Anthropic has confirmed what Reuters first reported as speculation earlier this year: the AI lab behind Claude is building an in-house silicon team to design its own custom AI chips. The move puts Anthropic alongside Google, Amazon, Meta, and OpenAI in a growing race among major AI developers to control more of their own hardware stack — though the company has been careful to frame the effort as a cost and performance play rather than a break from Nvidia.
What Was Actually Announced
Anthropic has begun actively hiring for a dedicated chip design team, with job listings offering salaries of up to $485,000 and specifically seeking engineers who have personally shipped finished semiconductor designs. The listings emphasize machine-learning accelerators, high-bandwidth memory subsystems, and hardware-software co-design — details suggesting a chip built specifically around how Claude actually trains and runs inference, rather than a general-purpose processor. Notably, the company hasn’t disclosed a timeline for when a first chip might be ready, nor confirmed whether it plans to manufacture the processors itself, though reports suggest early conversations with Samsung.
Crucially, this isn’t a pivot away from existing hardware partners. Anthropic has been explicit that it will maintain a multi-chip strategy, continuing to rely on processors from AWS, Google, Nvidia, and AMD even as its custom silicon program develops. As one industry analysis put it, the move is less an escape from Nvidia than a matter of arithmetic — the economics of serving billions of tokens a day at Anthropic’s current scale.
Why Now: The Numbers Behind the Decision
The timing tracks closely with Anthropic’s growth trajectory. The company’s run-rate revenue surpassed $30 billion in April 2026, up sharply from roughly $9 billion at the end of 2025, with the number of customers spending more than $1 million annually more than doubling to over 1,000 in under two months. That kind of scale changes the math on custom hardware: developing a cutting-edge AI chip typically costs somewhere around $500 million before manufacturing expenses even enter the picture — a bet only worth making once a company’s compute bill is large enough for the savings to outweigh the upfront cost.
Anthropic has been building toward this on the supply side for a while. The company took over SpaceX’s Colossus 1 data center earlier this year to help power Claude, and has separately announced plans to expand its use of Google Cloud technology, including access to up to one million TPUs as part of a deal worth tens of billions of dollars, expected to bring more than a gigawatt of additional compute capacity online in 2026.
The Bigger Industry Pattern
Anthropic’s announcement fits neatly into a broader shift already underway across the frontier AI industry. Google has built its own TPU line for years; Amazon has developed Trainium and Inferentia chips for training and inference; Meta has expanded a custom-chip partnership with Broadcom through 2029; and OpenAI has already unveiled its own inference chip, called Jalapeño, built with Broadcom, which reportedly delivered per-token cost savings of roughly 50% compared to standard GPU inference in early testing. Anthropic is reportedly targeting similar cost reductions — cutting per-token inference costs by roughly half — by having its own models and chips designed together from the start, rather than fitting Claude onto general-purpose hardware built for a wide range of workloads.
The underlying driver connecting all of this, according to industry observers, is a persistent shortage of advanced AI chips that constrains nearly every major lab simultaneously — a bottleneck the industry is trying to solve through some combination of expanded manufacturing capacity, custom chip design, and hardware-software efficiency gains, often all three at once.
What Doesn’t Change
It’s worth being clear about what this announcement is not. Anthropic’s compute today runs entirely on other companies’ silicon, and that won’t flip overnight — there’s no confirmed timeline for a finished chip, and even once one exists, it will supplement rather than replace the company’s existing Nvidia, AMD, Google, and AWS partnerships. Custom silicon in this industry tends to redistribute spending across the chip ecosystem rather than eliminate it entirely, since the underlying manufacturing still runs through the same handful of foundries and suppliers regardless of whose logo is on the final design.
The Bottom Line
Anthropic’s confirmation of an in-house chip team is less a dramatic break from Nvidia and more a signal of how mainstream vertical integration has become across the AI industry. When serving billions of tokens a day is the baseline, cutting per-token cost by even a modest percentage compounds into serious money — and designing hardware and models together, rather than adapting models to whatever hardware happens to be available, is becoming the default strategy for labs operating at Anthropic’s scale. Whether this yields a finished chip in the near term remains an open question, but the company has clearly decided the arithmetic is worth the bet.
Do you think custom silicon will meaningfully shift the balance of power away from Nvidia, or is this more about incremental cost savings for the biggest players? Share your take in the comments, and subscribe for more coverage of the infrastructure race shaping AI in 2026.