DeepSeek restarts its funding round, plans to raise nearly $8 billion. The core issue is not that it “can’t make ends meet,” but that after the large-model race shifts from “training-driven” to “inference-driven,” compute consumption has surged exponentially 📉.
The company’s global API usage has already hit 80 trillion tokens in a single day. The compute burned by an Agent per task is 10–100 times that of a conversational request. Add to that the near-complete rewrite of underlying code required for V4 to fully adapt to Huawei Ascend; the cost of single-round training has already exceeded $500 million. With the buildout of a GW-scale intelligent computing center, in-house inference chips, and the core team expanding by at least twofold—this system-level expansion path of “model + compute + chips” can no longer be covered by its own funds.
The deeper driving force is talent and commercialization: since the end of 2025, key researchers such as Guo Day Ya have been poached by big tech with high salaries. External valuations can turn stock options from “paper promises” into priced assets. Meanwhile, under a low-price strategy, API revenue can’t support a $74 billion valuation. Raising prices and securing financing at the same time is essentially buying entry tickets for the long-term table.
On the compute side, Nvidia (NVDA) 📈 and Tencent (0700.HK) 📈 directly benefit; on the application side, SenseTime (0020.HK) is under short-term pressure 📉.
Short term 📉: cash burn accelerates, valuation triples-to-septuples within three months, and skepticism about the bubble is hard to dispel. Long term 📈: if data centers + in-house chips + an open-source ecosystem can be made to run smoothly, it will become the valuation anchor for China’s AI—long in the long run, short in the short run. #DeepSeek重启融资拟募近80亿美元
$NVDA
The company’s global API usage has already hit 80 trillion tokens in a single day. The compute burned by an Agent per task is 10–100 times that of a conversational request. Add to that the near-complete rewrite of underlying code required for V4 to fully adapt to Huawei Ascend; the cost of single-round training has already exceeded $500 million. With the buildout of a GW-scale intelligent computing center, in-house inference chips, and the core team expanding by at least twofold—this system-level expansion path of “model + compute + chips” can no longer be covered by its own funds.
The deeper driving force is talent and commercialization: since the end of 2025, key researchers such as Guo Day Ya have been poached by big tech with high salaries. External valuations can turn stock options from “paper promises” into priced assets. Meanwhile, under a low-price strategy, API revenue can’t support a $74 billion valuation. Raising prices and securing financing at the same time is essentially buying entry tickets for the long-term table.
On the compute side, Nvidia (NVDA) 📈 and Tencent (0700.HK) 📈 directly benefit; on the application side, SenseTime (0020.HK) is under short-term pressure 📉.
Short term 📉: cash burn accelerates, valuation triples-to-septuples within three months, and skepticism about the bubble is hard to dispel. Long term 📈: if data centers + in-house chips + an open-source ecosystem can be made to run smoothly, it will become the valuation anchor for China’s AI—long in the long run, short in the short run. #DeepSeek重启融资拟募近80亿美元
$NVDA