In a leaked nearly four-hour investor talk, DeepSeek's Liang Wenfeng says the main US-China gap is compute power, Nvidia's CUDA moat is disintegrating, and more
In a rare four-hour talk, the reclusive founder reveals an almost Daoist philosophy of AI—AGI as a tide no company can own …
Inside ChinaFred Gao
Context & Ripple Effects
DeepSeek had already emerged from a High-Flyer research branch as a credible challenger to U.S. AI leaders, while its research-first posture and willingness to share work were portrayed as competitive advantages. DeepSeek’s rise from a quant-fund research unit and its research-led, more open posture make Liang’s remarks consequential beyond a single company.
The company’s founder had previously identified U.S. chip restrictions as a bottleneck. This talk refines that earlier warning about export-control constraints into a broader thesis: access to compute, rather than ownership of AGI, is the decisive strategic divide.
First-order effects
DeepSeek publicly anchors its competitive narrative in constrained compute access, putting hardware availability and efficiency at the center of how its progress will be assessed.
Liang’s assertion that CUDA’s moat is weakening directly challenges Nvidia’s software-lock-in narrative, though the remarks alone do not establish a change in Nvidia’s market position.
Second-order effects
Chinese AI developers and infrastructure providers have greater incentive to prioritize software portability, alternative accelerators, and efficiency work that reduces dependence on the Nvidia stack.
Nvidia and its ecosystem face added pressure to demonstrate that CUDA’s developer and deployment advantages remain durable, especially as DeepSeek’s work has already prompted debate over how openly AI advances should be shared. the debate over publishing AI breakthroughs
Third-order effects
If compute access remains the binding constraint, AI competition may increasingly turn on national hardware supply, allocation, and financing rather than solely on model research quality.
If credible alternatives to CUDA gain practical adoption, AI infrastructure could shift from a single dominant software ecosystem toward a more fragmented, hardware-specific deployment landscape; that outcome remains contingent on real-world developer uptake.
The trend: This is one data point in the shift from frontier-model competition toward compute access and hardware-software independence as strategic AI advantages.
“My preferred narrative is “a group of ordinary people doing extraordinary things,” rather than “a group of geniuses doing extraordinary things."" - Liang Wenfeng, CEO of DeepSeek, in his 4hr Investor meeting [image]
Some other observations: AGI as the main line: they ignore areas like video generation and world models, and focus on language models. China's role in AI: Same in the manufacturing sector, Chinese AI will make services cheaper, leveraging structural advantages the US lacks.
I've made a transcript of DeepSeek's nearly 4-hour meeting with investors by Liang Wenfeng. Liang is closer to Daoist thinking. He keeps saying “restraint” in the meeting, setting limits on his own gain. A very different logic from his US counterparts.🧵 https://www.fredgao.com/..…
52 Quotes of Wenfeng. Would be a good set of fortune cookies or cards, one for every week. He agrees with me, or I with him. In short: gradual singularity near-term, still little interest in business, and, importantly - *commitment to always open source their best model.* [image]