OpenAI is a genuinely improving business — scored against the Burn Mask the month its IPO filing went public. The finding is not the headline you would expect. The danger is in the financing, not the company — and one event settles it.
The danger is in the financing, not the company.
Start with where you keep your money, not where OpenAI keeps its losses. If you hold an index fund, a 401(k), or a pension, you are already invested in the artificial-intelligence build-out — and that build-out has quietly become a large part of the economy itself.
Here is the uncomfortable arithmetic. Big Tech is spending hundreds of billions a year building AI infrastructure, and analysts estimate that to justify it the industry needs to generate roughly $2 trillion in annual revenue by 2030 — many times what it earns today. The companies doing the spending are now about 35% of the S&P 500, a heavier concentration than the dot-com peak. JP Morgan Asset Management has noted that AI-linked stocks have driven the lion's share of the index's returns, earnings growth, and capital spending since ChatGPT launched.
If the AI story falters, it is not only venture investors who are exposed. It is ordinary people — through index funds, retirement accounts, the consumer spending propped up by paper stock gains, and the share of GDP growth this spending currently supplies. The risk has been quietly socialized. That is why this investigation leads with the economy, not the company.
And OpenAI is the keystone. Much of that infrastructure spending is built on the assumption that OpenAI's demand arrives as promised — a $250 billion Azure purchase commitment, the $500 billion Stargate plan, large deals with Nvidia and Amazon. If OpenAI's economics work, the spending is justified. If they do not, the most important number in the market loses its anchor. So the right question is not "will OpenAI survive." It is "what is the company actually worth, and who is holding the risk if the answer is less than the price."
Investigation #1 used the Goodwill Mask — built to catch mature companies hiding a dead core behind acquisition goodwill they refuse to write down. OpenAI does not fit it. It is young, it carries no acquisition mask, and its product demand is real. A wrong instrument produces a wrong verdict, so we built a second one.
The Burn Mask catches a different thing: a company hiding broken or unproven unit economics behind a profitability story that keeps outside capital flowing. The test is not "is it losing money" — almost every young company does. The test is whether the losses are narrowing on a real path to self-funding, or whether the story is doing the work the economics cannot yet do.
The Goodwill Mask is a dead tree with its leaves glued on — the canopy looks full until the storm. The Burn Mask is etiolation: a seedling growing fast and pale, stretching tall toward a light it has not reached, putting its energy into height instead of roots. It can grow impressively for a long time. The question is whether it develops roots — economics that fund themselves — before the support is removed.
The instrument runs four gates before it scores. Gate 0 — does the product exist (a fake product is fraud, not a Burn Mask). Gate A — is it really burning outside capital. Gate B — does it depend on continuous funding. Gate C — the Amazon Gate: if the core economics are visibly improving on a clear path, the company is cleared, exactly as Amazon would have been in 2001. Only what passes all four gets the nine-signal score.
Before OpenAI was scored, the Burn Mask was tested against six known cases and matched every one: WeWork (18/20, weeks from zero before its rescue), MoviePass (20/20, selling dollars for cents), Webvan (19/20), and the disconfirming controls — Amazon 1997–2002 (cleared at Gate C), Uber 2016–2023 (a Watch that earned its way out as the core turned profitable), and Theranos (screened out at Gate 0 as fraud). This is a methodology, not a hot take.
In June 2026, leaked audited financials — obtained by journalist Ed Zitron, independently verified by the Financial Times, and reported by Fortune, Bloomberg and others — gave the first detailed look under OpenAI's valuation. The company declined to comment. The numbers are extraordinary in both directions.
Read the two columns honestly. Revenue tripled, the loss per dollar earned is falling, and 900 million people use the product every week — that is a real, growing business, not a mirage. But total costs hit $34 billion, research and development alone exceeds the entire top line, and the company's own target for profitability is around 2029–2030. The widely shared "$38.5 billion net loss" is real but misleading — it is inflated almost entirely by a one-time, non-cash $41.55 billion charge from converting to a for-profit structure, money that never left the company. The honest operating figure is the $20.9 billion.
| Line | 2024 | 2025 | What it says |
|---|---|---|---|
| Revenue | $3.7B | $13.07B | Growth with almost no precedent at this scale |
| Operating loss | $(8.78B) | $(20.92B) | Widening in dollars |
| Cost of revenue | $2.65B | $7.5B | Inference — the cost of serving every answer |
| R&D | $7.8B | $19.18B | $10.59B of it paid to Microsoft for compute |
| Sales & marketing | $1.11B | $5.73B | Up more than fivefold — buying growth |
One number is the crux, and it is genuinely contested: the gross margin. On the paid product alone, OpenAI's "compute margin" reportedly roughly doubled toward 70% (The Information) — the engine of serving a paying user is improving fast. But a fuller, blended view that carries the cost of the free tier and the heavier "reasoning" models puts the margin closer to 33% and possibly falling (Sacra). Whether margins are improving has three credible answers depending on which one you pick — and that ambiguity is itself a Burn Mask signal, not a footnote.
OpenAI passes the first three gates and does not cleanly clear the fourth — so, by the instrument's own rule, it is scored as a live case, with the expectation that the core looks like a Watch.
Real product, real audited revenue, 900M users. The opposite of Theranos. No fraud here.
A $20.9 billion operating loss, funded by outside capital. Unambiguously burning.
Needs continuous capital — no cash-flow breakeven expected before ~2030 — but sits on the largest war chest in startup history, which lowers the acuteness.
The paid-serving engine is improving, but the company-level path to self-funding is not yet demonstrated, and a looming price war threatens margins. That is the Uber situation: score it, and expect a Watch-shaped core.
| # | Signal | Evidence | Score |
|---|---|---|---|
| 1 | Burn-to-breakeven trajectory | Ratio improving ($2.37→$1.60 loss per dollar), but absolute burn growing (~$27B projected 2026). Closing in ratio, widening in dollars. | 1 / 3 |
| 2 | Unit economics | Opposite of below-cost: paid "compute margin" reportedly ~35%→~70%. But blended margin contested (~33%, possibly down). Improving too slowly to call clear. | 1 / 3 |
| 3 | Runway depth | ~$50B in assets, roughly half cash, plus the $122B round. Years of runway. The strongest disconfirming signal. | 0 / 2 |
| 4 | Valuation-to-cash gap | $852B now, ~$1T IPO target, against zero company-level profit. Bridgewater reportedly told clients the price implies "a monopoly outcome that does not yet exist." | 2 / 2 |
| 5 | Narrative dependency | Valuation rests on an unproven 2029–2030 profitability story while R&D alone exceeds revenue. The "70% compute margin" is the textbook custom metric over a murkier reality. | 2 / 2 |
| 6 | Moat / commoditization | Real brand and distribution moat (900M users), but eroding: token prices fell ~$10→~$2.50 in a year, and OpenAI is weighing cuts to fend off Anthropic. | 1 / 2 |
| 7 | Capital-structure fragility | A $250B Azure commitment and 20% revenue share to Microsoft; circular vendor financing across Nvidia/Amazon/Microsoft; SoftBank funding its stake with heavy leverage. | 2 / 2 |
| 8 | Insider incentive to sustain the story | A $6.6B employee share sale at a $500B mark; the Foundation's warrant pays only if value rises tenfold; SoftBank's balance-sheet motive. Structural, not fraud. | 2 / 2 |
| 9 | The catalyst | Dated and specific: a confidential S-1 filed ~May 22, an IPO as early as September 2026 at up to $1T — into which SoftBank's refinancing wall is wired. | 2 / 2 |
This is the whole finding in one picture. OpenAI is a genuinely improving business wrapped in a fragile, narrative-funded, IPO-dependent capital structure. The seedling is finding some light — it is the borrowed, circular, reflexive capital holding the ladder that could give way. The zombie risk is in the financing, not the machine.
Stated as hard as the bear case: OpenAI clears toward healthy if the paid-margin gains keep outrunning the heavier reasoning models, if the Microsoft revenue share (now capped through 2030) stops dragging cash flow, if the enterprise mix keeps improving, and if the IPO lands cleanly and refinances the fragility. None of that is fantasy — which is exactly why this is a 13 and not a 17.
The most distinctive thing about OpenAI's risk is not on its income statement. It is in the shape of the money around it. The companies that supply OpenAI are also the companies funding it — and in some cases booking its spending as their own revenue.
Then there is SoftBank — not an operating company but a leveraged holding company, and the distinct fragility in the structure. It has committed tens of billions to OpenAI and is funding it by burning its own balance sheet: selling its entire Nvidia stake and part of its T-Mobile stake, carrying a $40 billion bridge loan due within a year, with its loan-to-value near the ceiling it set for itself. That is fragility stacked on fragility. It does not make OpenAI WeWork — the fundamentals are completely different — but it is the same enabler pattern, and it is wired directly into the IPO.
Because the risk is in the financing, the thing that resolves it is a financing event: the IPO. It is the fork. The scorecard is explicitly provisional, and this is why.
A well-priced offering at a defensible multiple demonstrates the path, refinances the SoftBank fragility, and converts the narrative into proof. The seedling puts down roots. The capital-structure flags ease.
A postponed or weak offering — especially into a token price war — fires the catalyst. The refinancing wall arrives with no exit. The support is removed before the roots are down.
This is the Uber rule from the calibration set: a Watch is a moment in time, not a sentence. The honest move is to re-score OpenAI the day after the IPO prices — the inputs to half the signals change that day. Until then, 13/20 is a reading with a date on it, not a verdict for all time.
Put the two investigations side by side and the AI market stops looking like separate companies and starts looking like one connected system — with two different disguises operating inside it.
A mature structure folding xAI's losses into a space-infrastructure valuation story through merger goodwill it has not written down. The disguise is on the balance sheet.
A young, improving business funded and priced on an unproven profitability story, inside a circular capital structure. The disguise is in the financing.
xAI is the bridge — it wears both masks. As the all-stock SpaceX merger, it is a Goodwill Mask. As an AI lab, it is a textbook Burn Mask: SpaceX's own IPO filing shows xAI lost $6.36 billion on $3.2 billion of revenue in 2025, with $12.7 billion of capital spending — more than the rest of SpaceX combined — and 117 million Grok users of whom only about 1.9 million pay.
And the rivals are financially entangled. In May 2026, Anthropic agreed to pay xAI about $1.25 billion a month for compute through xAI's Colossus data center — terminable on 90 days' notice. xAI needs the revenue to offset its burn; Anthropic reaches profitability partly by renting compute from a competitor rather than building its own. Nvidia funds both. It is one capital-recycling system, not three companies — and SpaceX's own IPO, with an analyst fair value roughly 55% below its price, shows how index-inclusion mechanics push retirement money into it.
The finding cuts against both easy narratives. "OpenAI is a fraud" is wrong — the business is real and improving. "OpenAI is unstoppable" skips the part where the price assumes an outcome that has not happened yet, funded by capital that has to keep believing. Here is what the verdict implies for the people actually exposed.
So the one thing to watch is not OpenAI's next model or its next user milestone. It is whether the financing keeps converting narrative into proof — starting with how the IPO prices. The machine is doing its job. The question has always been about the money around it.
In the interest of being straight with you: this investigation was researched with the help of an AI assistant (Claude) made by Anthropic — one of the companies discussed here. Anthropic, OpenAI, and xAI were all held to the same standard, and the analysis deliberately treats the circular financing as a property of the whole system, not a charge against any one company. Anthropic is noted fairly as the one of the three projecting an operating profit. Where a comparison touches Anthropic, weigh it accordingly.
This is research, not investment advice, and not a recommendation to buy, sell, or hold anything. The Burn Mask Scorecard produces flags for investigation — questions worth asking — not accusations and not a verdict on any company's solvency.
OpenAI's 2025 figures come from leaked financials verified by the Financial Times and reported by multiple outlets; OpenAI has not confirmed them. xAI's figures come from SpaceX's public IPO filing. Macro and margin figures are attributed to their sources in the text and reflect estimates that reasonable analysts dispute. Every figure is dated; private-company numbers change. Nothing here is invented.
This verdict is provisional and time-stamped. It should be re-scored after OpenAI's IPO prices. Do your own work, and consult a licensed professional before making any financial decision.