DeepSeek Makes $500 Million a Year and Is Valued at 163 Times Revenue

The compute gap is closing. The revenue gap is not.

Yesterday Huawei set its chip roadmap against Nvidia↗: the 960DT lands in the first quarter of 2027, the Ascend 960PR in the third, and its "superclusters" scale toward a million processors ⁵. Today the other ledger got priced: every Chinese AI model together generates roughly 10% of the revenue reported by OpenAI and Anthropic, Rhodium Group finds in estimates published Thursday↗ ². Chinese models also cost far less per task than their mostly closed US rivals, per Artificial Analysis ². Cheap inference has bought capability, not a business.

A tenth of the revenue, lab by lab

Rhodium puts OpenAI at $40 billion of annual recurring revenue and Anthropic at $65 billion ². ARR multiplies a recent month by 12, so these are run-rates, not audited sales ². China's leaders: ByteDance at $4 billion, Alibaba at $2.4 billion, Z.ai at $1.8 billion as told to investors on Wednesday, Moonshot at $1 billion, MiniMax at $800 million, DeepSeek at $500 million ².

The six sum to about $10.5 billion against a combined $105 billion, roughly the tenth Rhodium describes. A wider tally↗ of all Chinese models reaches $10.7 billion ³.

The richer multiples sit in China

The standard bear case says US investors overpay for AI. Per dollar of revenue, the richer price sits in China: Rhodium puts DeepSeek at 163 times revenue and Moonshot at 50, against 34 for OpenAI and 21 for Anthropic, and calls the first two "exorbitant" ¹².

DeepSeek's ratio implies a mark near $80 billion on half a billion of sales, arithmetic the report leaves to the reader.

Cheap weights are hard to charge for

Open weights let anyone with capable hardware run the model without paying its developer, and Chinese labs are still hunting for a larger cut of third-party access ². Closed US models, which Artificial Analysis prices far higher per task, meter use and convert it into recurring revenue; that incumbency in monetization is who benefits from this gap ².

The capital behind each side widens the difference. More than 60% of equity investment in Chinese AI chips and servers came from state-affiliated sources ². Beijing's AI capital spending doubles this year to $139 billion ³. Wright expects government funding, helpful on hardware, to "probably balk" at bankrolling frontier labs directly ².

That is the financing gap: it "will be far more difficult for Chinese frontier AI labs to scale sustainably," Wright said, leaving them "heavily dependent upon a favorable climate in the equity market," historically not an easy bet in China ². Government money built the compute. It has not bought the customers.

The gap can still close: Rhodium's data is summer-vintage

Rhodium's figures could draw only on data from this summer, and Chinese usage is compounding from a low base ². Z.ai lifted its year-end target from $2.4 billion to $3 billion and rose more than 5% on Thursday ². The US tower has its own crack: the top 1% of customers supply 80% of revenue at both OpenAI and Anthropic, per Ramp ³.

The listings will now price both ledgers: Anthropic reportedly next month, Moonshot's $3 billion Hong Kong raise, DeepSeek's STAR Market filing↗ through CITIC ⁴. Until revenue converts, capability is the only gap closing, and the investors overpaying for AI are the ones in China.