Anthropic Builds Custom AI Chips for Claude: What It Means for Nvidia, AMD, Samsung, and the AI Trade

Anthropic has confirmed it is building an in-house chip team to design custom AI processors for Claude. The stated goal is speed, efficiency, and scalability. The company said it will co-design hardware and models while continuing to buy chips from AWS, Google, Nvidia, and AMD under a multi chip strategy. Reports also point to talks with Samsung as a potential manufacturing partner. The move follows OpenAI unveiling its own custom AI chip earlier this year.

This is not a rumor about a startup dabbling in hardware. It is the second frontier lab in a year telling the market that the cost of inference is now a first order business problem, not an engineering footnote.

Why is Anthropic building its own AI chips?

Because the margin on a frontier model lives and dies on cost per token. Training gets the headlines. Inference pays the bills. When a model serves billions of requests a day, a twenty percent improvement in tokens per watt flows straight to gross margin, and it compounds every quarter as usage grows.

General purpose accelerators are built to serve everyone. A chip co-designed with one model family can strip out what that family never uses and spend the transistor budget on what it uses constantly. That is the same logic that pushed Google to build TPUs and Amazon to build Trainium and Inferentia. Anthropic is following a proven path, not inventing one.

There is a second motive that nobody puts in a press statement. Supply. Access to leading edge accelerators has been the binding constraint on frontier labs for three years. Owning a design gives you a seat at a second table.

Does this mean Anthropic is dropping Nvidia?

No, and the language in the announcement is deliberate. Anthropic said it will continue to rely on chips from AWS, Google, Nvidia, and AMD through a multi chip strategy. Custom silicon is additive capacity aimed at the highest volume, most predictable inference workloads. Frontier training runs and spiky new workloads stay on merchant silicon because that is where the flexibility and the software maturity are.

The practical timeline matters here. A first generation custom accelerator announced today is a 2028 or later volume story, and that assumes the first tape out works. Between now and then, Anthropic buys more accelerators, not fewer, because demand for Claude is growing faster than any internal program can absorb.

What does this do to Nvidia stock?

In the near term, very little on the fundamentals. Nvidia is supply constrained, not demand constrained, and every hour a custom chip is not shipping is an hour Nvidia is still the default. The risk is not this quarter. The risk is the terminal value argument.

The bear case that custom silicon supports is narrow but real. If the four or five largest buyers of accelerators each carve off their steady state inference volume into internal designs, the Nvidia revenue mix shifts toward training and toward the long tail of smaller customers. That is still an enormous business, but it is a lower share of a growing pie, and the market pays a different multiple for that.

Our view is that the moat is CUDA and the software stack, not any single chip generation. Custom accelerators consistently win on narrow workloads and consistently underdeliver on everything else. Watch the software, not the silicon.

Is this good or bad for AMD?

Mixed, and the read is more interesting than the Nvidia one. AMD was named as a continuing supplier, which is worth something to a company still proving it belongs in frontier AI conversations. Being on the list at all is validation.

But AMD is the merchant vendor with the least software lock-in, which makes it the most substitutable once internal silicon comes online. The AMD path is to keep winning inference sockets on price and memory bandwidth while ROCm closes the gap. This announcement raises the urgency on that clock.

What about Samsung as the manufacturing partner?

If Samsung Foundry lands the Anthropic account, that is the most consequential detail in the whole story and it has nothing to do with AI models. Samsung has spent years losing leading edge logic share to TSMC. A named frontier AI customer is exactly the reference win that reverses a foundry narrative, because foundry customers follow other customers.

Treat it as unconfirmed until there is a signed agreement. Talks are not contracts. But note the strategic logic. Every hyperscaler and lab designing custom silicon wants a credible second source to TSMC, and there are only two candidates on the planet.

Who actually captures the value here?

Follow the physics, not the press release. Custom silicon does not reduce the amount of high bandwidth memory required, the amount of advanced packaging required, or the amount of power and cooling required. It changes whose logo is on the die.

That points at the layer investors underweight. Memory makers, advanced packaging capacity, power generation, and electrical equipment get paid whether the accelerator says Nvidia or Anthropic on it. Toll roads on commerce beat bets on which brand wins a race.

How does this fit a value investing framework?

It does not, if you are trying to buy the winner of the chip design war. Nobody can underwrite a 2029 silicon roadmap with any honesty, and a business whose competitive position resets every eighteen months is not a business you can value with confidence.

What is underwritable is the constraint. Power, grid equipment, cooling, memory, and packaging are capital heavy, slow to add, and already sold out. Those businesses have pricing power because supply cannot respond quickly. We buy the bottleneck at a sane price and let other people argue about architectures.

The discipline is the same as always. Buy a durable franchise when the market is worried about the wrong thing, size the position so a bad year does not force a sale, and then do very little.

What does this say about the Anthropic IPO story?

It reinforces the capital intensity argument. A company designing its own accelerators is telling you its infrastructure spend is going up, not down, and that it intends to control its own cost curve before a public listing rather than after. That is the correct order of operations, and it supports a premium revenue multiple only if the efficiency gains show up in reported gross margin.

The number to watch is not the valuation headline. It is cost per token over the next four quarters. If that line bends, the program is working and the multiple is defensible. If it does not, the market will eventually charge Anthropic for the capex.

The bottom line

Anthropic building custom AI chips is a margin decision dressed as a hardware announcement. Nvidia and AMD keep shipping into a shortage for years. Samsung has the most to gain if the talks convert. And the quiet winners are memory, packaging, and power, because none of that gets cheaper no matter whose name is on the chip.

Commentary and opinion only. Nothing here is financial advice.