Anthropic Wants Its Own AI Chips

The company behind Claude is building an in-house chip design team and has hired a veteran of Google’s TPU programme. The move signals that the AI race is expanding beyond models and into the hardware underneath them.

Anthropic wants more control over the machines that run Claude.

The company has announced plans to build an in-house team focused on custom AI chip design, confirming a major shift for one of the world’s leading AI companies.

Then came a significant hire.

Anthropic has brought in Amir Salek, a veteran of Google’s custom chip programme and a key figure in the development of Google’s Tensor Processing Units, or TPUs. Salek previously led Google’s TPU business and oversaw the delivery of its first seven generations of chips. (Business Insider)

The message is clear.

Anthropic does not want its future to depend entirely on buying computing hardware from other companies.

Why AI companies care so much about chips

The AI boom has created an enormous demand for computing power.

Every time someone asks Claude a question, generates code or runs an AI agent, powerful processors somewhere in a data centre have to perform the calculations.

Training a frontier AI model requires even more.

That has made advanced AI chips one of the most strategically important resources in the technology industry.

Nvidia currently dominates the market for high-end AI accelerators, while companies such as Google, Amazon and Meta have also developed or commissioned their own specialised chips.

Anthropic has so far followed a diversified strategy.

The company uses hardware from Nvidia, Google and Amazon Web Services, as well as other suppliers. It has stressed that its custom silicon initiative will not replace that approach. Instead, it describes the strategy as multi-chip, with different hardware used for different requirements. (Reuters)

Why build a chip instead of just buying one?

The obvious question is: if Nvidia already makes extremely powerful AI processors, why would Anthropic spend money developing its own?

Control is one answer.

Custom silicon can be designed around the specific workloads a company cares about.

Instead of building a general-purpose accelerator that must serve thousands of different customers, an AI company can design hardware around the way its own models operate.

That can potentially improve performance and efficiency.

It can also help address supply constraints.

The demand for AI computing has become so intense that access to advanced chips and data-centre capacity has become a strategic issue for the largest AI companies. Reuters reported that Anthropic’s chip initiative is partly driven by the continuing shortage of hardware needed to develop and operate more advanced AI systems. (Reuters)

Anthropic is not alone

This is becoming a broader industry trend.

Google has been developing its TPU architecture for years.

Amazon has its own Trainium and Inferentia chips.

Meta has been developing custom AI accelerators.

Microsoft has also invested heavily in its own AI silicon.

And OpenAI has been moving toward greater control over its hardware requirements as its computing needs grow.

The result is a gradual change in the structure of the AI industry.

The biggest AI companies are increasingly trying to control multiple layers of the stack:

Models. Software. Chips. Data centres. Energy.

The more layers they control, the less dependent they become on outside suppliers.

Nvidia is still very much in the picture

Anthropic’s move should not be interpreted as the end of Nvidia’s dominance.

Far from it.

Anthropic says it will continue using hardware from Nvidia alongside chips from Google, Amazon and other suppliers. (Reuters)

Designing a competitive AI processor is also extremely difficult.

It requires specialised semiconductor engineers, software engineers, manufacturing partners and enormous amounts of capital.

Reuters reported that industry sources estimate designing an advanced AI chip can cost around US$500 million, although actual costs vary significantly depending on the complexity and manufacturing process. (Reuters)

Anthropic has not announced a timeline for when its own chips will be ready, nor has it said that it intends to manufacture them itself.

So this is a long-term strategy, not an imminent replacement for Nvidia.

The real prize is efficiency

The most interesting part of Anthropic’s strategy may not be about replacing Nvidia at all.

It is about getting more useful work from every dollar spent on computing.

AI companies are facing an unusual economic problem.

Their products are becoming more popular, but serving those products requires enormous amounts of computing power.

If an AI company can make its models run faster or more efficiently on specialised hardware, even a relatively modest improvement can become financially significant at enormous scale.

Consider the difference between running an AI system for a few million users and running it for hundreds of millions.

A small efficiency improvement multiplied across that scale can translate into enormous savings.

Claude is becoming more than a model

There is another strategic dimension.

Anthropic increasingly operates a full AI platform rather than simply selling access to a language model.

Claude is being used for coding, enterprise work, research and increasingly autonomous tasks.

Those workloads place different demands on computing infrastructure.

An AI system that spends substantial amounts of time reasoning, writing code or running tools may have very different hardware requirements from a conventional application.

Designing hardware and models together could therefore allow Anthropic to optimise the entire system.

That is the thinking behind the growing industry trend toward co-design, where software and hardware are developed together rather than independently.

A new battle is forming underneath the AI race

For years, the public AI competition was mostly about models.

Who had the smartest chatbot?

Who could write the best code?

Who had the strongest reasoning?

Now another competition is developing underneath it.

Who controls the computing infrastructure?

That question matters because advanced AI cannot exist without enormous amounts of computation.

The industry is already spending vast sums on data centres, networking, power generation and specialised processors.

Broadcom, for example, is reportedly negotiating more than US$60 billion in debt financing connected to AI chip development and infrastructure, illustrating the extraordinary amount of capital now flowing into the hardware side of the AI economy. (Reuters)

What Anthropic’s move really means

Anthropic building its own chip team does not mean Nvidia is about to disappear.

It does mean something more important.

The AI industry is maturing.

The companies building the most capable models are discovering that controlling the software alone may not be enough.

As AI becomes more computationally demanding, access to chips, data centres and electricity becomes part of competitive advantage.

Anthropic’s hiring of a Google TPU veteran is therefore more than a personnel announcement.

It is a signal that the next phase of the AI race will be fought not only in model labs, but inside semiconductor design teams and data centres.

And the winners may ultimately be the companies that can make the entire AI stack work together most efficiently.