Automated translation could eliminate NVIDIA's software moat within a year.
TileLang's explicit tiling and backend compilation let developers write once and hit Hopper, MI300X, or Ascend targets, with public examples already matching FlashMLA throughput on H100 in under 100 lines of Python.
Public repos and model releases confirm the technical path exists, but nothing establishes timelines, chip procurement volumes, or whether automated translation will play any role in broader adoption.
Many users stay bullish on NVIDIA because the CUDA gap with DeepSeek and Huawei still perpetuates its lead while others flag Huawei hardware aging and power bottlenecks.
No Digg Deeper questions have been answered for this story yet.
I have said this before, but distillation is the least of the US's worries. Gpt 5.6 sol, a model whose tokens are basically free, can entirely rewrite all CUDA programs to whatever the fuck you want them to be written to
On domestic compute: bullish within a year. «Domestic AI chips have no problems in hardware or ecosystem—the only problem is insufficient production capacity» «Previously, domestic GPU adaptation had a problem called poor ecosystem… The moat of NVIDIA's CUDA is being rapidly dismantled, for probably three reasons. - with AI, building this ecosystem is much easier than before, because AI can write code. - second, some new technologies. For example, our company produced a technology called TileLang—a high-level language. Using this high-level language to write CUDA operators, you can quickly rewrite NVIDIA's entire ecosystem, and combined with AI, there seem to be no obstacles. - Another point: because CUDA—NVIDIA evolved from gaming GPUs, so in many places the gaming GPU design and settings carried through»
Wenfeng: - DeepSeek is working closely with Huawei and believes it can secure approximately 16,000 Huawei AI chips. - AI-powered code generation and advanced programming languages such as TileLang are expected to rapidly lower the barriers to entry created by the CUDA ecosystem. - Although NVIDIA GPUs were used to train V3, DeepSeek has significantly reduced its software dependence on NVIDIA’s ecosystem by using its own compiler and a TileLang-based environment. - If TileLang and DeepSeek’s proprietary compiler are ported to Huawei chips, the ecosystem issues surrounding Chinese chips could largely be resolved within about a year. The main remaining bottleneck would be production capacity. - Huawei’s 950 SuperNode could replace workloads currently handled by the GB200 and GB300, although roughly four Huawei cards are needed to match the performance of one NVIDIA card, and Huawei’s products remain about two years behind. - The chip gap between China and the United States is best understood as roughly a 4x gap in hardware efficiency and a two-year time lag—not primarily as an ecosystem gap. The end of CUDA’s moat is approaching. The ecosystem problem for Chinese chips could be solved within a year. The only real bottleneck left is production itself. Very bearish on $NVDA.
Wenfeng articulated the Nvidia bear case quite well in the leaked investor call
@jukan05 Not so bearish tbh His original words are “willing to buy as much as possible and as soon as possible from nvda” and the model scale is mostly limited by computer resource But of course they are working on tilelang and it might work in one year
It's so sad the way the US decided to force China to fund direct competition to our most valuable companies over some scary stories, while being largely ineffective at the stated goal of slowing their AI progress. And it was all foreseeable.
Wenfeng: - DeepSeek is working closely with Huawei and believes it can secure approximately 16,000 Huawei AI chips. - AI-powered code generation and advanced programming languages such as TileLang are expected to rapidly lower the barriers to entry created by the CUDA ecosystem. - Although NVIDIA GPUs were used to train V3, DeepSeek has significantly reduced its software dependence on NVIDIA’s ecosystem by using its own compiler and a TileLang-based environment. - If TileLang and DeepSeek’s proprietary compiler are ported to Huawei chips, the ecosystem issues surrounding Chinese chips could largely be resolved within about a year. The main remaining bottleneck would be production capacity. - Huawei’s 950 SuperNode could replace workloads currently handled by the GB200 and GB300, although roughly four Huawei cards are needed to match the performance of one NVIDIA card, and Huawei’s products remain about two years behind. - The chip gap between China and the United States is best understood as roughly a 4x gap in hardware efficiency and a two-year time lag—not primarily as an ecosystem gap. The end of CUDA’s moat is approaching. The ecosystem problem for Chinese chips could be solved within a year. The only real bottleneck left is production itself. Very bearish on $NVDA.
@jukan05 The CUDA moat got obliterated by the AI that it helped make. I've ported 2 CUDA projects to MPS in a couple hours, not massive projects by any means. But we're talking weeks of work in a couple hours.
I have said this before, but distillation is the roast of the US's worries. Gpt 5.6 sol, a model whose tokens are basically free, can entirely rewrite all CUDA programs to whatever the fuck you want them to be written to
On domestic compute: bullish within a year. «Domestic AI chips have no problems in hardware or ecosystem—the only problem is insufficient production capacity» «Previously, domestic GPU adaptation had a problem called poor ecosystem… The moat of NVIDIA's CUDA is being rapidly dismantled, for probably three reasons. - with AI, building this ecosystem is much easier than before, because AI can write code. - second, some new technologies. For example, our company produced a technology called TileLang—a high-level language. Using this high-level language to write CUDA operators, you can quickly rewrite NVIDIA's entire ecosystem, and combined with AI, there seem to be no obstacles. - Another point: because CUDA—NVIDIA evolved from gaming GPUs, so in many places the gaming GPU design and settings carried through»
@0xWendy99 @jukan05 他说的是尽可能多的买算力卡，但没说必须从英伟达买，而且你想买也没不到或者说价格太贵了不划算也不可能乱买
@0xWendy99 @jukan05 而且英伟达的卡还有一个问题，它的卡是针对专有格式nvfp8进行优化的，而DeepSeek和kimi之类的国模现在都是用的mxfp8格式
Moat of software is not a moat anymore. When we truly step into the era of AI, the largest and only moat is manufacturing of hardware. Even you can build robots to manufacture anything you need, you still need to build millions of robots first, and this is still a big moat. Today, there is no one busy in manufacturing robots except @elonmusk .
@jukan05 You sure it's not 16K racks? And even then, that doesn't sound like a lot.
@jukan05 @zephyr_z9 Bearish until you realize AI is a worldwide phenomena.
@jukan05 Wenfeng thinks 16k chips is security. It’s a lease with an expiration date. HuaWei hardware ages like milk in a datacenter. You’ll be paying rent on yesterday’s compute by Q3.
@jukan05 Not bearish, gap is gap. 2 years gap would be substantial in case of RSI. I would argue, it is enough to be THE gap to perpetuate the lead of US AI models
@jukan05 CUDA 护城河这下有点压力了
@jukan05 no imminent impact. but, yes, moat is gone. AI alao good at kernel coding.
@jukan05 CUDA moat is a under threat from various angles! $QCOM already announced plans, now China is doing their own thing. Ecosystem advantage is being chipped away
@jukan05 @zephyr_z9 Very bullish NV
@jukan05 Is it bearish for NVIDIA to lose a market where they were already prohibited from selling? Are you still repeating the same FUD nonsense every day?