--- title: "Liang Wenfeng Says CUDA's Moat Falls in a Year. The Wall He Can't Climb Takes Three." pubDatetime: 2026-07-24T20:13:00.000Z description: "A leaked four-hour call with DeepSeek's founder got read as a bear case for NVIDIA. The software claim in it already happened, in public, on GitHub. The capacity number is the one that settles the export-control argument." tags: [ai, china, deepseek, export-controls, nvidia, policy, open-weight, 2026, 2026-q3, 2026-07] --- > [!tldr] TL;DR > A four-hour call between DeepSeek founder Liang Wenfeng and his investors, recorded 20 May, was published yesterday and immediately read as a bear case for NVIDIA. Liang did say AI code generation and TileLang are collapsing the barrier CUDA built, and that porting his compiler stack to Huawei silicon would close the ecosystem gap in about a year. He also said he needs 200,000 Huawei 950 chips to train a frontier model and can get 16,000, that Huawei needs roughly four cards to match one NVIDIA card and runs about two years behind, and that the capacity shortfall is "basically unsolvable" for at least three years. The software claim is verifiable and mostly already happened in public on GitHub. The capacity number is the one that decides the export-control argument, and it points the other way. The post that made this leak travel was [Jukan's](https://x.com/jukan05/status/2080205767861510343), which summarised Liang's chip remarks in six bullets and signed off: "The end of CUDA's moat is approaching… Very bearish on $NVDA." It has 1,255 likes. The top reply, from [@0xWendy99](https://x.com/0xWendy99/status/2080209109182546225), points out that Liang's actual position is he is "willing to buy as much as possible and as soon as possible from nvda." Both things are in the transcript. Only one of them is a trading thesis. ## What leaked, and how hard to lean on it Fred Gao published the [transcript](https://www.fredgao.com/p/deepseeks-liang-wenfeng-breaks-his) on his Inside China newsletter yesterday. The source is an audio file, `deepseek 0520.m4a`, running 3 hours 44 minutes, from a 20 May investor meeting. Mirrors went up fast: an [abridged 52-point English version](https://elsewhere.news/en/elsewhere/wenfeng-liangs-four-hour-investor-meeting-full-transcript) at Elsewhere News, and a [PDF on GitHub](https://github.com/demo-zexuan/liang-wenfeng-investor-meeting-2026-7-22). Gao's transcript carries its own health warning at the top: auto-transcribed by speech recognition, organised by AI, and "individual proper nouns and numbers may contain recognition errors, please refer to the original recording as authoritative." So an entire week of chip-policy commentary is now resting on figures that came out of a machine transcription of a machine translation of a leaked private call, with a disclaimer attached specifically to the numbers. The 200,000, the 16,000, the 4x, the two years: every one of those is being quoted to three significant figures by people who have not heard the audio. Hold them loosely. ## The CUDA claim already happened Here the leak is checkable, and it checks out better than the framing suggests. [TileLang](https://github.com/tile-ai/tilelang) is a real, mature open-source project: a Pythonic DSL for writing high-performance GPU kernels, built on Apache TVM, sitting at 6,878 GitHub stars. It was open-sourced on 20 January 2025, eighteen months ago. Its acknowledgments credit Peking University's Prof. Zhi Yang, with part of the work done during an internship at **Microsoft Research**. The tool Liang expects to dissolve NVIDIA's ecosystem advantage was partly raised inside Microsoft. The Huawei port is already underway too. TileLang [added AscendC and Ascend NPU IR backends on 29 September 2025](https://github.com/tile-ai/tilelang-ascend), ten months ago. And DeepSeek made its own bet public before this call leaked: [deepseek-ai/TileKernels](https://github.com/deepseek-ai/TileKernels), "a kernel library written in tilelang", landed on 22 April 2026 and has 1,661 stars. When Liang says the ecosystem problem could be resolved "within about a year", he's quoting the remaining distance on work that started in 2025. One detail cuts against the anti-NVIDIA reading. In December 2025 TileLang added a [CuTeDSL backend](https://github.com/tile-ai/tilelang/pull/1421) that compiles down to NVIDIA CUTLASS. It also supports AMD MI300X and Apple Metal. TileLang is a portability layer, and portability layers get used most by whoever writes the most kernels, which today is still people with NVIDIA hardware. ## 16,000 against 200,000 The part of the call that actually constrains DeepSeek gets far less airtime. Liang said he needs roughly 200,000 Huawei 950 chips to train a frontier model and that Huawei can supply about 16,000. Huawei's [expected output for this year](https://www.reuters.com/world/china/huaweis-new-ai-chip-find-favour-with-bytedance-alibaba-which-plan-place-orders-2026-03-27/) is around 750,000 chips, split across every Chinese AI company. At Liang's own arithmetic, China's entire 2026 domestic supply covers under four frontier training runs, for everyone. Huawei's [own roadmap](https://www.huawei.com/en/news/2025/9/hc-xu-keynote-speech) makes the squeeze worse than the headline. The 950 line splits in two: the 950PR for prefill and recommendation, and the 950DT for decode and training. The 950DT, the one you need for a training run, is scheduled for Q4 2026. The chip Liang wants 200,000 of is still, in the part that matters to him, a launch. Then the efficiency gap. Four Huawei cards to match one NVIDIA card is also a 4x power bill for the same compute. As [one reply](https://x.com/MTorygreen/status/2080250560968614144) put it: "Software moats fall in a year. Power takes a decade." China has more grid slack than the US to absorb that, which is one of the few places this story runs in Beijing's favour. Liang's summary was flat: "Huawei's problem is still insufficient capacity… This problem is currently basically unsolvable." He expects it to hold for at least three years. ## Everyone is quoting the same man Transformer's Shakeel Hashim [read the leak](https://www.transformernews.ai/p/deepseek-ceo-liang-wenfeng-export-controls-china) as the strongest available argument *for* export controls: the CEO of China's best-regarded lab says compute is the only thing holding him back, so restrict compute. The counter-argument, that controls accelerate indigenization, points at TileLang and the Ascend port as exhibit A. Liang supplies evidence for both and a timeline that arbitrates between them. One year to fix the software. Three years, unsolvable, on the silicon. If you think transformative AI is close, that gap is the whole policy question, and it lands roughly where the [open-weight fight in Washington](/posts/openai-anthropic-china-open-weight-alliance) has been heading all month. For the market read, some scale: NVIDIA guided to $91 billion in Q2 revenue for its 26 August report, with China data-centre compute already excluded from the outlook. A cost curve that bends in 2029 is not what moves that number. There's one more thread worth pulling. Liang described his API pricing as recovering hardware cost in ten months, roughly sixfold profit, and said flatly that Alibaba and Tencent can't reach his cost structure. That doctrine has a downstream effect nobody on the call mentioned: AISI recently [measured](/posts/aisi-open-weight-cyber-gap) DeepSeek V4-Pro at $0.28 per reliably solved offensive cyber task against Opus 4.5's $12.50. Restraint as a pricing philosophy is also, from a defender's chair, a 45x discount on attack attempts. The transcript went up on Hacker News seven separate times in 24 hours, across five different mirrors. Top score: [14 points](https://news.ycombinator.com/item?id=49017758), zero comments. The most detailed account we have of how China's most influential AI founder thinks about compute, and the internet scrolled past it to argue about the tweet. --- **Sources** - Fred Gao, [DeepSeek's Liang Wenfeng Breaks His Silence](https://www.fredgao.com/p/deepseeks-liang-wenfeng-breaks-his), Inside China (23 July 2026) - [Wenfeng Liang: Four-Hour Investor Meeting Transcript](https://elsewhere.news/en/elsewhere/wenfeng-liangs-four-hour-investor-meeting-full-transcript), Elsewhere News (abridged English version) - Jukan (Citrini Research), [thread on the leaked chip remarks](https://x.com/jukan05/status/2080205767861510343) (23 July 2026) - Shakeel Hashim, [DeepSeek's boss just made the case for export controls](https://www.transformernews.ai/p/deepseek-ceo-liang-wenfeng-export-controls-china), Transformer (24 July 2026) - David Moadel, [A Chinese CEO Just Outlined the Bear Case for NVIDIA](https://247wallst.com/investing/2026/07/23/a-chinese-ceo-just-outlined-the-bear-case-for-nvidia-it-should-terrify-owners-of-the-stock/), 24/7 Wall St (23 July 2026) - [tile-ai/tilelang](https://github.com/tile-ai/tilelang) and [deepseek-ai/TileKernels](https://github.com/deepseek-ai/TileKernels) on GitHub - Eric Xu, [Leading a New Paradigm for AI Infrastructure](https://www.huawei.com/en/news/2025/9/hc-xu-keynote-speech), Huawei Connect keynote (September 2025) **Related on this blog** - [The White House Says Kimi K3 Is Distilled Fable](/posts/moonshot-fable-distillation-kimi-k3) - [OpenAI and Anthropic Agree on Exactly One Thing](/posts/openai-anthropic-china-open-weight-alliance) - [The Open-Weight Cyber Gap Is Four Months. The Price Gap Is 45x.](/posts/aisi-open-weight-cyber-gap)