이미지: Ars Technica
Summary
- Anthropic has officially confirmed the formation of a custom silicon team to design chips dedicated to its Claude models.
- The company plans to maintain a "multi-chip strategy," using its own chips alongside hardware from other vendors.
- Major AI firms including OpenAI, Google, and Meta are already pursuing similar vertical integration moves.
Anthropic officially confirms custom silicon team
Anthropic has officially acknowledged plans to directly design dedicated semiconductors to power its AI model, Claude. The news came to light after Business Insider spotted job postings on Anthropic's careers page seeking senior engineers with semiconductor design experience; an Anthropic spokesperson subsequently confirmed the plans to both Business Insider and TechCrunch. Anthropic's careers page currently lists openings for a silicon engineer and a technical program manager (silicon).
The spokesperson made clear that Anthropic will continue to pursue a "multi-chip approach" — meaning the company will not rely solely on chips it designs itself, but will scale up by using hardware from other companies alongside its own. Prior to this announcement, The Information had reported that Anthropic was considering Samsung as a hardware manufacturing partner; this official confirmation lends substance to that earlier report.
Why now, why in-house silicon
There are two structural reasons behind Anthropic's decision to build a silicon team at this point in time. The first is dependence on NVIDIA. Many companies in the AI industry today rely heavily on NVIDIA GPUs for both model training and inference. In an environment where demand consistently outstrips supply, this kind of concentration can become a strategic vulnerability — and that risk grows as competition for compute infrastructure intensifies.
The second reason is the potential for optimization between models and hardware. Designing chips tailored to a specific model, or conversely developing models that take advantage of particular hardware characteristics, can yield better performance than general-purpose chips. Anthropic has clearly recognized this, stating that it plans to have its silicon team and model team work side by side in a co-design structure. While the company has previously co-designed some hardware with partners, the key difference this time is that it intends to bring that expertise fully in-house.
An industry-wide trend toward vertical integration
Anthropic's move fits within a broader industry pattern. OpenAI recently unveiled a custom chip called "Jalapeño," designed specifically for large-scale language model inference in data centers, developed in partnership with Broadcom. Google has long run its models on its own hardware, and Meta likewise designs its own chips and deploys them in production services. Mistral is also reportedly considering a similar direction.
The wave of major AI companies moving to bring chip development in-house isn't simply about cutting costs. Underlying it is a strategic calculation to boost both model performance and infrastructure efficiency through vertical integration, while reducing supply chain risk. For Anthropic, there is also the backdrop of needing to respond to a growing trend of software developers running smaller or open-weight models on their own hardware or edge devices.
Current status and limitations
That said, tangible results will take considerable time to materialize. Anthropic itself has stated that hiring for key team positions is still underway, and given the time required to go from chip design to mass production and real-world deployment, visible outcomes are unlikely in the near term. Specific terms of any manufacturing partnership, as well as specifications and release timelines for a first chip, have not been disclosed.
Pricing, billing models, and access pathways for external developers were also not mentioned in available sources. This announcement pertains to internal infrastructure strategy, and is better understood as a move to secure medium- to long-term competitiveness rather than something users will notice as an immediate change in service.
