DeepSeek is pausing a $1.5B funding round after its founder's leaked remarks on the US compute gap

DeepSeek, one of China's leading AI labs, has told the investors in its second funding round, worth about $1.5 billion, to pause the signing process after a 4-hour recording in which founder Liang Wenfeng appeared to concede that China is roughly two years behind the United States on training compute. The episode turns a private technical admission, surfaced on Chinese social media, into a public fundraising freeze without any change in the underlying chip supply.

The trigger is a transcript of the 22 July investor meeting that circulated on Chinese platforms last week and was mirrored on GitHub. DeepSeek has not confirmed the record, and a fact-checked English translation by Recode China AI and a thematic breakdown by Hello China Tech are the working surfaces for non-Chinese readers. The substance is consistent across the two: Liang argued, in remarks that Hello China Tech flags as conversational rather than precise, that "all the differences we see, including talent, model capability, and applications, can be attributed to differences in compute resources." The cleanest line is his summary: "to put it simply, two years behind, using one-twentieth of the compute."

Compute, in this context, is the dollar-and-chip budget for training large language models, the same budget that has put U.S. labs like OpenAI and Anthropic on a multi-billion-dollar annual spending curve. By that measure, "two years behind, one-twentieth of the compute" is a public admission that the China-U.S. AI race is now defined less by who has the better model and more by who can still buy the silicon to train one. Liang's framing was blunt: "Talent is not the bottleneck. Resources are the biggest bottleneck. The talent gap is fundamentally a compute gap."

The pause is provisional. Bloomberg, reported by Fortune, says DeepSeek has told prospective investors in the second round that it is suspending the signing process, and that "negotiations remain fluid" because the company may still choose to proceed. The timing is what matters: the suspension came in the days after Liang's investor-meeting remarks went viral on Chinese platforms, and Bloomberg attributes it in part to Liang's own frustration over the leaks. A founder who can describe his own lab's hardware disadvantage on the record is, in effect, telling existing and prospective investors that the U.S. chip lead is now a fundraising risk factor.

Liang spent part of the meeting on hardware workarounds. He claimed that Huawei's 950 super-node, Huawei's domestic equivalent of an Nvidia rack-scale training system, "can fully replace Nvidia's GB200 and GB300 in performance and price," and that four Huawei chips match one Nvidia chip on the relevant workloads. Neither claim cites a specific chip model, benchmark, or workload. The Recode China AI translation preserves the same ratio without a workload anchor, and the Hacker News parsing thread is already treating the "4-to-1" line as a public-relations claim rather than a benchmark result. Liang also said DeepSeek's V3 was trained on Nvidia chips but built on a homegrown high-level compiler called TileLang rather than Nvidia's broader software stack, a useful reminder that China's compute gap is about both the silicon and the developer ecosystem that runs on top of it.

DeepSeek's first external round closed in June at roughly $50 billion pre-money, Fortune reports, raising $7 billion from Tencent, battery maker CATL, and a state-backed national AI fund. The second round had been targeting at least 10 billion yuan, roughly $1.4 billion, at a pre-money valuation that Fortune and the Financial Times put at $71 billion to $74 billion ahead of a planned 2027 IPO. A pause on those terms does not change what DeepSeek has already raised, but it does change the price discovery for the next dollar in.

For the rest of China's AI labs, the Liang transcript sets a new price on a private admission. A founder can now concede a hardware gap, the recording can leak, and a funding round can stall, all before the next chip export rule or model release. The next test is whether the second round closes at the original $71B to $74B band, drops, or walks away, and whether the same calculation shows up at the cap tables of DeepSeek's peers.