Anthropic has confirmed that it is assembling a team to design its own custom chips, marking one of the clearest signs yet that the company behind Claude intends to control more of the hardware its models run on. The news was first reported by Business Insider and later confirmed directly to TechCrunch, with Anthropic saying it plans to co-design hardware and models so that its technology runs faster and more efficiently. A public job listing for the effort describes engineers being recruited into a "custom silicon team," a phrase that quietly signals long-term ambition rather than a one-off experiment.
The timing is not accidental. Demand for Claude has climbed steadily, and across the industry AI companies are locking up every scrap of computing capacity they can find. Anthropic already relies on an unusually broad set of suppliers, having signed deals with Amazon Web Services, Google, Nvidia, and AMD to secure the accelerators that train and serve its models. Designing its own silicon does not replace those relationships, but it does change the company's posture: from a buyer of compute to a co-author of the machines that produce it.
A separate report from The Information last month added an important detail, noting that Anthropic had been scouting Samsung as a potential manufacturing partner for the chips. That would fit the standard playbook for a company that wants a custom design without building a fabrication plant of its own, a path that would be prohibitively expensive and slow.
Why It Matters
For most of the current AI boom, the story has been about models. Whose system scores highest on a benchmark, whose chatbot feels smartest, whose weights are open. Anthropic's move is a reminder that the real competition is increasingly happening one layer down, in the silicon that decides how cheaply and quickly those models can actually be served to millions of users.
General-purpose graphics processors are extraordinary machines, but they are designed to do many things reasonably well rather than one thing perfectly. A chip built specifically for running a particular family of models can strip away that generality, trading flexibility for efficiency. When a company is serving inference at enormous scale, even small per-query gains in speed or power draw compound into very large savings. That is the logic behind co-design: shape the model and the chip together so that neither wastes effort on capabilities the other does not need.
There is also a strategic dimension that has little to do with engineering. Access to cutting-edge accelerators has become a genuine bottleneck, and pricing power sits largely with a handful of suppliers. A lab that designs its own chips gains leverage, insulation from shortages, and a measure of independence from the roadmaps of companies that are, in some cases, also its competitors. Compute is quietly becoming part of a frontier lab's intellectual property, not merely a line item it purchases.
The Reaction
Anthropic is not breaking new ground so much as joining a movement that its largest rivals have already embraced. The pattern across the industry is now consistent enough to look like a rule: once a lab reaches sufficient scale, it starts designing silicon.
In June, OpenAI unveiled a Broadcom-built chip, internally nicknamed Jalapeño, aimed specifically at inference workloads rather than training. Google DeepMind has for years leaned on Alphabet's Tensor Processing Units, the in-house accelerators that have quietly powered much of Google's AI work. Meta, meanwhile, has been developing its own MTIA accelerators to handle the enormous inference load generated by its apps. Seen against that backdrop, Anthropic's announcement reads less like a surprise and more like a company catching up to a competitive baseline.
Observers were quick to note the careful framing. Anthropic is hiring a design team and co-designing hardware; it has not announced a finished processor, a production timeline, or a claim that it is abandoning its existing suppliers. The multi-vendor strategy remains firmly in place. What has changed is that custom silicon has shifted from something Anthropic buys to something it also intends to shape.
What Comes Next
The practical questions now are about execution. Designing a competitive AI chip is one of the hardest problems in modern engineering, and turning a design into working, high-yield silicon requires manufacturing partners with rare expertise. The reported interest in Samsung points toward a fabrication relationship, but any first chips would likely be years away and aimed narrowly at inference, where the efficiency gains are most immediate and the design challenge is more tractable than for training.
Talent will be the near-term battleground. The job listing seeking experienced chip designers puts Anthropic into direct competition for a small pool of specialists already being courted by every major cloud provider and hardware firm. Building a credible silicon team from scratch is a multi-year commitment, and the pace at which Anthropic can hire will say a great deal about how serious and how fast this effort really is.
It is also worth watching how suppliers respond. Nvidia and AMD have every incentive to keep large customers close even as those customers begin designing around them, and cloud partners like AWS and Google already offer their own accelerators. The relationships are becoming more layered, part partnership and part quiet competition, and Anthropic will have to manage all of them at once.
Closing Thoughts
Step back, and Anthropic's decision fits a larger reshaping of what an AI company is. The stack that matters now runs from models down through inference software, chips, data centers, and ultimately the energy that powers them. For a while it was possible to compete by owning only the top of that stack. That era is ending.
By reaching toward the silicon layer, Anthropic is betting that durable advantage in AI will come not from any single clever model but from controlling how the whole system fits together. It is an expensive, slow, and uncertain bet, and there is no guarantee the company will ship a chip that beats what it can already buy. But the direction of travel is unmistakable. The most valuable AI companies increasingly want to own the machine, not just the mind that runs on it.
한글 요약
클로드를 만든 앤스로픽이 자체 AI 칩을 설계하기 위한 '커스텀 실리콘 팀'을 꾸리고 있다고 확인했습니다. 비즈니스 인사이더가 먼저 보도했고 앤스로픽이 테크크런치에 직접 확인한 내용으로, 하드웨어와 모델을 함께 설계해 기술을 더 빠르고 효율적으로 구동하겠다는 구상입니다. 지난달 디 인포메이션은 앤스로픽이 삼성을 잠재적 제조 파트너로 검토했다고 전했습니다. 앤스로픽은 이미 AWS·구글·엔비디아·AMD와 컴퓨팅 하드웨어 계약을 맺고 있지만, 클로드 수요가 급증하는 상황에서 공급사에만 의존하기는 어렵다는 판단으로 읽힙니다.
핵심은 경쟁의 무게중심이 '모델'에서 그 아래 '칩'으로 내려오고 있다는 점입니다. 범용 그래픽 프로세서는 다재다능하지만, 특정 모델을 겨냥한 전용 칩은 유연성을 포기하는 대신 효율을 얻습니다. 대규모 추론(inference)에서는 질의당 작은 속도·전력 이득도 누적되면 큰 비용 절감으로 이어집니다. 여기에 최신 가속기 확보가 병목이 되고 가격 결정권이 소수 공급사에 쏠린 현실이 더해지면서, 자체 칩 설계는 공급 부족에 대한 방어막이자 경쟁사이기도 한 공급사들로부터의 독립성을 뜻하게 됩니다.
앤스로픽만의 행보는 아닙니다. 오픈AI는 6월 브로드컴과 만든 추론 전용 칩(내부명 할라페뇨)을 공개했고, 구글 딥마인드는 오랫동안 알파벳의 TPU를, 메타는 자체 MTIA 가속기를 활용해 왔습니다. 다만 앤스로픽은 설계팀 구성과 하드웨어·모델 공동설계를 밝혔을 뿐, 완성된 칩이나 양산 일정, 기존 공급사와의 결별을 발표한 것은 아닙니다. 첫 칩은 몇 년 뒤 추론용으로 등장할 가능성이 높고, 희소한 칩 설계 인재 확보 속도가 이 시도의 진정성과 속도를 가늠하는 잣대가 될 전망입니다. 참고: TechCrunch, Reuters, Business Insider.