Anthropic has formed an internal engineering team to develop custom AI chips for Claude, giving the company more control over the hardware behind its models. Nvidia, AMD, Google and Amazon technologies will remain part of its computing infrastructure.
Anthropic is moving beyond model development and into chip design.
The company has confirmed the creation of an in-house team that will work on custom processors for Claude. Hardware and software engineers are being recruited to develop chips alongside the AI models that will eventually run on them.
Anthropic says the work is intended to help Claude operate faster and more efficiently at the scale required by customers. The company has not disclosed when its first design could be completed, which manufacturing process it might use or who would physically produce the chips.
A processor designed around Claude
Most leading AI models run on processors developed for a broad range of customers and workloads. Anthropic’s plan is to design hardware more closely around Claude’s own computational requirements.
A custom chip could be configured around the model’s memory use, data movement and most frequently performed mathematical operations. That may allow Anthropic to extract more performance from the same amount of computing capacity.
The company has not released technical targets, so it remains unclear how much faster or more energy-efficient Claude could become. Any improvement would also depend on the software developed to connect the models with the new hardware.
Anthropic is joining a wider industry shift toward custom AI silicon. Large technology companies have increasingly designed their own accelerators as demand for computing power has risen and access to the most advanced processors has become strategically important.
Nvidia, Google and AWS will remain in use
The new project does not mean Anthropic is abandoning its existing chip suppliers.
The company describes the initiative as the latest step in a multi-chip strategy that includes technologies from Amazon Web Services, Google, Nvidia and AMD. Claude will continue to run across different hardware platforms while Anthropic develops its own processors.
Using several suppliers reduces the risk of depending on a single source of computing capacity. It also allows different workloads to be assigned to the hardware best suited to them.
Anthropic’s own chip could eventually become another option within that infrastructure rather than a complete replacement for commercially available processors.
This distinction matters because the company would still need substantial external capacity even after introducing a custom design. Training and operating large AI models requires enormous numbers of chips, extensive data-centre infrastructure and reliable access to electricity.
Developing an AI chip can cost hundreds of millions
Building an advanced AI processor is a costly and technically demanding undertaking. Industry estimates suggest that designing a competitive chip can require an investment approaching $500 million.
The cost includes specialist engineering teams, design verification, test production and the work needed to identify errors before large-scale manufacturing begins. A flaw discovered late in the process can delay production and significantly increase the final bill.
The chip itself is only one part of the project. Anthropic will also need software tools that allow Claude to use the processor efficiently and enable engineers to deploy it inside large data centres.
This combination of hardware and software development is one reason custom chip programmes can take several years to move from initial design to commercial use.
Anthropic has not said whether it intends to manufacture the processors itself. Most companies that design custom chips rely on specialist semiconductor manufacturers to produce them.
What the project could mean for Claude
It is too early to determine whether Anthropic’s chips will affect Claude’s speed, pricing or availability.
A successful design could give the company more control over how its models are trained and operated. It may also help Anthropic secure additional computing capacity as demand for Claude grows.
However, those benefits will depend on the chip’s performance, production cost and availability at scale. A custom processor must compete not only with today’s hardware but also with newer generations of chips released while the project is under development.
Anthropic has not announced a launch date or shared specifications for the planned processor. It has also not confirmed when customers might first encounter a version of Claude running on the company’s own silicon.
For now, the clearest change is organisational: Anthropic is building a dedicated team to design the hardware behind future Claude systems while continuing to rely on its existing chip partners.