NVIDIA 必须比前沿实验室更快地变成芯片公司——一场围绕定制 ASIC 的攻防
这篇讲透了前沿实验室为什么都在自研芯片:模型和芯片要一起迭代才有效果,买同款 GPU 拿不到任何专属优势。看完你会明白 NVIDIA 收购 Groq、Hugging Face 这些动作背后的焦虑,做基础设施或算力决策的人都该看看。
一篇关于 NVIDIA 未来命运的分析指出,其最大威胁来自自家客户:Google 已有 TPU,OpenAI 和 Meta 正与 Broadcom 合作开发定制 ASIC,Anthropic 组建了自研芯片团队,SpaceX/xAI/Tesla 则在推进 Terafab 项目。在递归自我改进(RSI)的框架下,每家前沿实验室和超大规模云厂商都必须成为芯片公司,否则就会输掉竞赛。
论证的核心机制是:纯软件层面的自我改进存在物理极限,软件无法改变芯片的带宽、内存容量和互联。要真正获得 RSI 的复利效应,必须让模型与芯片协同设计,形成双向反馈——更聪明的模型设计出更好的定制芯片,这些芯片再训练下一版模型。购买与所有人相同的 NVIDIA GPU 无法带来专有优势,而向 NVIDIA 提交下一代架构需求则意味着提前数年共享路线图。同时,AI 已能自动编写编译器、把 CUDA 内核移植到其他芯片,这直接侵蚀了 CUDA 护城河。
Jensen Huang 曾反驳称模型迭代太快,固定硬件的定制 ASIC 不比 GPU 更划算,且 Broadcom 等合作伙伴抽成不低。但文章认为,当 AI 参与协同设计时,硬件不会落后于模型变化;加上 RSI 需要 7×24 小时的内部推理,定制 ASIC 在成本和能效上远优于通用 GPU(NVIDIA 毛利率超 75%)。NVIDIA 的应对是收购 Groq(快速推理 IP)、Poolside + Nemotron(内部编码模型)和 Hugging Face(分发渠道),并押注开源权重模型来留住企业客户——否则企业将直接调用运行在实验室自有芯片上的 API。
“NVIDIA has to become a frontier lab faster than frontier labs are becoming chip companies.” Mo @PersianMacroGuy Ok I do have a view on NVIDIA without Jensen. I’ll tell you the real reason $NVDA will likely fail without him and why the frontier labs and hyperscalers are the biggest winners. (Elon has been in a really good mood recently. Have you noticed?) The biggest threat to NVIDIA is its own customers, the frontier labs. NVIDIA has to become a frontier lab faster than frontier labs are becoming chip companies. And with Jensen out, they likely won't maintain the momentum to get there. With recursive self-improvement (RSI), every frontier lab and hyperscaler must become a chip company or lose the AI race. It's that simple, and they've all laid the groundwork (Google TPUs, OpenAI and Meta building custom ASICs with Broadcom, Anthropic building an in-house silicon team, and SpaceX/xAI/Tesla building Terafab). That's because software-only RSI has clear physical limits. Software can reduce how much computation and data movement a model needs, but it can't change the chip's physical bandwidth, memory capacity, or interconnect. Custom silicon gives labs another way to improve performance: designing the memory system, compute, and networking around their own workloads. To get the real compounding effect of RSI, you have to co-design the model and the silicon so it becomes a two-way feedback loop: smarter models design better custom chips, and those chips train and run the next model faster. You can't do that on off-the-shelf NVIDIA GPUs because you get zero proprietary edge buying the same chip as everyone else, and getting your next-gen architecture into NVIDIA's silicon means sharing your roadmap with them years before tape-out (which they then sell to your rivals). Plus, AI now automates writing compilers and porting CUDA kernels to other chips, which kills the CUDA moat. Jensen argued custom ASICs aren't more cost-efficient than GPUs because models change too fast for fixed hardware, and partners like Broadcom charge fat margins too. But hardware doesn't fall behind model changes when AI co-designs the next model and chip together instead of humans designing them separately (which Google and OpenAI are already doing). And labs aren't building custom silicon just to save on NVIDIA's 75%+ gross margins. Closing that hardware-software loop is the main goal, though the cost savings are still massive since RSI requires 24/7 internal inference where custom ASICs are way cheaper and more power-efficient than general-purpose GPUs. That's why NVIDIA has been trying so hard to become a frontier open-weight lab, buying Groq (fast inference IP), Poolside + Nemotron (internal coding models), and Hugging Face (distribution). If closed labs win, enterprises just call their APIs running on the labs' own custom silicon and NVIDIA gets cut out. Subsidizing free open-weight models is the only way to make enterprises buy or rent NVIDIA GPUs to run the models themselves. With Jensen gone, it's unclear if the Hugging Face deal even survives DOJ approval (can a GPU company own the whole stack?), and a non-founder is unlikely to take the massive risks needed to become a frontier lab: putting billions of capex into internal clusters and free open-weight models instead of just selling GPUs, while competing directly with its own biggest customers $GOOG, $META, $SPCX, $ANT, $MSFT 🔗 View Quoted Tweet📊 38 ⚡
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Key entities and technical topics under continuous tracking include NVIDIA, TPU, ASIC, Broadcom.
Source attribution: curated from Naval (@naval). Cross-source corroboration: 11 independent sources have covered this development (including theverge, rohanpaul_ai, @cb_doge), confirming industry-wide signal strength.