DeepSeek Wants Fresh Billions Weeks After Its First Round — Cheap AI Is Expensive

DeepSeek, the Chinese AI lab famous for underpricing everyone, is already back asking for money. Just weeks after closing its first-ever funding round — roughly $7 billion at a $52 billion valuation in late May — the company is in early talks with investors for a new round at a pre-money valuation of about $71 billion, the Financial Times reports.

The fresh capital has a clear destination: DeepSeek wants to build its own data centers and buy AI chips, and it’s designing its own inference chip — the hardware that runs a model once it’s trained — to reduce dependence on Nvidia and Huawei. Founder Liang Wenfeng put in about $3 billion of his own money in the first round, making him the largest backer, alongside investors like Tencent, CATL, and JD.com. The context: DeepSeek’s new V4 models are the largest open-weights models around at up to 1.6 trillion parameters, and their prices are brutal — roughly eleven times cheaper than GPT-5.5 on input, a discount the company recently made permanent. It’s working, too. Payment data from US financial firm Ramp showed DeepSeek among the fastest-growing software vendors at American companies in June, though Ramp also flagged the security question of sending company data through DeepSeek’s platform.

What’s behind this? Selling AI far below the competition’s price only works if you keep driving your own costs down — and that takes infrastructure you own rather than rent. Hence the raise. The performance picture is honest but nuanced: OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos-tier models have pulled ahead into a league DeepSeek can’t match yet, but the gap in capability is far smaller than the gap in price, and for many everyday tasks a good-enough model at a fraction of the cost wins. Meanwhile DeepSeek’s domestic rivals — Zhipu, MiniMax, Moonshot — are all raising and shipping aggressively, so standing still isn’t an option.

What this means for you: If you just chat with an AI app, little changes directly. But this price war is quietly setting the cost of everything built on AI — the tools at your workplace, the features in your apps. When capable open-weights models cost pennies, AI features stop being premium add-ons. The caveat to keep in mind: where a model runs matters. Companies (and cautious individuals) should weigh that data flowing through any provider’s platform is subject to that provider’s jurisdiction — one reason some businesses run open-weights models on their own hardware instead.

Sources

• Financial Times
• The Decoder