跳到正文
原文
The Decoder:AI News· Matthias Bastian·· 5 天前AI 评分67

Anthropic 称 AI 正在侵蚀 Nvidia 的 CUDA 优势,移植 AMD 芯片从数月缩短到 3 天

Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days

文章要点
  • 1Anthropic 称其前沿模型仅 3 天就把 Claude 移植到 AMD GPU,几乎无人工干预
  • 2官方称内核近 100% 由模型 Mythos 自己编写,传统方式上线新芯片需 6 到 12 个月
  • 3媒体称若数字属实,Nvidia 靠 CUDA 构筑的软件锁定优势将明显缩水,其硬件毛利率超 70%
  • 4核心数字出自投资者路演未经独立验证;Nvidia 承诺至多投 100 亿美元,AMD 至多 50 亿美元
决策要点

Anthropic 宣称 AI 本身可以瓦解 Nvidia 的 CUDA 软件护城河,若成立,芯片选择的逻辑将从生态绑定转向按性价比和可得性调度。

  • 影响:若数字属实,算力采购可跨厂商按性价比与可得性比较,绑定单一芯片生态的切换成本会明显下降;但核心数据出自上市前投资者材料,Nvidia 与 AMD 均是 Anthropic 投资方,引用该口径时应留有余地。
  • 参考:可考虑先盘点自家模型与推理软件栈对特定芯片生态的依赖深度,估算真实迁移成本;值得在下一轮算力采购中要求供应商提供跨架构方案与报价,把切换成本计入总拥有成本后再比较。
  • 依据:核心数字出自 Anthropic 投资者路演,暂无独立验证;传统路径 6 到 12 个月亦为该公司口径,披露文件显示 Nvidia 与 AMD 分别承诺至多 100 亿与 50 亿美元投资。
正文

Anthropic claims it has broken free of its Nvidia dependence with a multi-chip strategy. The company says it ported its own AI model, Claude, to AMD GPUs in just three days, with almost no human intervention.

Nvidia's biggest competitive edge is the combination of hardware and software, specifically CUDA, the software ecosystem that locks developers into Nvidia chips. Porting AI models to other chip architectures takes so much effort that many companies won't even try. Anthropic now claims in an investor presentation that it has solved this problem using AI itself.

According to Tom Brown, co-founder and Chief Compute Officer at Anthropic, the company now trains and runs Claude on all four major chip families. Anthropic made an "unconventional bet" on a multi-chip strategy early on, at a time when most of the industry considered the effort not worth it. Each chip needs its own custom software and inference stack, which is expensive and time-consuming. But Anthropic has been investing in this work for years and has since become "incredibly good" at it, Brown says.

Anthropic uses AI to port Claude to AMD

The company's most concrete test came with AMD GPUs. Until now, bringing up a new chip platform kept dozens of engineers busy for months. This year, Anthropic used AMD GPUs for the first time and let its own frontier model, Mythos, do the work. According to Anthropic, Claude was running on AMD after just three days, with almost no human intervention. The model ported or wrote nearly 100 percent of the kernels itself.

Kurvendiagramm: Mythos erreicht ~100 % Leistung in ~3 Tagen, traditioneller Ansatz benötigt ~6–12 Monate auf AMD-GPUs.
With Mythos, Anthropic brings new hardware to a usable performance level in about three days, instead of the six to twelve months traditional engineering approaches need to onboard new accelerators. The figures refer to prompt processing of an Anthropic model on AMD MI355X GPUs. | Image: Anthropic

If these figures are accurate, the software overhead, long considered Nvidia's strongest competitive advantage, would shrink dramatically because of AI. Companies could switch more easily between chip providers and pick the most affordable or readily available option.

In a compute market where supply is tight, a multi-chip approach lets companies focus on the best price-performance ratio, regardless of the vendor. Some chips are better for training, while others are better for inference. And because of its dominance in the AI chip market, Nvidia currently enjoys a margin of more than 70 percent, which is unusually high for hardware.

Rasterdiagramm: Fungibler Compute-Pool verteilt AWS-Trainium-, Google-TPU-, Nvidia- und AMD-GPU-Ressourcen auf vier Workloads.
Anthropic allocates computing capacity from a fungible pool across four hardware platforms—AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs—for tasks such as research, model development, customer demand, and internal Claude usage. | Image: Anthropic

Anthropic treats compute from different vendors, chip architectures, and locations as one interchangeable pool. The company says it gets more out of every dollar of compute than any other frontier AI lab.

Compute multipliers aim to widen the lead

Beyond its multi-chip strategy, Anthropic relies on what it calls compute multipliers, research breakthroughs that squeeze more performance from the same physical computing capacity. According to Anthropic, these multipliers build on one another so that each new breakthrough amplifies the effect of the previous one.

Eventually, Claude should be able to discover such compute multipliers on its own. That would amount to recursive self-improvement, a concept that's intensely debated in AI research but hasn't been proven yet.

Anthropic is also building its own data center capacity through the "Anthropic Direct" program. Instead of buying computing power only as a cloud service, the company leases chips and operates them in purpose-built data centers. The goal is to reduce total cost of ownership.

The figures come from Anthropic's investor presentation and haven't been independently verified. But they show that at least one major AI lab no longer considers Nvidia's CUDA dominance insurmountable and that AI models themselves could become a tool for breaking hardware dependencies.

Both major GPU makers now have deep financial ties to Anthropic. In November 2025, as part of a $45 billion deal with Microsoft, Nvidia committed to investing up to $10 billion in Anthropic. According to Anthropic's S-1/A registration statement filed with the SEC on November 10, 2026, Nvidia is also participating in a conditional private placement totaling $26 billion, in which seven technology companies are subscribing to shares at the IPO price.

AMD announced a strategic investment of up to $5 billion in July 2026; in return, Anthropic plans to deploy up to two gigawatts of AMD Instinct MI455X GPUs in Helios rack-scale systems, with the first gigawatt phase beginning in the first half of 2027.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

来源:The Decoder:AI News · the-decoder.com