DeepSeek CEO reveals 85% inference margins, $1B API revenue, and 10-month GPU payback in rare investor call

DeepSeek's Infrastructure Economics Point to Scaled, Self-Sustaining AI Operations

DeepSeek CEO Liang Wenbong spent four hours on an investor call laying out the operational math behind the Chinese AI lab's rapid ascent. The metrics he shared suggest a company hitting profitability faster than its Western competitors—and one that may be solving the capital intensity problem that has constrained the broader AI industry.

The headline figures are striking. DeepSeek claims $1 billion in annualized API revenue, 85% gross margins on inference, and a 10-month GPU payback period. Those last two numbers compound into the core insight: the company is generating enough cash from API sales to replace its compute infrastructure in less than a year, creating a self-funding flywheel for capacity expansion.

For context, the typical concern with AI infrastructure is that companies must raise staggering amounts of capital upfront—OpenAI recently raised its projected data center spending to $750 billion through 2030—and then spend years scaling usage before reaching cash flow positive. DeepSeek's 10-month payback implies it has largely solved that sequence. Once a GPU pays for itself in revenue within ten months and is depreciated over three to five years, reinvesting margin dollars back into more capacity becomes math, not faith.

The 85% margin on inference is the lever. It means the lab is running a tight operation on the compute side—either through exceptional efficiency, aggressive cost management, or both. Paired with the API revenue figure, it suggests DeepSeek is not a lab burning money on frontier model training in the hope of eventual dominance. It is a business.

Wenbong also signaled a focused strategy. He said the lab would not pursue world models, image generation, video generation, or a consumer chat app—a deliberate narrowing that echoes the "focused culture, talent dense teams" operating model that Anjney Agarwal noted at AMP when comparing frontier labs to larger incumbents. DeepSeek is not chasing Alibaba's sprawl. It is staying on path to AGI.

The disclosure came during a capital roadshow—DeepSeek is now raising outside capital, moving beyond its original funding from a hedge fund parent—and represents a rare window into the financial mechanics of a Chinese AI lab. Western labs typically do not share this level of operational detail with investors, particularly around margins and payback periods. The contrast itself signals confidence, or at minimum, a willingness to compete on economics rather than narrative alone.

The broader implication is that if these numbers hold, the capital constraint on AI development may not be as binding as hyperscalers fear. A company with cash flow positive inference and sub-year GPU payback can fund its own scaling. That reframes the competition not as a race to raise the most capital, but as a race to build the most efficient operations.