OpenAI crossed twenty-five billion dollars in annualized revenue while projecting fourteen billion in losses. The defining question: does AI inference follow the AWS playbook or the WeWork playbook? The structural difference is in the marginal cost curve.
OpenAI reached twenty-five billion dollars in annualized revenue in February 2026, making it the fastest software company to cross that threshold. It projects fourteen billion dollars in losses for 2026. It raised one hundred and twenty-two billion dollars in March at an eight-hundred-and-fifty-two-billion-dollar valuation. It plans to go public before the end of the year at a price approaching one trillion dollars.
These numbers are not contradictory. They are a bet. The bet is that AI inference costs follow the same trajectory as cloud computing costs: hemorrhage cash for years while the marginal cost curve bends toward zero, then collect monopoly rents once competitors exhaust their capital. Amazon Web Services ran this playbook from 2006 to 2015. The question is whether the analogy holds.
It does not.
The Structural Difference
AWS launched in 2006 and first reported a profit in the first quarter of 2015. Nine years of losses. Revenue grew from twenty-one million to nearly eight billion dollars over that period. The losses were real. But the critical structural fact is that AWS had gross margins above eighty percent by 2013. Amazon simply chose not to report the segment separately until the margins were undeniable. The profitability was always there at the unit level. Each additional customer cost almost nothing to serve because compute infrastructure has near-zero marginal cost once built.
OpenAI's inference costs are moving in the opposite direction. Gross margins fell from forty percent in 2024 to thirty-three percent in 2025. Inference costs quadrupled year over year: from roughly two billion dollars in 2024 to eight point four billion in 2025 to a projected fourteen point one billion in 2026. The company spends roughly one dollar and fifty-six cents for every dollar it earns at its current run rate.
This is the structural difference that breaks the AWS analogy. Cloud hosting is a fixed-cost business. Once the servers are built, serving another customer costs electricity and bandwidth. AI inference is a variable-cost business. Each query consumes compute proportional to model complexity. Each new model generation is more expensive to run, not less. GPT-5 costs more per query than GPT-4 cost per query. The marginal cost curve is not bending toward zero. It is bending away from it.
The Anthropic Gap
Anthropic reached thirty billion dollars in annualized revenue by April 2026, surpassing OpenAI while spending roughly one quarter as much on model training. It claims gross margins near forty percent versus OpenAI's thirty-three. It projects profitability by 2027. OpenAI's internal projections push profitability to 2029 or 2030.
The gap is not merely operational. It is architectural. Anthropic derives eighty percent of revenue from business customers. OpenAI derives roughly two-thirds from consumer subscriptions, a segment now under severe price pressure as users migrate from a twenty-dollar plan to a new eight-dollar tier. Replacing a twenty-dollar subscriber with an eight-dollar subscriber requires two and a half times the users to maintain revenue while compute costs scale linearly with usage.
OpenAI disputes the comparison. It argues that Anthropic's gross revenue accounting overstates by roughly eight billion dollars. Even granting the adjustment, the cost structure divergence remains. One company is approaching margins that support independence. The other is planning to raise capital continuously through 2030.
The Forcing Function
The IPO creates a hard disclosure event. OpenAI has targeted the fourth quarter of 2026 for a public listing, though CFO Sarah Friar has told industry insiders the company may not be ready by then. Amazon committed thirty-five billion dollars of its investment contingent on an IPO by the end of 2028. SoftBank committed thirty billion. The capital structure contains a clock.
An S-1 filing will force disclosure of unit economics that private markets have not required. Quarterly gross margins. Customer acquisition costs. Revenue per user trends. The consumer-to-enterprise mix. The gap between inference costs and inference revenue at the model level. These numbers exist inside the company. They do not yet exist in public.
The filing transforms a narrative into an accounting statement. If the margins are improving toward forty percent, the AWS analogy has evidence. If they continue declining, the WeWork comparison becomes unavoidable: revenue growth masking unit economics that deteriorate with scale rather than improving.
What to Watch
Four predictions, each falsifiable within twelve months.
First, OpenAI will not achieve gross margins above forty percent in any quarter of 2026. Inference cost growth will outpace revenue growth because model complexity increases faster than efficiency gains reduce serving costs.
Second, the IPO will be delayed to the first half of 2027. The S-1 preparation process will reveal unit economics that do not support a trillion-dollar valuation in the current interest rate environment.
Third, Anthropic will report positive free cash flow before OpenAI does. Lower training spend, higher enterprise mix, and a business model that does not depend on consumer subscription volume create a structurally shorter path to profitability.
Fourth, at least one investor from the one-hundred-and-twenty-two-billion-dollar round will attempt to sell secondary shares at a discount within twelve months of the investment. The round's size and valuation create paper gains that institutional risk management will seek to realize before the IPO resolves the unit economics question.
The bleed rate is the metric that separates infrastructure bets from broken businesses. AWS bled cash while margins were structurally sound. WeWork bled cash while margins were structurally broken. OpenAI's margins are declining. The IPO will force the answer into public view. Until then, the fastest-growing software company in history is spending more than fifty cents beyond what it earns on every dollar of revenue, and that number is getting worse.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)