability is short, clear, or cheap, and a listing at $1 trillion prices in the assumption that it is all three.

Before turning to the audited statements, it is worth fixing the moment they landed in. On the morning of July 8, 2026, ten days before this report is dated, a single piece of reporting reset the price of the entire artificial-intelligence complex, and it did so without a single number from OpenAI's accounts changing.

The New York Times, citing three people involved in the deliberations, reported that OpenAI was weighing a delay of its initial public offering into 2027 in order to hold out for a valuation near $1 trillion. The company's advisers had reportedly laid out two paths: accept a lower valuation and list before the end of 2026, or wait, and aim higher. Sam Altman reportedly regarded any reduction of the trillion-dollar target as a non-starter, while the chief financial officer, Sarah Friar, had signalled to associates that a 2027 listing was the more probable outcome. The last private valuation, set in the March 2026 primary round, was roughly $852 billion.

The market did not treat the news as a scheduling matter. It treated it as a signal about whether investors still had the appetite to pay premium multiples for assets whose value rests on profits expected years out. SoftBank, which has built much of its recent thesis on its OpenAI stake, fell as much as 13% in Tokyo. The damage spread through the memory and semiconductor names that supply the AI buildout. SK Hynix and Samsung Electronics each fell 8% to 9%, Advantest dropped roughly 10%, and South Korea's KOSPI triggered its market-stabilising sidecar. The Nasdaq logged a fourth consecutive down session and was on course for its worst month since March 2025, even as the Dow set record highs, a clean rotation out of long-duration growth.

Two unrelated pressures compounded the move in the same week. On June 25, Apple raised prices on MacBooks and iPads by as much as 20%, citing the surging cost of memory chips, and its stock fell 6.1%; Microsoft raised Xbox prices by $100 to $150. The same memory shortage that lifts the cost of consumer hardware lifts the cost of AI inference. And the May reading of the Federal Reserve's preferred inflation gauge came in hot, pushing the discount-rate assumptions underneath every richly valued growth stock in the wrong direction.

OpenAI has never published audited financial statements. Until June 2026, every figure in public circulation was a leak, an estimate, or a projection: annualised run-rates extrapolated from a single strong month, board decks described secondhand, third-party reconstructions of a cap table. The disclosure changes the evidentiary basis of the entire discussion.

On June 15, 2026, the journalist Ed Zitron published what he described as OpenAI's audited statements for fiscal years 2024 and 2025, stating that the documents had been independently verified by the Financial Times. Within hours, Fortune, Ars Technica, and Benzinga had confirmed the central figures, and OpenAI declined to comment. The provenance matters because it separates this disclosure from an earlier, less reliable one: a reconstructed cap table that circulated in April was an estimate assembled from public and secondary data. The June statements are a different class of evidence, audited and verified by a major financial newspaper.

The top-line picture is a business that more than tripled its revenue and lost a fortune doing it. The composition of the cost base, charted below, shows why.

In 2024, OpenAI reported $3.7 billion of revenue against $12.48 billion of total costs and expenses, for a loss from operations of $8.78 billion and a net loss attributable to the company of $5.09 billion. In 2025, revenue rose to $13.07 billion while total costs and expenses reached $34 billion. The loss from operations widened to $20.92 billion, and the net loss attributable to the company came to $38.53 billion.

2025 net loss includes a one-time non-cash charge discussed in Section 04. Percentages computed by YRS.

Two features deserve immediate attention. First, research and development in 2025, at $19.18 billion, exceeded the company's entire revenue, a single expense line larger than every dollar the business took in. Second, sales and marketing rose more than fourfold. For a company whose narrative rests on demand pulling itself in faster than the cost to serve it, a customer-acquisition line growing at that rate is a detail worth sitting with. At year-end 2025 the company held just over $50 billion in assets, with roughly half in cash.

A number as large as $38.53 billion invites misuse in both directions. Read carelessly, it becomes "OpenAI burned $38 billion in cash," which is false. Read defensively, it becomes "most of it is just accounting, so ignore it," which is also wrong. The honest reading sits between, and getting it right is the difference between analysis and headline.

The headline net loss is reconciled from a much larger gross figure through a sequence of adjustments, and the single largest item is non-cash. When OpenAI converted from a nonprofit to a public benefit corporation in October 2025, early-investor rights and warrants converted into equity. As the value of those instruments rose, driven by the rising value of the company itself, the conversion generated a fair-value charge of roughly $41.55 billion. A person familiar with the matter told the Financial Times that the increase in OpenAI's worth alone produced a charge of about $30 billion. No money left the building. It is an accounting consequence of becoming more valuable, not a cash cost of operating.

Strip the conversion charge away and the relevant figures are these. The operating loss of $20.92 billion is the cleanest description of the business: what it costs to run, set against what it earns. On that basis the company spent roughly $1.60 for every dollar of revenue in 2025. Below that, accounting for interest income, interest expense, and the losses allocated to noncontrolling and redeemable interests, the company's true cash burn for the year was estimated near $8 billion. That is the figure that actually depletes the balance sheet.

So which number belongs in the analysis? All three, used for what each describes. The $38.53 billion is the audited headline and will anchor the prospectus; ignoring it is not an option. The roughly $8 billion of cash burn is what matters for runway. The $20.92 billion operating loss should anchor any judgment about the underlying business, since it sits apart from the conversion accounting and the one-time items.

The reconciliation also contains a smaller item worth noting. In both years, OpenAI allocated a portion of its losses to "noncontrolling members' capital" and, in 2025, additionally to "redeemable noncontrolling interests," $17.87 billion and $3.95 billion respectively. These allocations reduce the group net loss of roughly $60.35 billion to the $38.53 billion attributable to the company. The mechanics are ordinary for a complex capital structure, but the scale is a reminder that OpenAI's ownership is unusually layered.

A loss is a relationship between two numbers, and the revenue side of OpenAI's is healthy by any standard. The pressure is entirely on the other side. Understanding why the company loses money means understanding why its cost base grows faster than the revenue it supports.

The clearest way to see the problem is to place every category of spending against the revenue it has to be funded from. In 2025, total costs ran to $34 billion against $13.07 billion of revenue: everything beyond the first $13 billion is spending the business did not earn back.

Research and development dominates at $19.18 billion, more than double the prior year and, as noted, larger than total revenue. For a frontier-model developer this is the core of the business: training successive generations of models consumes enormous compute, and "research and development" here is in large part the cost of that training. The audited documents indicate a substantial share of this line was paid to Microsoft for compute, developed in the next section.

Cost of revenue, at $7.5 billion, is the expense of serving the product, the GPU compute consumed every time a user submits a query. With roughly 800 million weekly users on the consumer product alone, much of it free, this line scales directly with usage; in the first half of 2025 the company spent $5.02 billion on Azure inference. Sales and marketing is the line that should give a bull pause: at $5.73 billion it rose nearly fivefold. The bull case for any platform is that the product becomes its own distribution; a sales-and-marketing line growing faster than revenue points the other way.

This is the structural heart of the problem, and it cuts against the intuition that software margins improve with scale. The next model is larger; training it costs more; serving it to more users at higher quality costs more still. The clearest single test of whether the business is bending toward profitability is to compare the growth rates of revenue and cost directly.

Total costs grew 2.72× against revenue's 3.53×, which means that in 2025, for the first time clearly visible in the data, revenue outpaced costs on a growth basis. If that relationship holds and widens, the operating loss eventually narrows in absolute terms. The bull case lives in that gap. The bear case observes that the gap is narrow, the base is enormous, the operating loss still grew 2.38× in absolute terms, and sales and marketing grew faster than any other line, 5.16×. A single year does not establish a trend.

The audited statements expose a relationship that estimates could only gesture at: the degree to which OpenAI's cost base flows through a single counterparty. Of the $34 billion OpenAI spent in 2025, $17.2 billion went to Microsoft. Roughly half of the most-watched private company's spending is routed to one supplier, which books much of it as revenue.

The composition tells the story. OpenAI paid Microsoft $10.59 billion classified as research-and-development expense, most plausibly the cost of training its models on Azure. A further $6.047 billion fell under cost of revenue, the inference compute that serves the product. Smaller amounts covered sales and marketing ($527 million) and general and administrative expense ($42 million). The total, $17.2 billion, means that for every dollar OpenAI spent in 2025, roughly fifty cents reached Microsoft.

The flow in the other direction is a rounding error by comparison: Microsoft paid OpenAI $303 million in 2025, and SoftBank paid it $867 million. At year-end, OpenAI carried liabilities to Microsoft of $3.64 billion. The dependency is not a partnership of equals in cash terms; it is a customer relationship in which OpenAI is overwhelmingly the payer.

This concentration carries two consequences. The first is recapture: a large fraction of every dollar an investor puts into OpenAI to fund its compute is recaptured by Microsoft as Azure revenue. The second is fragility. A cost base that runs through one counterparty is exposed to that counterparty's pricing, capacity, and strategic priorities. The October 2025 recapitalisation reworked the commercial terms, capping OpenAI's revenue-share payments to Microsoft at $38 billion through 2030, removing Azure exclusivity, and extending Microsoft's IP rights to OpenAI's models through 2032, but the operational dependency on Azure compute, as the 2025 numbers show, remains the spine of the cost structure.

If the cost base explains why OpenAI loses money today, the trajectory of unit costs is where the argument about tomorrow is fought. The bull case rests on a single empirical claim: that the cost of serving a model falls fast enough, for long enough, that the unit economics eventually invert. The evidence for that claim is real. So is the catch.

The price of inference has collapsed. A model in the GPT-4 class cost roughly $30 per million input tokens at launch in early 2023; by 2026 the comparable figure was a fraction of that. Industry-wide, the average cost per million tokens has fallen by roughly an order of magnitude per year since 2021.

Here is the catch, and it is the most important second-order point in the unit-economics debate. Per-token prices are falling, but per-task token consumption is rising even faster. The newer, more capable models, particularly the agentic systems that reason in long chains and call tools repeatedly, burn dramatically more tokens to complete a given task. Goldman Sachs has projected that agentic AI could lift token consumption roughly twenty-four-fold by 2030, to on the order of 120 quadrillion tokens per month. When the number of tokens per task grows faster than the price per token falls, the bill goes up even as the unit price goes down.

This is not theoretical. Enterprises are seeing AI bills rise, not fall, despite per-token deflation. One large technology company reportedly exhausted its entire annual budget for AI coding tools within four months. A senior figure at the dominant chip supplier conceded that for his team, the cost of compute now exceeds the cost of the employees it was meant to augment. For OpenAI specifically, the implication cuts both ways: falling per-token costs help its own margins on every query served, but the same deflation invites a price war. Reporting indicates OpenAI has weighed steep cuts to its own token pricing to defend against Anthropic and cheaper open-weight models. A sustained price war would compress the very revenue line the bull case needs to grow.

Every section to this point has examined whether OpenAI can afford to supply artificial intelligence. This one examines the assumption underneath the entire valuation: that enterprises will demand it, at a price, in volumes, and with a return that justifies the spending. The demand side is where the bull case is most exposed, and the most damaging evidence comes from the most committed adopters.

Consider the single most telling episode of 2026. Microsoft, the company that put roughly $13.75 billion into OpenAI, that writes a meaningful share of its own code with generative AI, and that has staked its product strategy on the technology, cancelled most of its internal licences for a leading AI coding tool roughly six months after rolling them out, and redirected thousands of engineers to a cheaper in-house alternative. The reported reason was cost: internal use at the scale employees had adopted it became hard to justify against the bill.

The episode crystallises an uncomfortable question: what if using AI at scale costs more than the labour it is meant to streamline? A study of the economics of automating vision-related work found AI cheap enough to make economic sense for only about a quarter of the wages tied to those tasks; for the remaining three-quarters, keeping a human was still cheaper. Independent productivity research has struggled to find measurable, economy-wide gains from AI adoption at scale.

The willingness to pay is thinner than usage suggests. In a large 2026 survey, the share of US adults expecting AI to have a negative effect on society over the next two decades exceeded the share expecting a positive effect by more than two to one. Adoption is rising, but much of it is involuntary (an AI summary atop a search result, an automated voice on a support line) rather than a paid choice. Even at Microsoft, only a low-single-digit percentage of the users who interact with its AI assistant in its productivity suite actually pay for it.

None of this proves that AI demand will disappoint. Enterprises are spending real money, the technology is genuinely useful, and the providers are iterating quickly. But the demand side is an assumption, not a fact, and the audited supply-side losses can only be justified by demand that materialises at scale and at a profit. The evidence that it will is, at the moment, mixed, and the most committed adopters are the ones supplying the doubts.

Every valuation is really a bundle of assumptions expressed as a single figure. The case for $1 trillion does not rest on OpenAI being worth that today; on its current financials it is not. It rests on the claim that the company will be worth far more later. An underwriter has to decide how aggressive those assumptions really are, and whether they cross from aggressive into implausible.

The valuation has roughly doubled every six to twelve months for two years, from $157 billion in late 2024 to $852 billion in early 2026, with the trillion-dollar target one step beyond.

Start with the multiple. On $13.07 billion of audited 2025 revenue, a $1 trillion valuation is roughly 76 times trailing revenue. On the annualised run-rate the company prefers to cite, on the order of $20 billion, it is closer to 50 times. Either way it is a multiple with almost no precedent among companies of this size, applied to a business with a $20.92 billion operating loss. The chart below sets those multiples against the large-cap technology names a public investor would compare them to, and against the most important live comparison of all, Anthropic, which is examined in detail in Section 12.

Sources: company announcements and reporting.

That ascent has been funded by private capital willing to underwrite the growth story at escalating prices. A public listing changes the buyer. Public markets price on disclosed financials, demand quarterly accountability, and reprice violently when assumptions slip, exactly the dynamic the June 26 selloff demonstrated. There is a quieter signal in the private market itself: reporting on secondary transactions indicates that since the March round, pricing has skewed below the $852 billion mark, with sellers outnumbering buyers by a wide margin in some venues. When the people closest to the company are net sellers below the last primary mark, it is at minimum a data point against the idea that $1 trillion is a floor.

A company that burns cash and loses money on operations survives on its ability to raise capital. OpenAI has raised more, faster, than almost any private company in history, but the machinery that funds it has become intricate, leveraged, and in places circular, and the audited losses raise the stakes on every part of it.

The scale of the forward need is the place to start. With cash on hand of roughly $25 billion at year-end 2025 and burn that several analysts expect to accelerate sharply, the runway is not long relative to the spend. One major bank has estimated that OpenAI may need more than $207 billion in additional capital through 2030, even under optimistic assumptions.

That capital has become concentrated. SoftBank is the central node. Its total commitment reaches roughly $65 billion by October 2025, for a stake of around 13%. To fund the bet, SoftBank sold its entire stake in the dominant AI chipmaker and the remainder of a major telecom holding. But the structure underneath is leveraged: SoftBank signed a $40 billion bridge loan in March 2026, maturing in March 2027, whose twelve-month tenor implies that lenders expect an OpenAI liquidity event within the year. A delay of that IPO into 2027 is therefore not a neutral event for SoftBank; it is a direct complication of the refinancing the bridge loan assumes.

The strain is already visible. A planned margin loan of around $6 billion, itself cut from a larger target, stalled in mid-2026 after lenders balked at pricing a private-company stake as collateral. A credit rating agency cut SoftBank's outlook to negative, citing the liquidity strain from its OpenAI exposure. SoftBank's own equity now trades as a leveraged, illiquid proxy for OpenAI sentiment, which is why a report about OpenAI's IPO timing moved SoftBank's stock 13% in a session.

Around this sits a broader pattern that the skeptics call circular financing. The dominant chip supplier has committed to invest up to $100 billion into OpenAI progressively as OpenAI deploys that supplier's systems, capital that flows out to buy the chips it came in to fund. Critics liken the arrangement to the vendor financing of the late-1990s internet bubble. The audited losses sharpen the concern: the more OpenAI loses, the more external capital the circle must keep supplying, and the more the whole structure depends on the IPO window staying open.

A valuation premium of OpenAI's magnitude implicitly assumes durable competitive advantage, that the company will hold its lead and convert it to profit before rivals erode the prices that profit depends on. The competitive evidence of 2026 complicates that assumption, most pointedly in the enterprise market, where the money is.

The central development is that Anthropic, OpenAI's nearest competitor, has flipped the enterprise narrative. By one widely-cited estimate of enterprise large-language-model spending, Anthropic's share rose to roughly 40% while OpenAI's fell to about 27%, a reversal from the position three years earlier.

Anthropic also filed confidentially for its own IPO, one week before OpenAI, and reached a private valuation that, by mid-2026, had eclipsed OpenAI's in at least one round. Its reported revenue climbed steeply through the year, and it even projected a first quarterly operating profit, though on a contested basis examined in Section 12. Banks reportedly told both companies that whoever lists first will define the category.

The consumer position is more secure but not static. ChatGPT remains the dominant consumer AI product by a wide margin, and consumer subscriptions are the core of OpenAI's revenue. But its share of the consumer AI market has slipped from its peak, competitively-priced offerings from a major search incumbent are gaining, and open-weight models continue to commoditise the underlying capability. A premium valuation assumes pricing power; commoditisation is the enemy of pricing power, and the model layer is commoditising.

Figures are estimates and reported targets, not audited for Anthropic.

Every valuation needs a comparison. OpenAI's cleanest one comes from its nearest private rival rather than from any public technology company. Anthropic is racing the same course, on the same timetable, toward the same public market, and on several of the metrics that matter most to an underwriter, it is ahead. Because the two companies will likely price within months of each other, Anthropic is the benchmark against which OpenAI's $1 trillion will actually be judged.

The headline reversal happened in the spring of 2026. Anthropic closed a financing round that valued it at roughly $965 billion, eclipsing OpenAI's $852 billion mark for the first time, and one week before OpenAI's own confidential submission it filed a draft registration statement of its own with the SEC. The company that was, two years earlier, widely regarded as the underdog is now the higher-valued of the two and, by the consensus of bankers and prediction markets alike, the more likely to list first.

The growth comparison is the part that should most concern an OpenAI underwriter. Anthropic has reported an annualised revenue run-rate climbing from roughly $9 billion at the end of 2025 to about $47 billion by May 2026 — a fivefold increase in under six months, and an order-of-magnitude pace of compounding that outruns OpenAI's. Where OpenAI roughly tripled revenue year over year, Anthropic has reportedly been compounding at something closer to ten times annually since each company first reached a billion dollars of annualised revenue. If those trajectories hold even approximately, Anthropic passes OpenAI in absolute revenue, having started far behind.

The divergence is not only about growth rate; it is about the shape of the business. Anthropic's revenue is roughly four-fifths enterprise, sold to businesses on subscriptions and API access, and in the spring of 2026 its share of enterprise large-language-model spending surpassed OpenAI's. OpenAI's revenue, by contrast, leans heavily on its consumer product. For a public investor weighing durability, enterprise revenue is generally regarded as stickier and higher-quality than consumer subscriptions, which churn.

Anthropic also appears to be nearer the thing OpenAI cannot yet show: a quarterly operating profit. In May 2026 the company told investors it expected to post roughly $559 million of operating profit on $10.9 billion of revenue in the second quarter, its first ever. That figure deserves to be read closely rather than at face value, because the picture is contested. The profit is a non-GAAP operating figure, projected mid-fundraise, for a single quarter, and it excludes items such as stock-based compensation that an audited filing would carry. It also coincides with a quirk of timing. SpaceX's own registration statement disclosed that Anthropic had agreed to pay roughly $1.25 billion a month for compute at the Colossus cluster, about $15 billion a year and $45 billion over the life of the contract, with a reduced ramp-up rate during precisely the months Anthropic used to project the profit. The company's reported compute cost per dollar of revenue fell from 71 cents to 56 cents in a single quarter, and Anthropic itself warned investors that profitability might not hold through the year as compute spending rises. A reasonable reading, and the one this desk takes, is that the quarter is real but temporary: revenue is compounding faster than the cost base, and a one-time discount offered an early glimpse of what positive economics might eventually look like, well before they are structurally established. The bear reading, argued forcefully by some, is that the profit is an accounting artifact of a sweetheart supplier rate rather than evidence of a changed business model. Either way, a single discounted quarter is not durable profitability, and it is not audited.

Most recent reported financials, side by side

The two companies do not report on the same basis, which is the first thing to notice. OpenAI's 2025 figures are audited (as reported secondhand); Anthropic's are quarterly projections shared with investors during a fundraising round and are not audited. The tables below place each company's latest available numbers next to the other's, with the basis of each figure labelled rather than blended.

Figures from company statements and press reporting (WSJ, CNBC, The Information, SpaceX S-1, and the Zitron analysis); none audited by YRS. The two columns are deliberately not made to look alike, because they are not: one is an audited full year, the other a projected single quarter on a non-GAAP basis during a fundraise. They are placed together for reference, not for direct subtraction.

The valuation chart in Section 9 contains the single most counterintuitive fact in this comparison: even though Anthropic is valued higher than OpenAI in absolute terms, it trades at a lower multiple of revenue, roughly 21 times its run-rate against OpenAI's 50 — because its revenue base is larger. An investor who believes the AI thesis but is disciplined on price could reasonably conclude that Anthropic, not OpenAI, is the better-priced way to own it. A roadshow will invite exactly that comparison, and it does not flatter the larger ask.

None of this is a verdict that Anthropic is the superior business or the safer investment. It carries its own acute risks, most visibly an unresolved dispute with the U.S. Department of Defense that led to a Pentagon blacklisting of its models and litigation that remains ongoing, a concentration in enterprise customers that cuts both ways, and the same unproven long-run profitability that hangs over the entire sector. Its run-rate figures are reported rather than audited, and run-rate is a generous lens. The point is narrower and more structural: OpenAI is no longer the default frontier-AI investment by valuation, growth rate, enterprise position, or proximity to profit. When its prospectus reaches investors, the first question many will ask is not whether OpenAI is a good business in the abstract, but why it should be worth more than a competitor that is growing faster, priced more cheaply against revenue, and closer to making money. The audited losses make that question sharper, and Anthropic is the reason it cannot be deferred.

A company priced almost entirely on profits expected years from now is, in valuation terms, a long-duration asset, and long-duration assets are the most sensitive in the market to the rate at which the future is discounted. The macro environment of mid-2026 is, for that reason, an active headwind rather than a neutral backdrop.

The proximate pressure is inflation. The May reading of the Federal Reserve's preferred gauge came in at 4.1% on a headline basis, above 4% for the first time since April 2023. Higher-for-longer rates raise the discount rate applied to distant cash flows, which compresses the present value of exactly the kind of far-off profits OpenAI's valuation depends on. The rate path itself is genuinely uncertain: the futures market, as of early July, priced a meaningful probability of a hike at the Federal Reserve's late-July meeting rather than a hold.

The reading is straightforward: with the market assigning roughly a one-in-three chance to a hike at the next meeting and a tightening bias beyond it, the rate environment is not supplying the falling-discount-rate tailwind that lifted growth valuations in easier years. For a business whose entire equity value rests on cash flows arriving toward 2030 and beyond, a higher and more uncertain discount rate is a direct and continuous headwind, independent of anything in the company's own operations.

The interaction is what makes the moment precarious. Hotter inflation lifts discount rates, compressing AI valuations from above. The same inflation is partly driven by the memory-chip shortage that the AI buildout itself created, which raises infrastructure costs, compressing AI margins from below. The technology that is supposed to justify the valuations is contributing to the cost pressure that undermines them. That is an unusual and unstable configuration, and it is the environment into which OpenAI is contemplating the largest listing in market history.

The audited statements do not predict the future; they constrain it. They establish what is true now, enormous growth, enormous losses, a concentrated and leveraged funding base, a contested competitive lead, and from that base a small number of distinct paths open. This section sketches three, with the conditions that would distinguish them.

Scenario A, The window holds

OpenAI lists, at or near $1 trillion, into a market that remains willing to fund the growth story. Revenue continues to compound, the cost-growth gap that opened in 2025 widens, inference deflation flows through to margins faster than price competition erodes revenue, and the company reaches positive cash flow toward the end of the decade roughly as projected. In this path the audited losses read, in hindsight, as the expected cost of building a category-defining asset. The conditions to watch: a successful listing at or above the last private mark, the operating loss narrowing in absolute terms in 2026, and enterprise share stabilising.

Scenario B — The window narrows

The IPO slips to 2027 or lists below the trillion-dollar target. The funding machine, built around an assumed near-term exit, strains: SoftBank's bridge-loan refinancing becomes harder, the circular-financing structure tightens, and OpenAI raises additional private capital at terms less favourable than the trajectory implied. The business is fine, it keeps growing and keeps spending, but the financing becomes the binding constraint, and the premium compresses toward something the financials can more nearly support. This is, on the current evidence, the path the market began pricing on June 26. The conditions to watch: the timing and pricing of the listing, the health of SoftBank's stock, and secondary-market pricing relative to the last primary mark.

Scenario C, The thesis breaks

Demand disappoints at the margin that matters, enterprises conclude, as the most committed adopter already has in one instance, that the bill exceeds the value at scale, or a price war compresses revenue while costs stay sticky, or the macro environment forces a broad repricing of long-duration assets. Any one of these pushes the date of profitability out rather than in, and a company that depends on continuous capital access faces that access on worse terms precisely when it needs it most. This is a tail risk rather than the base case, but a genuine one, and the audited losses are what give it weight: a profitable company can wait out a bad market, while a company losing $20 billion a year at the operating line cannot wait indefinitely.

What an investor should monitor

• The listing itself — timing, valuation, and the disclosures in the public prospectus, which will reveal audited gross margin, the full Microsoft economics, and the scale of forward compute obligations.
• The operating loss trajectory, whether the $20.92 billion narrows in absolute terms in 2026, the single cleanest test of whether the cost-growth gap is durable.
• SoftBank's equity and credit — the most liquid public proxy for OpenAI sentiment and the clearest early indicator of strain in the funding machine.
• Secondary-market pricing, where existing shares actually change hands relative to the last primary mark, and the balance of buyers to sellers.
• Enterprise demand signals — paid conversion, retention, and any further instances of major adopters concluding the economics do not work at scale.

This desk does not issue a price target on a private company, and nothing here is a recommendation to transact. The central conclusion is narrow and defensible. OpenAI is growing faster than almost any company of its scale ever has, and losing more money in the process than almost any company of its scale ever has. A $1 trillion valuation asks the market to bet that the growth wins out, on a timetable the company has not demonstrated, funded through years of loss by capital that has to keep arriving on good terms. The audited statements do not settle whether that belief is correct. They make it possible, for the first time, to see exactly what is being bet.

Step back from the arithmetic for a moment, because the scale of what is unfolding deserves to be named plainly. Whatever the outcome of OpenAI's listing, and whatever becomes of the company in the years after it, this is an event without a true modern precedent. The numbers in this report are large enough that the ordinary frame of reference stops being useful. Markets have not seen anything of this magnitude, on this timeline, in decades, and arguably never in this exact form.

Consider the bookends. A company that did not sell a commercial product a decade ago is contemplating the largest initial public offering in the history of finance, at a valuation that would place it among the fifteen most valuable enterprises on earth from its first day of trading. It is doing so while losing more money, faster, than any company of comparable scale on record. Its nearest rival is racing it to the same milestone within the same quarter. Behind both sits a financing structure measured in hundreds of billions of dollars, threaded through the largest technology companies in the world, and a buildout of physical compute so vast it is reshaping the price of memory, the cost of electricity, and the balance sheets of nations. There is no clean analogue for this in the railway booms, the internet listings of the late 1990s, or the largest debuts of the modern era. The closest comparisons all undershoot.

And that is the honest shape of the conclusion. The magnitude is certain; the direction is not. If the bulls are right, OpenAI's listing will mark the largest act of value creation of the era, the moment a new general-purpose technology converted into one of the most valuable franchises ever assembled, and the audited losses in this report will read, in hindsight, as the price of admission to something genuinely transformative. If the skeptics are right, it will mark the high-water line of a historic mania, the largest and most public test of whether the economics of artificial intelligence ever cohered, and the same losses will read as the early evidence everyone had and chose to look past. Either way the figure is the same: this is the defining financial event of the decade, and possibly of far longer.

What this desk can say with confidence is narrow and, we think, durable. The event is too large to ignore, the stakes are too high to take on faith, and the audited statements have made it possible, for the first time, to weigh both. Whether OpenAI becomes the most important company of the century or the most instructive cautionary tale of it, the listing will be remembered as a hinge point, the moment the market was forced to put a number on the most consequential technology of its time and to decide, in public and with real capital, what it was actually worth. History of that size does not come along often. We are watching it priced in real time.

This report is built on a hierarchy of evidence. The central financial figures derive from audited statements verified by the Financial Times. Secondary figures are sourced to named outlets and flagged by reliability. Forward projections are treated as projections, and conflicts are named rather than resolved.

Provenance of the core figures

All 2024 and 2025 income-statement figures are sourced to audited financial documents first published by Where's Your Ed At on June 15, 2026, and independently verified by the Financial Times, with corroborating reports from Fortune, Ars Technica, and Benzinga the same day. The documents themselves are not public; the figures reported from them are specific and consistent across outlets. OpenAI declined to comment.

Treatment of the headline loss

This report deliberately separates three measures: the operating loss ($20.92 billion), an estimate of true cash burn (approximately $8 billion), and the net loss attributable to the company ($38.53 billion). The gap between the operating and net figures is driven principally by a one-time, non-cash fair-value charge of roughly $41.55 billion arising from the October 2025 conversion. Any analysis that presents the $38.53 billion as a cash loss is in error; any analysis that dismisses the operating loss because the headline is "just accounting" is equally so.

Conflicts explicitly carried, not resolved

• Forward revenue. Reported internal targets (~$200B by 2030, all-in) conflict with lower third-party estimates (~$39B by 2030 for the core business). Different bases; not reconciled.
• Forward burn. 2026–2027 cash-burn projections vary by source and definition. Shown as ranges.
• Enterprise share. Competitive share figures are survey-based estimates that shift quickly and are contested.
• Anthropic figures. All Anthropic revenue, run-rate, valuation, margin, and profitability-timing figures are reported by the company or press and are not audited by YRS. Run-rate annualises recent sales and is not comparable to OpenAI's recognised, audited $13.07B. Anthropic's own audited numbers will not be public until its S-1 is released.
• Comparative multiples. The SpaceX (priced ~94x, first-day close ~112x), Alibaba 2014 (~20x), and Meta (~9x) multiples in Section 9 are drawn from reported figures and standard market data; OpenAI's and Anthropic's are computed against run-rate, not recognised revenue, and are therefore generous to both.
• Filing date. OpenAI publicly confirmed its confidential S-1 on June 8, 2026; reporting indicates the draft was submitted on or around May 22. Both are consistent.

On the charts

Charts are rendered by YRS from the figures cited. Where a series is directional rather than precise, the token-price trajectory, the FedWatch history, and the run-rate ramps in particular, the caption says so. The FOMC probability charts reflect CME FedWatch data as of early July 2026. The 2023 revenue and cost figures are approximate; all 2024 and 2025 figures are from the audited statements.

What this report does not claim

It does not value OpenAI, which is private and which this desk does not cover for transaction purposes. It does not forecast the IPO's timing or pricing. It does not assess the quality, safety, or capability of OpenAI's technology. And it does not constitute investment, legal, or tax advice. YRS holds no position in OpenAI.