AnthropicがClaudeを支える独自ハードウェアの設計に乗り出すAnthropic will design its own hardware to power Claude

Anthropic has confirmed that it plans to build an in-house silicon team to design custom chips for running its Claude models, a move aimed at reducing its reliance on Nvidia's dominant GPUs. The decision matters because it signals that the cost and scarcity of AI accelerators have become strategic concerns for even the best-funded AI developers, and it places Anthropic alongside rivals that increasingly treat hardware as a competitive lever rather than a commodity to be purchased.

The effort mirrors similar work reported at OpenAI, and together the two companies illustrate a broader race among leading AI labs to gain more control over their hardware supply chains. Both are trying to scale up available compute while lowering their exposure to a single vendor. For now, the source material frames this as a race to "scale up while reducing dependence on Nvidia," rather than a claim that either company has finished or shipped a working chip. Building custom silicon is a multi-year undertaking, so the practical impact of Anthropic's team is likely to unfold gradually rather than immediately.

To understand why this is happening, it helps to look at Nvidia's position. The company's data-center GPUs, such as the H100 and its successors, have been the default hardware for both training and running large language models, and demand has consistently outstripped supply. That scarcity translates into high prices, long lead times, and significant leverage for Nvidia. Equally important is the software layer: Nvidia's CUDA platform and its surrounding ecosystem have become deeply entrenched, which raises the cost of switching to alternative hardware even when chips are available.

Anthropic is not the first to pursue an alternative. Google has developed its Tensor Processing Units, or TPUs, over multiple generations and uses them internally and through its cloud. Amazon has built its Trainium and Inferentia chips specifically for training and inference workloads. Microsoft has introduced its Maia accelerator, and Meta has been developing its MTIA silicon for internal use. Against that backdrop, Anthropic's move looks less like an outlier and more like the AI industry converging on a common conclusion that owning at least part of the hardware stack is worth the investment.

Anthropic's situation is notable because of its existing relationships with two of those chip makers. The company has taken substantial investment from Amazon and has also partnered with Google, and it has used both companies' custom accelerators to run its models. Designing its own silicon would give Anthropic a path that is more tightly optimized for its specific architectures and workloads, potentially improving performance per dollar and per watt, while still leaving room to keep using third-party chips where that makes sense. The company appears to be aiming for greater flexibility rather than a clean break from any single supplier.

The technical logic behind custom silicon is that general-purpose GPUs are designed to handle a wide range of tasks, whereas an application-specific integrated circuit, or ASIC, can be tuned for the narrower set of operations that dominate AI inference and training. That specialization can yield efficiency gains, which matter enormously at the scale these companies operate, where electricity and hardware costs are among the largest line items. Inference in particular, the process of actually serving model responses to users, is a repetitive and high-volume workload that is well suited to purpose-built chips.

The challenges are considerable. Designing competitive chips requires deep expertise, large budgets, and long development cycles, and firms typically depend on partners such as Broadcom or Marvell for design help and on foundries like TSMC for manufacturing. Even a strong chip is only useful if it is paired with mature software, compilers, and tooling, which is where Nvidia's incumbency remains strongest. For that reason, custom silicon is best understood as a complement to a diversified hardware strategy rather than an immediate replacement for Nvidia.

For the wider industry, Anthropic's confirmation reinforces a clear trend: as model training and deployment costs climb, the companies at the frontier are seeking to internalize more of their infrastructure. Whether these in-house chips ultimately match Nvidia's performance is unproven, but the strategic motivation, controlling cost, supply, and optimization, is now shared across nearly every major AI developer.