DeepSeek's Playbook Leaked, Down to the GPU Math. Days Later, a $1.5 Billion Round at $71 Billion Went on Ice.

TL;DR

On May 20, DeepSeek founder Liang Wenfeng spent three hours and forty-four minutes walking the lab's first outside investors through the playbook: the compute math, a claim that four Huawei cards do the work of one Nvidia card, a homegrown compiler stack he says has nearly written CUDA out of the story, and a pledge that DeepSeek's strongest models will keep shipping as open weights. In late July the talk leaked, and a lightly edited transcript went viral in translation. On July 25, Bloomberg reported the aftermath: DeepSeek verbally told prospective backers it is suspending its second funding round, a deal targeting at least 10 billion yuan (about $1.5 billion) at a valuation near $71 billion. That makes it a strong candidate for most expensive audio file of 2026.

A four-hour debrief, on the record forever

The meeting was Liang's debrief for the incoming shareholders of DeepSeek's first external financing ever: roughly 50 billion yuan, about $7 billion, closed in June 2026 at a valuation of roughly $52 billion. Bloomberg calls it one of the largest startup financings China has seen. Backers reportedly include Tencent and battery giant CATL, and Bloomberg's sources say Liang put around $3 billion of his own money in alongside them.

Then the room leaked. A recording circulated through Chinese tech and investor circles, Tencent's tech desk published a lightly edited transcript in late July, and English translations from AI Proem, Recode China AI, and HelloChinaTech landed within a day. The original WeChat link has since been taken down. DeepSeek has not confirmed the transcript is genuine, and Bloomberg says it has not verified the viral posts either. Hold that caveat: everything in the next three sections is quoted from a document the company will not vouch for.

The compute math Liang laid out

The through-line of the whole talk is that compute is the only gap that matters. In the transcript Liang attributes essentially every US-China difference, including the apparent talent gap, to differences in compute resources, which matches Bloomberg's summary of the remarks that went viral. The most repeated framing has DeepSeek producing comparable models 12 to 18 months after the US frontier on roughly one-twentieth of the compute, though his estimate of the lag wanders between about six months and two years depending on which question he is answering.

He is also blunt that the constraint is not money. In the transcript he says spending the lab's multibillion-yuan GPU budget in full this year would count as an exceptional performance by the procurement team. The bottleneck is getting cards at all, not paying for them.

Four Huawei cards, one Nvidia card, and a compiler

The line already being quoted at Nvidia's expense: "Four Huawei cards equal one Nvidia card," which the AI Proem translation extends with "and a two-year lag." No workload, precision, or interconnect caveats attached, so treat it as a founder's round number rather than a benchmark.

The more consequential claim for builders is the software one. Liang says DeepSeek wrote a high-level kernel compiler, TileLang, built its stack on top, and now "barely depends on Nvidia's ecosystem," with CUDA's moat eroding fast. CUDA's lock-in works like a city where every appliance is hardwired into one power grid; a compiler layer like TileLang is the universal adapter, so you write the kernel once and plug it into Nvidia silicon today and Huawei silicon tomorrow, and the grid quietly loses its leverage. That is precisely the future Nvidia's valuation assumes will not arrive, described matter-of-factly by the lab with the best claim to be living in it.

Open weights as identity, not a funnel

Our strongest model will probably be open-sourced too. I can't see what good closed-source does.

For anyone deploying open models, that is the most load-bearing sentence in the transcript. It comes with two supporting claims: the open weights are identical to what DeepSeek deploys internally, no dumbed-down community edition, and the business math works anyway because inference hardware pays for itself in about ten months at DeepSeek's prices. His pricing philosophy is a proverb: "Those who take more will be beaten by those who take less." On AGI he is unhurried, arguing that a model given a clearly described problem with complete context already surpasses humans, and that the lab's single non-negotiable interest is keeping the team stable.

Then the bill arrived

On July 25 Bloomberg reported that DeepSeek verbally told prospective investors it will not sign the agreements for its second round: at least 10 billion yuan at a pre-money valuation around 480 billion yuan, roughly $71 billion, a step-up of about a third over the June round. The trigger, per Bloomberg's sources, was Liang's frustration at seeing his closed-door remarks circulate online. The lab that invented rush-hour token pricing has discovered the one thing it cannot meter: a live microphone.

Why this matters if you build on open weights

DeepSeek is the lab whose releases have twice reset what "free" means in this market, and this is the first detailed look at the strategy underneath, however involuntarily published. If the transcript is genuine, the open-weight pipeline you may be building on is not a loss leader awaiting an enterprise pivot; it is the whole thesis, priced to a ten-month hardware payback and philosophically committed at the founder level. That lands in a crowded week for the thesis: Kimi K3's full weights publish on July 27, and DeepSeek's answer to it now has a paper trail.

The countervailing risk is chilling. The most transparent look inside a Chinese frontier lab ever published just cost that lab a $1.5 billion round, at least temporarily. Every founder in Beijing and Hangzhou watched that happen. The next four-hour strategy debrief will be shorter, vaguer, and swept for phones.

The caveats, straight

• DeepSeek has not confirmed the transcript's authenticity, and Bloomberg explicitly says it has not verified the viral posts. Every quote above inherits that asterisk.
• The English versions are translations of a lightly edited Chinese transcript of a recording. Specific phrasings, especially the Huawei ratio and the gap estimates, have visibly drifted between translations.
• The pause was communicated verbally, not in writing. A verbal pause can un-pause just as quietly, and none of the reporting says the round is dead.
• The valuation and round figures come from Bloomberg's unnamed sources; DeepSeek has published none of them.

Key Takeaways

• A recording of DeepSeek founder Liang Wenfeng's May 20 investor debrief leaked; Tencent's tech desk published a lightly edited transcript in late July and English translations spread within a day. DeepSeek has not confirmed it is genuine.
• On July 25, Bloomberg reported DeepSeek verbally suspended its second funding round: at least 10 billion yuan (about $1.5 billion) at a valuation near $71 billion, after closing about $7 billion at roughly $52 billion in June.
• The transcript's core claims: the US-China gap is compute, not talent; DeepSeek ships comparable models 12 to 18 months later on roughly one-twentieth the compute; and four Huawei cards currently equal one Nvidia card.
• Liang says DeepSeek's TileLang-based compiler stack has left it barely dependent on Nvidia's ecosystem, and calls CUDA's moat fast-eroding.
• The strongest signal for builders: DeepSeek's best models will keep shipping as open weights, identical to internal deployments, backed by an inference business Liang says pays back hardware in about ten months.
• The meta-lesson cuts the other way: candor leaked, and it froze $1.5 billion. Expect Chinese labs to get quieter, not louder.

Sources: Bloomberg, Fortune, Unite.AI, AI Weekly, AI Proem (translation), Recode China AI (translation), HelloChinaTech (translation), transcript PDF mirror on GitHub