Leaked transcript: DeepSeek's Liang Wenfeng says compute gap with US, not talent, is the real problem; fundraise paused
- A leaked transcript of an investor meeting has DeepSeek founder Liang Wenfeng framing China's AI disadvantage as a resource problem, not a talent one, saying "the biggest gap between us and the US is in resources."
- DeepSeek has told prospective investors it is suspending its second fundraising round, according to people familiar with the matter cited by Bloomberg, days after the remarks circulated online.
- Liang reportedly said he needed 200,000 Huawei 950 chips to train a frontier model but received only 16,000, and expects the chip crunch to last at least three years.
- Liang is said to have floated narrowing the gap with US labs to three to six months using a fraction of the compute those labs use.
- A 118-item version of the transcript published by Tencent's tech outlet reportedly covers AGI strategy, chip supply, pricing, and staff retention.
Hacker News 의견들
This bodes well for smaller open models. There's a delusion that what US AI companies are doing is 'best'; the frontier is bloated and there are hundreds of viable speedups. Open weights will keep grinding down that bloat.
People are misreading the headline. It's not that leaked comments caused the pause, it's that DeepSeek is pausing fundraising because they perceive a compute gap with the US, and the leak just revealed that plan.
China could close the compute gap by renting hyperscaler compute through shell entities, like how the US procured titanium from the USSR for the SR-71 during the Cold War.
Actually it's the reverse now. Trump reversed course on the Nvidia ban, and it's China that's blocking its own companies from buying Nvidia chips, so any shell game would be to dodge Chinese restrictions, not US ones.
Everything in this transcript reads completely different from what the people running Anthropic or OpenAI say publicly.
Liang basically says the resource gap, not the talent gap, is the real problem: they have 'virtually no difference' in people versus the US but can't afford to train the largest models. Trump's chip bans pushed China's domestic semiconductor progress, similar to what happened with EVs and solar.
If Chinese open models are catching up at a fraction of the cost and AI gets commodified anyway, why is DeepSeek still chasing the frontier when the lead looks temporary and expensive to hold?
They want AGI first because once someone gets there, nobody knows what the world looks like after, so missing that jackpot could end your company even if the payoff is uncertain.
AI's already commoditized. The winner-take-all fundraising story only makes sense if AGI is some real technological step change, otherwise this ends up looking like a steel mill business, not software, and countries treat steel as a national security issue for a reason.
DeepSeek is funded by their own hedge fund, High-Flyer, and they cap token prices basically at cost to recover server expenses. The transcript frames this as a moral stance: they don't chase image/video gen hype and only care about reasoning, chain of thought, and continuous learning.
You have to be careful criticizing China here since criticism of specifics gets read as criticism of the Party's policies.
Funny thing is the current US administration works the same way. Anthropic didn't cooperate with the military and its new model got 'paused' weeks later, with officials openly hinting beforehand that non-cooperation would hurt business. Not defending China, just noting the same ugly pattern shows up here.