- DeepSeek, which has repeatedly rejected venture capital, now plans to raise no less than $300 million in external funds.
- The sudden shift in funding strategy comes on the eve of the company's planned launch of its highly anticipated next-generation flagship V4 model later this month.
Chinese AI startup DeepSeek is engaged in a series of close commercial talks for its first injection of external capital, The Information reported on Friday, citing four people familiar with the matter.
The company is seeking to raise its first round of funding at a valuation of more than $10 billion, aiming to replenish its cash reserves to compete in the high-cost development of AI models.
The startup, which had previously rejected investment offers from top venture capitalists and tech giants on multiple occasions, now plans to raise at least $300 million in external funding, according to the report.
This rare move marks a major strategic shift by founder Liang Wenfeng, breaking the company's long-held practice of not raising external funds.
In the cash-burning race to develop large AI models, even DeepSeek, known for its low-cost algorithmic miracles, needs to replenish its financial ammunition.
The sudden shift in funding strategy comes just as the company prepares to officially release its highly anticipated next-generation flagship V4 model later this month.
The total parameter count of the V4 model is expected to jump significantly to the trillion level, featuring an ultra-long context retrieval window of up to a million tokens.
The release of the flagship model has been delayed multiple times, primarily due to an extremely complex and arduous migration of its underlying computing hardware ecosystem.
DeepSeek is attempting to completely break its long-term reliance on Nvidia chips, fully migrating its new generation of models to Huawei's latest Ascend AI chips.
Engineers have spent significant time rewriting underlying core codes to resolve the deep integration issues of the V4 model with Huawei's domestic computing hardware ecosystem.
Nvidia CEO Jensen Huang recently said bluntly in an interview that such frontier AI models optimized for Chinese hardware are bad news for the US.
To support the full-scale deployment of its new generation model, DeepSeek is no longer relying solely on software-level optimization, but is expanding into the asset-heavy infrastructure sector.
The company is planning to build a large physical data center in Ulanqab, Inner Mongolia, in northern China, to meet the massive and growing computing power demands of future models.
DeepSeek recently posted several senior on-site job openings for the data center, with maximum monthly salaries reaching 30,000 yuan ($4,400).
Ulanqab is widely considered one of the most suitable cities in China for establishing data centers, boasting abundant power supplies and optimal natural climate conditions.
The extremely low average annual temperature in the area can significantly reduce the daily cooling costs of servers, thereby effectively improving the overall power usage effectiveness of the new physical data center.
Late last month, DeepSeek suffered a severe system crash lasting nearly 12 hours, further highlighting the urgency for the company to build its own underlying physical infrastructure.
As the performance gap between top global models continues to narrow, the marginal cost for AI companies to catch up technologically is inevitably rising sharply.
If this large-scale fundraising is ultimately successful, DeepSeek will possess more computing resources and be able to offer higher salaries to prevent the loss of top AI researchers.
($1 = 6.8180 yuan)
