DeepSeek's Annualized Revenue Tops $1 Billion, Plans $7.5 Billion Fundraise Ahead of IPO
nashnova research
DeepSeek's annualized revenue has topped $1 billion, more than doubling in months; the company is raising $7.5 billion and preparing a Shanghai listing — but a compute shortage and a regulatory probe hang over the lofty valuation.
Revenue doubled — where is the money coming from?
Annualized revenue run rate has crossed $1 billion, up from under $500 million just a few months ago.
Two drivers: a price hike last month that raised customer costs 2.3 to 4.5 times, and sustained demand — users stayed despite the increase.
This means → DeepSeek proved that sharp price increases stick when your model is strong enough and still the cheapest mainstream option.
Why is its margin higher than OpenAI's?
Gross margin on API sales hit 82.9% in the first seven months of this year — above both Anthropic and OpenAI over the same period.
In plain terms = for every dollar earned, DeepSeek keeps more than 80 cents; its rivals keep less.
The edge comes from infrastructure: optimized efficiency and lower chip usage push down the cost of running each model.
This reflects a structural advantage baked into the architecture, not just a pricing tactic.
A $7.5 billion raise at a $70 billion valuation — how big is this bet?
The second funding round targets 50 billion yuan (roughly $7.5 billion), at a 500-billion-yuan valuation, with a deadline of late October this year.
The company is simultaneously preparing to list on the Shanghai Stock Exchange.
This means → if the round closes, DeepSeek joins the very top tier of global AI startups by valuation, on par with OpenAI's scale.
Seventy percent of compute goes to training — what about inference?
More than 70% of DeepSeek's compute is allocated to training new models; under 30% is left for inference — running existing models to serve user queries.
CEO Liang Wenfeng said internal tests show smaller models run well on consumer gaming GPUs, handling most everyday tasks.
In plain terms = top-tier chips train the frontier models; cheaper gaming cards handle the day-to-day — a workaround for a compute-starved environment.
Chips are the bottleneck — can Huawei fill the gap?
U.S. export controls have cut off Nvidia's advanced chips, making compute scarcity DeepSeek's biggest challenge.
Liang said the company will prioritize training on domestic Chinese chips and expects Huawei to begin delivering training chips as early as Q4 this year.
This means → Huawei's actual chip performance will directly determine whether DeepSeek can break its Nvidia dependency and keep iterating at pace.
A regulatory probe and rising rivals — pressure on two fronts?
Chinese internet regulators are investigating DeepSeek over a potential data leak, after Anthropic accused DeepSeek and Moonshot AI of routing sensitive user data to the Claude model.
Competition is intensifying: Z.ai says its annualized revenue has reached $1.8 billion; MiniMax surged from $150 million in February to $800 million by August.
This reflects a reality beyond technology — compute constraints, regulatory heat, and fast-closing rivals converge into the central test of whether this round's lofty valuation can hold.
市场有风险，内容仅供研究参考，不构成投资建议。