Anthropic weighs making in-house AI chips, sources say

Anthropic is exploring whether to design its own AI chips in response to a shortage of chips needed to power and develop more advanced AI systems, according to three people familiar with the matter.

The effort is still preliminary, and the company has not chosen a design or formed a dedicated team, one source said.

The company may still decide to only buy AI chips.

Anthropic uses a mix of chips for Claude, including Google TPUs and Amazon’s chips.

Earlier this week, it also signed a long-term deal with Google and Broadcom, which helps design the TPUs.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Anthropic’s new deal already brings custom chip perks

The custom chip idea follows a commitment to expand with Google and Broadcom for 3.5 gigawatts of next-generation Tensor Processing Unit (TPU) capacity, with rollout expected to start in 2027 1.
The arrangement goes beyond a standard buyer relationship. It ties Claude software work to chip design so the models can run better on Google TPU lines such as Ironwood (TPU v7) and the planned Zebrafish (TPU v8) 2.
The partnership could lower spending. One analysis puts TPU infrastructure at US$30–35 billion per gigawatt versus about US$50 billion for comparable Nvidia systems 3.
Anthropic said the added compute will power frontier Claude models. It has not said it will deploy Broadcom-built TPU racks in its own data centers 4.

Supplier control matters more than chip building

CEO Dario Amodei has warned that a one-year mistake in capital spending could trigger bankruptcy. That makes a US$500 million chip effort a serious bet 5.
Talking publicly about a custom chip can strengthen negotiations, sending a message to Google, Amazon, and other providers that the company will not stick to one platform.
The infrastructure plan already spans multiple chip sources, including Amazon Web Services (AWS) Trainium, Google TPUs, and Nvidia graphics processing units (GPUs), which helps performance while limiting reliance on a single vendor 4.
This trend suggests large model builders like Anthropic are gaining more sway over hardware supply, with long-term compute costs as the driver 3.