Published 18 hours ago

Anthropic Confirms Custom Chip Plans for AI Inference and Model Scaling

Anthropic has confirmed it is building an in-house custom silicon team — a move that signals a deliberate shift toward greater control over the compute infrastructure that runs its Claude models. The confirmation came after job listings for a silicon engineer and a technical program manager appeared on Anthropic’s job board, prompting the company to clarify its intentions to multiple outlets.

The plan is not to abandon third-party hardware entirely. Anthropic’s spokesperson was explicit: the company will maintain a multi-chip approach, continuing to use hardware from external vendors while developing its own designs in parallel.

3 mins read

5 sections

Key Highlights

Anthropic is hiring an in-house silicon team to design custom chips for Claude inference.
The company will keep a multi-chip strategy while reducing strategic dependence on Nvidia.
Custom silicon marks a hardware-layer battle for frontier AI cost, latency, and scale.

Why AI Labs Are Building Their Own Chips

The pattern is now well established. Google has run its models on custom TPUs for years. Meta has designed and deployed its own silicon. OpenAI recently announced a chip called Jalapeño, developed with Broadcom, specifically targeting large language model inference in data centers. Mistral is reportedly exploring a similar path.

Two structural pressures are driving this trend.

Nvidia dependence is a strategic liability. A significant portion of the AI industry runs on Nvidia hardware. In an environment where compute demand consistently outpaces supply, that concentration creates real competitive risk. Any company that can reduce that dependency gains negotiating leverage and supply chain resilience.

Hardware-model co-design can improve performance. When the team building the model and the team designing the chip work side by side, there is potential to optimize both in ways that off-the-shelf hardware cannot accommodate. Anthropic has explicitly stated this co-design approach is central to its plan — not just procuring chips, but shaping them around how its models actually work.

This trend reflects a broader push for greater control over compute infrastructure.

What Anthropic Is Actually Doing

The company has previously co-designed certain hardware with external partners. The new direction is to bring that expertise inside Anthropic itself, building an internal team rather than relying on collaborative arrangements with third parties.

Earlier reporting suggested Samsung as a potential manufacturing partner, which would be consistent with a fabless design model — Anthropic designs the chip, a foundry produces it. That arrangement has not been officially confirmed, but it reflects a plausible path given how other AI labs have structured similar efforts.

The Competitive Context

As smaller, open-weight models become more capable, some developers are already shifting workloads to their own hardware or edge devices. Frontier model providers like Anthropic face pressure not just from each other, but from a broader ecosystem that is becoming less dependent on hosted inference.

Custom silicon could help Anthropic maintain a performance and cost edge in that environment — but the timeline is not short. The team is still being assembled, which means any practical benefit to Anthropic’s infrastructure or its users remains some distance away.

What to Watch

The confirmation itself is the meaningful signal here. Anthropic is not experimenting at the margins — it is making a structural investment in compute infrastructure at the same time its competitors are doing the same. For anyone tracking the AI tools ecosystem, this is a reminder that the competitive dynamics in frontier AI are increasingly being fought at the hardware layer, not just the model layer.

The companies that control their inference stack — cost, latency, and throughput — will have options that those dependent on shared infrastructure simply will not.