Why AI companies are building their own chips

For most of its life, Anthropic has done what every AI company does: rented computing power from other people's hardware. It has deals with AWS, Google, Nvidia, and AMD to run its Claude models on their chips. That arrangement has worked well enough to make Claude one of the most-used AI systems in the world. But on 5 August 2026, Anthropic quietly confirmed it is building a team to design its own custom chips.

It is easy to file this under "tech company does tech thing" and move on. But the decision tells you something important about where AI is heading, and about a constraint that most people outside the industry never think about: the chip underneath the model shapes almost everything about how that model behaves in the real world.

Why chips matter so much for AI

When you type a question into an AI assistant, the answer does not simply appear. The model has to perform an enormous number of calculations, multiplying huge arrays of numbers together millions of times in rapid succession, to turn your input into a coherent response. That process is called inference, and it happens every single time anyone uses the model.

The chip doing all that arithmetic matters enormously. A chip designed for general-purpose computing, like the CPU in your laptop, can do inference, but slowly and at great expense. Graphics chips (GPUs), originally built for rendering video games, turned out to be much better suited to the parallel arithmetic AI requires, which is why Nvidia became the dominant supplier. But a GPU is still a general-purpose tool. It handles thousands of different workloads, and its design reflects that flexibility.

A chip designed specifically to run one AI model, and nothing else, can be far more efficient. It can be laid out to match exactly how that model processes information, cutting out circuitry the model never uses and optimising the paths the data travels. The result is lower cost per query, faster responses, and less electricity consumed per answer.

This is not a theoretical benefit. Anthropic's stated goal is to co-design silicon and Claude models together, targeting roughly 50% cuts in per-token inference costs. Cut your cost per token in half and you can either make the product much cheaper, or run it twice as often for the same money. At the scale Anthropic operates, those numbers are significant.

What "co-design" actually means

The word that keeps appearing in Anthropic's announcements is "co-design". It sounds like jargon, but the idea is straightforward.

Normally, a chip company designs a chip, and an AI company designs a model, and the two meet somewhere in the middle. The AI company adapts its software to run on available hardware. Co-design means doing both at the same time, with the same team. The plan is to co-design hardware and models to help Claude run faster and more efficiently.

Think of it like designing a car and a road at the same time. A car built to travel a specific road can be lighter, faster, and more fuel-efficient than one designed to handle any road in any condition. If you know the road will always be smooth tarmac with gentle curves, you do not need heavy-duty suspension for potholes or a steering system calibrated for hairpin bends.

For AI chips, the equivalent of "knowing the road" is understanding the precise mathematical operations your model performs. Claude, like all large language models, relies heavily on a type of calculation called an attention mechanism, the process by which the model weighs how relevant each word in a sentence is to every other word. The co-design of hardware and software allows Anthropic to tailor chip architecture directly to Claude's attention mechanisms. A chip built around that specific operation can execute it far faster than a chip designed to handle everything.

Anthropic is not the first

It is worth noting that Anthropic is late to this particular race. Google has been running its AI models on its own custom chips, called TPUs (Tensor Processing Units), since 2016. Anthropic isn't the first AI company to decide to build its own chip. In June, OpenAI unveiled its Broadcom-built Jalapeño chip, which is designed specifically for AI inference.

The broader trend is accelerating. Custom AI accelerator shipments are on track to grow 44.6% in 2026, roughly three times the 16.1% growth rate projected for general-purpose GPUs. That gap tells you which way the industry is moving.

Anthropic's specific approach has an interesting twist. Anthropic is hiring for chip-design reinforcement learning work involving RTL generation, verification, and physical-design optimisation. The company is moving from exploring custom chips to hiring engineers who could use reinforcement learning to improve parts of the chip-design process itself. In other words, Anthropic wants to use AI to help design its AI chips. That is either a neat bit of recursion or a sign of how far AI-assisted engineering has come, depending on how you look at it.

This is not a break from Nvidia

The obvious question is whether Anthropic is trying to cut Nvidia out of its supply chain. The honest answer is: not really, at least not yet.

Anthropic emphasises it's not an "escape from Nvidia" but a cost optimisation within the existing ecosystem. The company maintains a multi-chip approach, continuing partnerships with major providers. Anthropic has inked deals with AWS, Google, Nvidia, and AMD to access AI computing hardware.

Custom chips take years to develop. The timeline from hiring initial semiconductor talent to deploying functional data centre hardware typically spans several years. Anthropic is not announcing a finished product. It is announcing a team, a direction, and a set of job listings. Anthropic has confirmed an in-house chip-design effort, but has not disclosed the team's size, a first-chip schedule, or a production partner. Samsung has been reported as a potential manufacturing partner, but nothing is confirmed.

So for the next few years, Anthropic will keep running Claude on the same mix of hardware it uses today. The custom chips, if and when they arrive, will be an addition to that mix rather than a replacement for it. The company's ambition is to reduce its dependence on outside suppliers over time, not to flip a switch overnight.

If this one was useful, there's plenty more on the site. Pieces on how AI works, plus coverage of AI news, the downsides included. All free to read, no account needed.

Why this matters to people who just use AI

If you use Claude, or any AI tool, you probably do not spend much time thinking about the chip running it. But the chip affects things you do notice: how fast the response arrives, whether the service goes down under heavy load, and how much the subscription costs.

The economics here are striking. The per-query cost of running a model equivalent to GPT-3.5 fell from approximately $20 per million tokens in November 2022 to $0.07 per million tokens by late 2024. That collapse in cost is what made it feasible to offer free tiers, to embed AI into other software, and to run models that answer billions of queries a day. Custom chips are one of the tools that could push those costs down further still.

There is also a less obvious implication. Right now, a handful of chip manufacturers, Nvidia especially, sit at a chokepoint in the AI industry. Every major AI lab depends on them. That dependence gives those manufacturers enormous leverage over pricing and supply. Anthropic's decision to design its own chips comes as demand for Claude rises while AI companies snatch up as many AI infrastructure deals as they can. Building your own chips is, among other things, a way to reduce exposure to that chokepoint.

Whether Anthropic's bet pays off will take years to see. Designing a chip from scratch is genuinely hard. Many companies have announced custom silicon programmes and quietly shelved them when the engineering reality proved more complicated than the press release suggested. But the direction of travel is clear. The AI companies that can control their own hardware will have an advantage in cost, speed, and resilience that those relying entirely on outside suppliers cannot easily match. Anthropic has decided it wants to be in that first group.