DeepSeek Resumes $74BN Mega-Raise - Then Warns It's Jacking-Up Prices
Chinese AI startup DeepSeek is looking to raise roughly 50 billion yuan at a valuation of 500 billion yuan (US$74 billion) - while the company warned developers a day later that it's raising prices. The resumed funding round, first reported by Chinese business outlet Caijing (via The Standard), is expected to be signed by the end of the month, with the process kept deliberately quiet; some participating institutions reportedly haven't even been formally told the deal is back on. The pricing notice, posted Thursday to DeepSeek's developer platform, warns that API rates will rise across the board and that the increase is expected to be "significant."
The round restarts barely two weeks after it was shelved, and roughly six weeks after DeepSeek's first-ever external raise in June, which pulled in some $7.4 billion from Tencent, CATL, Beijing's National AI Industry Investment Fund, NetEase, JD.com, Monolith Management and IDG Capital at a valuation north of $50 billion (corporate filings later implied a figure closer to $52 billion). The new target is a steep markup on a deal that closed weeks ago, and the prospective investor list reportedly blends returning backers with new money that couldn't get into round one.
As we noted last month, the round stalled because a transcript of founder Liang Wenfeng's May 20 closed-door meeting with investors leaked onto GitHub and went viral - a document in which Liang admitted China still trails the US in AI (a gap of compute and capital, not talent, per Liang), conceded DeepSeek remains heavily dependent on Nvidia hardware, described Huawei's substitutes as falling well short, and acknowledged the company could get its hands on processors he called "noncompliant."
Liang - reportedly furious over the leak - simply froze the deal.
Meanwhile, questions about how DeepSeek is obtaining restricted silicon go back to the very beginning - with federal investigators probing whether Singapore-based third parties funneled blacklisted Nvidia chips to the lab as far back as early 2025, and by that April the House Select Committee on the CCP had formally concluded that DeepSeek's model appeared to run on tens of thousands of export-restricted Nvidia chips, branding the company a tool for "subverting" US export controls. Congress alleged it and now the founder has all but confirmed it, on (leaked) record telling investors the workaround exists - even as the company lays groundwork for a Shanghai STAR Market IPO it reportedly hopes to file this year. Should be a fun prospectus to read.
Liang was able to hit the pause button so easily because DeepSeek's first round routed nearly every investor's money into a limited partnership Liang personally controls, complete with a five-year lock-up and zero voting rights. The lone exception was the state AI fund, which bought direct equity and kept its vote. Liang holds roughly 84% of the company and near-total voting control - meaning Tencent and CATL wired in billions for the privilege of watching, while Beijing holds the only outside seat at the table.
In the leaked transcript, he told investors that GPUs are the single best use of the company's capital, that he would happily convert the entire war chest into Nvidia hardware if supply allowed, and that spending 20 billion yuan a year on processors would count as a triumph for his purchasing team. He also told them the money doesn't go as far as it looks: even 50 billion yuan, he said, wouldn't train a model at the scale of the largest U.S. systems, which he estimated would take roughly 50,000 Nvidia GB300s or 200,000 Huawei Ascend 950s - before the experiments that precede any final training run. Beyond the hardware, the company plans to double headcount and push into data centers and AI agents, and Reuters reports it is quietly hiring chip designers to develop in-house inference silicon - an attempt to engineer its way out of the very dependence Liang lamented.
Price Hike
Which brings us to the other shoe - the company's announcement that API rates will rise across the board in a "significant" way - the company's second hike in under a month, following a "peak/off-peak" scheme TechNode reported in mid-July, which doubled rates during Chinese business hours.
No new rate card or effective date has been published - but this is the lab whose $0.14-per-million-token pricing dragged ByteDance, Tencent, and lately even the US majors into a price war. Demand has reportedly overwhelmed its infrastructure, and bargain-basement tokens are a hard pitch to pre-IPO investors. The catch: DeepSeek's weights are open, so any datacenter on earth can serve the same models. If they raise prices too far, customers may balk.