OpenAI's internal metric for inference profitability hit 70% in October 2025, according to financial data obtained by The Information. The figure, which measures revenue retained after covering the direct costs of running AI models for paying users, doubled from roughly 35% in January 2024. The company declined to confirm the numbers.
The margin that matters (and the one that doesn't)
"Compute margin" sounds like a profit metric. It isn't, really. The figure captures a narrow slice of OpenAI's cost structure: what's left after GPU rental, electricity, and maintenance for serving requests to paying customers. It says nothing about training new models, compensating researchers, or building data centers.
Still, 70% is notable. Traditional cloud computing businesses typically operate between 30% and 50% gross margins. For a company that was hemorrhaging money on every API call just two years ago, the improvement suggests OpenAI has gotten substantially better at wringing efficiency from its inference infrastructure.
The Information reported that OpenAI now outperforms Anthropic on compute margins for paid accounts, though Anthropic shows better overall server efficiency. That distinction matters: Anthropic may be spending its compute dollars more wisely across the board, while OpenAI extracts more from each paying customer specifically.
Where the money actually goes
The compute margin gains haven't translated into anything resembling profitability. OpenAI reported a $13.5 billion net loss for the first half of 2025 on roughly $4.3 billion in revenue. The company spent $6.7 billion on R&D alone during that period, a figure that likely includes the server infrastructure needed to develop new models.
A large portion of that net loss stems from non-cash factors, specifically the remeasurement of convertible interest rights issued to investors. But even the underlying cash burn tells an uncomfortable story: OpenAI consumed $2.5 billion in cash during H1 2025, roughly $14 million per day.
The company's spending breakdown for the first half reads like a startup burning through a war chest: $2 billion on sales and marketing, nearly $2.5 billion in stock compensation, plus whatever infrastructure investments didn't land in the R&D line item.
Revenue growth is real, sustainability is not
OpenAI's top line has grown rapidly. The company reported $12 billion in annualized revenue by July 2025, roughly double its December 2024 run rate. CEO Sam Altman said in November the company expects to exceed $20 billion in annualized revenue by year's end. The original 2025 projection was $12.7 billion.
But annualized revenue figures deserve skepticism. They represent a single month's performance multiplied by twelve, a metric that startup investors love because it smooths volatility and implies forward momentum that may or may not materialize. OpenAI's actual first-half revenue was $4.3 billion. The math doesn't lie, but the framing sometimes does.
The company projects positive cash flow by 2029, with revenue targets of $125 billion that year and $200 billion by 2030. Those numbers would require growth rates that dwarf even Google's early trajectory.
What the competition looks like
Google's Gemini model outperforming OpenAI on certain benchmarks prompted Altman to call a "code red" internally, redirecting resources toward ChatGPT improvements and delaying plans for an advertising product. Most ChatGPT users still access the free tier, which doesn't contribute to compute margin calculations at all.
OpenAI is pushing enterprise products and paid features in financial services and education, markets where it competes directly with both Google and Anthropic. The company is also in talks to raise at least $10 billion from Amazon, potentially at a $500 billion valuation or higher, a deal that would include access to Amazon's custom AI chips.
Anthropic, for context, reached $5 billion in annualized revenue by late 2025 and is reportedly raising at a $170 billion valuation. The AI infrastructure arms race shows no signs of slowing.
OpenAI's $40 billion funding round, led by SoftBank in early 2025, valued the company at $300 billion. A secondary share sale in October pushed that figure to $500 billion. The company has raised approximately $58 billion in total funding and has committed to over $1.4 trillion in infrastructure spending through the early 2030s.
Whether 70% compute margins on paid products can support those commitments remains the central question. The efficiency gains are real. So are the losses.
