The rumors turned out to be true. Anthropic says it’s building an in-house team to design custom AI chips for Claude, giving the company more control over the hardware behind its AI models.
It’s also not blowing up its existing infrastructure. Anthropic says chips from AWS, Google, Nvidia and AMD will remain part of the mix.
Anthropic told Reuters that its “hiring engineers with experience across the hardware and software stack to help co-design custom chips and AI models that can make Claude run faster and more efficiently at the scale required by customers.”
Hardware built around the model
Buying whatever AI chip is available has worked well enough so far, but Anthropic apparently thinks “well enough” isn’t good enough anymore.
Instead, the company wants hardware built specifically for Claude, which can make inference, the part where models actually answer users’ questions, faster and more efficient. It’s also one of the biggest bills AI companies pay.
Even so, Anthropic isn’t throwing out its existing playbook. It says AWS, Google, Nvidia and AMD will remain part of its “multi-chip approach” while it develops its own silicon.
Business Insider also reported that Anthropic has started hiring for a dedicated custom silicon team (what a thing to put on one’s business card). One job posting outlines that it’s looking for engineers with experience across chip design and verification. “This is a role for someone who has shipped silicon, has a realistic relationship with schedules, and is comfortable making consequential calls without a large organization behind them,” the posting states.
OpenAI, Broadcom round out new chip race
Anthropic isn’t the only AI company dipping its toes into the chip pond. OpenAI announced its Jalapeño inference chip with Broadcom earlier this summer, Meta has its own AI chip efforts in the fire, and Mistral has said it’s considering custom silicon too.
There’s a spicy reason for that. Reuters reports that designing a leading-edge AI chip can cost roughly $500 million, which is a staggering upfront investment. But, companies spending billions to build and run AI models are increasingly deciding it’s worth it if they can squeeze more performance and efficiency out of their infrastructure.
Anthropic hasn’t said when its chip might arrive or who will manufacture it. Early reports indicate it could be Samsung, but Anthropic isn’t confirming any of that at this point.
The channel angle: why chips matter to pricing and deployment strategies
For MSPs and solution providers, Anthropic’s silicon strategy could eventually influence how Claude-based services are priced, deployed and supported. Custom chips may give Anthropic more control over inference costs, capacity and performance, but partners will still need to understand how those benefits are delivered through AWS, Google Cloud and other infrastructure providers.
The immediate issue is not whether MSPs will purchase Anthropic-designed chips directly. It is whether changes below the application layer alter cloud consumption costs, regional availability, latency, service-level commitments or the architecture required to support customer workloads.
Providers building managed AI services around Claude should watch for changes to instance options, usage pricing, deployment models and integrations across Anthropic’s existing infrastructure partners.
The multi-chip approach also reinforces the need to avoid designing customer environments around a single hardware assumption. MSPs evaluating AI platforms should compare model capabilities alongside infrastructure availability, data residency, operational tooling, portability and the economics of running inference at scale.
As AI vendors exert more control over their hardware stacks, those infrastructure decisions may increasingly affect partner margins and the long-term flexibility of customer deployments.
Anthropic’s custom chip plans fit into a much bigger infrastructure push. Earlier this year, the company expanded its TPU capacity through Google and Broadcom as it looked to support growing demand for Claude, and now it’s adding in-house silicon to the mix as another way to optimize how those models run. Read more here.