OpenAI reported $13.1 billion in 2025 revenue against an estimated $9 billion net loss.1 The company was valued at $500 billion in an October 2025 share sale.1 Microsoft owns 27% of OpenAI Group PBC1 and restructured its deal to lock in a 20% revenue cut through 2030.2 When OpenAI eventually files to go public, its S-1 will contain the first audited breakdown of inference compute costs, per-segment gross margins, and the full text of the Microsoft commercial agreement. The entire LLM vendor cohort has operated under a private-market disclosure regime that makes margin opacity the default. A public S-1 ends that.
What an S-1 Actually Discloses and When
A confidential S-1, filed under the JOBS Act provisions for emerging growth companies, lets a company submit its registration to the SEC without immediate public release. The filing is reviewed in private, amendments happen behind closed doors, and the public only sees the document when the company chooses to release it, typically 15 days before its investor roadshow. OpenAI’s PBC structure, adopted in 2025,1 was widely interpreted as clearing the governance path for a future public offering.
The disclosure impact is delayed, not simultaneous with the IPO decision. When a public S-1 drops, analysts will have audited financials for the first time: revenue by segment, cost of revenue broken out (including inference compute), R&D spend, and the full text of the Microsoft commercial agreement. That last item is the one competitors should be worried about.
The Margin Picture: Revenue, Losses, and Microsoft’s Cut
OpenAI’s $13.1 billion in 2025 revenue against an estimated negative $9 billion net loss1 tells the story in two numbers: a business burning two-thirds of its revenue, valued at 38x trailing sales in its last private round.1
Microsoft’s revenue share compounds the margin pressure. Under the restructured deal, Microsoft takes 20% of OpenAI’s revenue through 2030.2 Microsoft also owns 27% of OpenAI Group PBC,1 and OpenAI depends on Azure for its compute infrastructure.1
For margin analysis, the key question is what sits between revenue and that 20% payout.2 Inference compute is the dominant cost. Model training is capital expenditure with a long amortization window; inference is recurring, variable, and scales with every user query. An S-1’s cost-of-revenue line would be the first credible data point on whether current pricing covers inference costs at any tier, from ChatGPT Plus subscriptions to API calls to enterprise deployments.
Microsoft’s Spending Signal and What It Implies About Agent Costs
Microsoft’s Q3 FY2026 results showed capex rising 49% to $31.9 billion.2 Azure grew 40% year-over-year,2 and the AI business hit a $37 billion annual run rate, up 123%.2 Those are headline beats. The subtext: Microsoft’s shares fell roughly 2%2 on concerns about sluggish Copilot adoption and competitive pressure. When the company selling the compute infrastructure for AI agents is simultaneously signaling that Copilot uptake is slower than expected, the unit-economics question answers itself.
Microsoft also launched its first-ever voluntary buyout program,3 targeting 7% of its US workforce, roughly 8,750 employees,3 while simplifying pay tiers from nine to five. This follows 9,000 layoffs the previous summer.3 A company spending $31.9 billion2 on infrastructure in a single quarter while trimming headcount is managing the margin pressure that infrastructure spending creates.
What a Public Comparable Does to Every Private LLM Vendor
An OpenAI S-1 creates a problem that Anthropic, Cohere, Mistral, and every other private LLM provider has avoided: a public, audited reference point for frontier-model financials.
Until now, the sector has operated with selective disclosure. Revenue figures appear in press interviews or leaked fundraising decks. Cost structures are discussed in vague terms. Gross margins are, across the board, undisclosed. Anthropic generates significant API revenue but has never filed a public document breaking out inference costs versus training costs versus R&D. The same is true for every competitor.
An OpenAI S-1 changes the equilibrium in three specific ways.
First, analysts get a real gross-margin number to model against. If OpenAI reports, say, 40% gross margins on API revenue, every private company telling investors they expect 60% faces a pointed question about why their cost structure is superior.
Second, the cost-of-revenue breakdown gives a benchmark for inference pricing. If OpenAI’s inference cost per token is public (even aggregated), competitors can no longer claim that margins are improving without providing their own evidence.
Third, the Microsoft commercial agreement, disclosed in full, becomes a template for understanding what infrastructure partnerships actually cost. The 20% revenue share2 is already known. An S-1 would reveal the fine print: minimum commitments, volume discounts, termination clauses, and the exact structure of the deal through 2030.
Microsoft’s Position: Revenue Share, Capex, and Regulatory Heat
Microsoft’s incentives in a future OpenAI IPO are not straightforward. The 20% revenue share through 20302 is a direct claim on OpenAI’s top line. Microsoft’s overhaul of the deal2 locked this in, presumably because the previous terms were less favorable or less certain.
But Microsoft is also deploying capital at an unprecedented rate. The $31.9 billion quarterly capex figure2 represents real spending on data centers, GPU clusters, and networking infrastructure, much of it designed to serve OpenAI’s workload and Microsoft’s own Copilot products. If Copilot adoption is sluggish, as the market reaction suggests, the return-on-investment timeline extends.
Simultaneously, Britain’s CMA launched an antitrust investigation4 on May 14, 2026, examining the bundling of Windows, Word, Excel, Teams, and Copilot. The probe concludes in February 2027. Active EU and US antitrust investigations are also underway. Microsoft’s strategy of embedding AI into its existing product suite is the exact behavior regulators are now scrutinizing.
An OpenAI S-1, when it eventually becomes public, will not just be a financing event. It will be the first time the unit economics of frontier-model inference are laid out in a document subject to SEC enforcement for accuracy. Every claim about margins, every fundraising deck asserting cost improvements, every analyst model built on estimated inference costs will be tested against a single audited filing. That is the structural consequence, and it applies regardless of whether OpenAI prices at $500 billion1 or half that.
Frequently Asked Questions
What is OpenAI’s total Azure spend commitment under the restructured deal?
OpenAI agreed to purchase $250 billion in Azure services over the life of the agreement. This contractual obligation would appear on the S-1 balance sheet as a committed expense and represents a fixed cost floor that compounds the margin pressure from the 20% revenue share to Microsoft.
Can the Microsoft revenue share end before 2030?
Yes — the deal includes an AGI termination clause: the 20% revenue share obligation ends if an independent panel verifies that OpenAI has achieved artificial general intelligence. The margin structure could shift overnight, but only via a determination that would simultaneously upend the commercial rationale for the entire partnership.
Who holds the rest of OpenAI’s equity besides Microsoft?
The OpenAI Foundation holds 26% of OpenAI Group PBC, and employees and investors hold the remaining 47%. The Foundation’s stake was designed to preserve the original nonprofit mission, but a post-IPO scenario creates a three-way governance dynamic — public shareholders, Microsoft at 27%, and the Foundation at 26% — that the PBC structure was specifically built to manage.
Could OpenAI avoid disclosing agent revenue separately in the S-1?
Yes. SEC segment reporting rules require disclosure of material business lines, but OpenAI could argue that ChatGPT, API, and enterprise deployments constitute a single reporting segment. If regulators accept that framing, the S-1 would show only blended margins — making inference cost-of-revenue a proxy for agent-tier economics rather than a direct measure.