Somewhere between $40 billion in private funding and a rumored $300 billion valuation sits a number nobody outside a handful of Microsoft finance executives has ever seen: what it actually costs OpenAI to run its models every single day. That number is about to become public record.
The Valuation Is the Distraction
Every conversation about OpenAI going public has collapsed into the same gravitational well — the headline valuation, the comparisons to Meta’s 2012 debut, the question of whether Sam Altman’s restructuring from nonprofit to capped-profit to full for-profit corporation makes the equity clean enough to sell. These are reasonable questions. They are also the wrong ones. The document that will matter most when OpenAI files its S-1 — which Reuters reported is being prepared for a speedy submission, with a September launch in view — is not a story about upside. It is a story about the burn rate required to produce that upside, and whether any public-market investor has ever been asked to bet on infrastructure costs at this scale before.
The mainstream take on OpenAI IPO economics treats the company as an extraordinarily expensive software business on its way to becoming a normal one. That is precisely backwards. The cost structure is not a transitional problem. It may be the permanent condition of frontier AI — a toll that rises with capability, not one that efficiency eventually eliminates.
$300 Billion Assumes a Cost Curve That Hasn’t Materialized Yet
Investors pricing OpenAI at $300 billion are implicitly betting on dramatic margin expansion. The logic runs: training costs fall as hardware improves, inference gets cheaper at scale, and the company eventually looks like a high-margin platform. Every part of that chain is contestable. The Wall Street Journal’s reporting on the IPO timeline focuses on structure and timing. It does not dwell on the possibility that the S-1’s cost disclosures could reset investor assumptions about what AI scale actually costs to maintain.
Training a frontier model does not get cheaper when your competitor trains a bigger one. OpenAI cannot stand still at GPT-4 while Google and Anthropic push capability forward — the product degrades in relative terms even if the absolute model hasn’t changed. That means the capital expenditure is not a one-time infrastructure buildout. It is an ongoing competition tax. Each new frontier model requires a new cluster, new runs, new evaluation cycles. The $300 billion valuation assumes that tax is manageable. The S-1 will tell us whether it is.
What the SEC Will Force Into the Open
September. If OpenAI hits that target, the filing arrives into a market still digesting the implications of every major tech company’s AI capital expenditure disclosures from 2025 earnings calls. Microsoft, Google, and Amazon collectively announced over $200 billion in planned AI infrastructure spending across 2025 and 2026. OpenAI, as a private company, has operated in the comfortable shadow of those disclosures — close enough to Microsoft’s balance sheet to borrow credibility, distant enough from public markets to avoid line-item scrutiny.
That ends with registration. SEC disclosure requirements for AI companies have grown significantly more demanding since 2023, covering not just financial performance but material risks tied to model dependency, compute concentration, and the governance of systems that drive revenue. OpenAI will have to disclose its compute agreements with Microsoft in enough detail that analysts can reverse-engineer the actual unit economics of running ChatGPT at 500 million weekly active users. That figure — 500 million — has been the company’s most-cited growth metric. The S-1 will attach a cost basis to it for the first time.
“The question isn’t whether the revenue is real. The question is whether the margin profile at scale looks like software or like a utility running on borrowed infrastructure.”
— Portfolio manager, large-cap technology fund
The Microsoft Variable Nobody Prices Correctly
A number that will define how analysts read OpenAI IPO economics: roughly 49%. That is the revenue-sharing percentage Microsoft reportedly holds as part of its multi-billion dollar investment arrangement, a structure that predates the current corporate reorganization and whose precise post-restructuring terms remain opaque. If that figure survives into the public entity in any recognizable form, OpenAI’s effective take-rate on its own revenue is dramatically lower than the gross numbers suggest. If it has been renegotiated away, the S-1 will tell us what OpenAI paid to buy its own economics back.
The dependency runs deeper than revenue share. OpenAI’s compute is substantially provisioned through Microsoft Azure. That is simultaneously a strength — preferential pricing, guaranteed capacity — and a structural exposure that public investors will price differently than private ones. Private investors bet on the relationship holding. Public investors will ask what happens if it doesn’t, and they will want a contractual answer, not a strategic narrative.
Why the Burn Rate Is the Real Moat Argument
Infrastructure. A datacenter that costs $10 billion to build and $2 billion a year to operate is not obviously a moat — unless you believe no competitor can replicate it. OpenAI’s actual argument for durable advantage is not intellectual property in any traditional sense. It is that the combination of proprietary training data, accumulated RLHF feedback from hundreds of millions of users, and the sheer scale of inference infrastructure creates a compounding lead that raw capital cannot easily replicate on a short timeline. That argument is coherent. It is also untested against a public-market cross-examination that will include analysts who covered Amazon Web Services and know what a commodity infrastructure business looks like at maturity.
The burn rate looks like a weakness. The burn rate is also the barrier. These two things are true at once, and the third truth — the one the S-1 will either validate or destroy — is whether OpenAI’s revenue is growing fast enough relative to that burn to make the ratio sustainable without continuous dilution. OpenAI reportedly hit $4 billion in annualized revenue in late 2024 and was targeting $11.6 billion for 2025. Those numbers are impressive in isolation. Against a cost structure that likely runs to multiple billions per year in compute alone, they leave the margin question genuinely open.
What Practitioners Should Watch in the Risk Factors Section
Model obsolescence risk. It will appear in the S-1 risk factors, probably buried in the middle of a long paragraph about competition, and it deserves more attention than it will get. The specific danger is not that OpenAI gets overtaken by a better model — it is that a sufficiently capable open-source alternative reduces enterprise customers’ willingness to pay for API access. Meta’s Llama releases have already begun compressing the low end of the market. The September IPO timeline means OpenAI is racing to go public before that dynamic becomes undeniable in its retention metrics.
Researchers watching the technology mature should pay particular attention to the compute efficiency disclosures. If OpenAI has achieved material gains in tokens-per-dollar over the past 18 months — and there is independent evidence suggesting inference costs across the industry have dropped significantly — the S-1 will show whether those gains are flowing to margin or being immediately redeployed into capability. A company that keeps reinvesting efficiency gains into bigger models is making a strategic choice that favors market position over profitability. That choice is defensible. It is not what a $300 billion valuation typically implies about a company’s proximity to earnings.
The Governance Discount That Isn’t Being Taken
OpenAI’s conversion from a capped-profit structure to a conventional for-profit corporation removed one layer of investor concern and introduced another. The nonprofit board retains certain oversight rights. The exact nature of those rights — whether they constitute a meaningful governance check or a ceremonial one — will be specified in the S-1, and the answer will matter to institutional investors who have spent three years watching AI governance frameworks produce more press releases than enforceable constraints. A dual-class share structure, if OpenAI adopts one, will concentrate voting control with Altman and early insiders in a way that gives public shareholders very limited recourse if the nonprofit remnant ever creates conflict.
That is not an argument against the IPO. It is a pricing variable. Governance discounts are real and measurable — Snap traded at a persistent discount to peers partly because of its share structure. Reuters’ sourcing on the IPO preparation suggests the timeline is being driven partly by market conditions and partly by the company’s need for permanent capital to fund infrastructure that private rounds can no longer comfortably cover. That latter motivation is the one worth taking seriously: when a company has raised $40 billion privately and still needs the public markets, the question is not ambition. The question is arithmetic.
FetchLogic Take
Within 90 days of the S-1 filing, at least two major sell-side initiations will peg OpenAI’s fair value below the IPO price — not because the business is weak, but because the compute cost disclosures will force analysts to model a margin profile that is structurally worse than the private-market valuation assumed. The company will price above those targets anyway, pop on day one, and then spend 18 months volatility-trading between the believers who see a platform and the skeptics who see a very expensive utility. The bet worth taking: OpenAI IPO economics, once fully disclosed, will reset industry assumptions about what frontier AI actually costs to sustain — and that reset will be more consequential for the sector than the IPO itself.
