Anthropic Taps Samsung to Build Its First Custom AI Chip

Welcome back to SavvyMonk, your daily dose of AI and tech news that actually matters.

Anthropic has spent the past year insisting it was content to rent its compute from everyone else. A new report says it is now trying to build some of its own, and it has picked an unexpected partner to do it.

Let's get into it.

Anthropic has begun early-stage work on its own artificial intelligence chip and has held talks with Samsung Electronics about manufacturing it, according to a report published on 2 July. The discussions are early enough that the company has not settled the basics, including what the chip will be used for, how powerful it should be, or how it would fit inside a server, and there is a real chance the effort never reaches silicon at all.

The Backstory

This is not the first sign that Anthropic wanted more control over its hardware. Reporting from April indicated the company was weighing whether to produce its own chips as a way to cope with the shortages squeezing every large AI lab. The Samsung conversations mark the point where that idea stops being a thought experiment and starts looking like a plan, even if a distant one.

The clearest evidence that Anthropic is serious is who it has been hiring. The company recently brought on Clive Chan, an early member of the team that designed OpenAI's own custom accelerator, and someone who spent more than two years learning what it takes to build an inference chip from the software layer up.

Hiring that kind of talent is expensive and deliberate, and it tends to signal intent long before any public announcement does.

Why Samsung

Samsung is an unusual choice on the surface, because the obvious foundry for a cutting-edge AI chip is Taiwan Semiconductor. The appeal comes down to two things. Anthropic is reportedly eyeing Samsung's 2-nanometre process and its advanced packaging facilities, the pieces of the puzzle that determine how efficient a finished chip can be.

And the relationship already runs deeper than a supplier arrangement, because Samsung put money into Anthropic's sixty-five billion dollar funding round in May, alongside SK Hynix and Micron, which gives both sides a reason to turn a financial tie into a manufacturing one.

The talks are not exclusive, which is worth holding onto. Anthropic is also in discussions to use chips from Microsoft and from the UK startup Fractile, and Google is separately talking to Samsung about producing part of a future tensor processing unit. What looks like one bet is really Anthropic spreading several at once.

The OpenAI Shadow

None of this is happening in a vacuum. Eight days before the Samsung story surfaced, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference processor, built in a startlingly fast nine-month cycle and claimed to deliver far better performance per watt than existing hardware.

The timing of Anthropic's move is hard to read as coincidence. Its closest rival had shipped a piece of the future that Anthropic did not yet own, and the reporting on the Samsung talks reads in part as an answer.

It also fits a much wider pattern. Google, Amazon, Meta and Microsoft have all spent years building proprietary silicon to run their own workloads, both to tune the hardware to specific tasks and to stop writing enormous cheques to a single supplier. Anthropic joining that group is less a surprise than a matter of when.

The Nvidia Problem

The supplier everyone is trying to route around is, of course, Nvidia. Even after a year of custom-chip announcements, one estimate puts Nvidia's share of the AI chip market at roughly 74 per cent, which is higher than it was before the arms race began. That statistic captures the strange shape of this moment.

Demand for AI compute is growing so fast that Nvidia is not actually losing ground in absolute terms, and its rivals are simply trying to grow faster than Nvidia alone can supply.

For a company like Anthropic the logic is about cost and control as much as capacity. Designing a chip around the exact way its models run means paying less to serve every answer, and owning part of the supply chain means being less exposed to the next shortage or price rise. The catch is that the cost of designing a single advanced chip runs to hundreds of millions of dollars, and the payoff, if it comes, sits years away.

What Anthropic Actually Says

Publicly, Anthropic is playing this down, and its statement is careful. The company pointed to its existing stack and said that Amazon Web Services' Trainium chips, Google's tensor processing units and Nvidia's graphics processors will remain central to how it scales, while declining to add anything about Samsung.

That statement is technically compatible with a parallel chip effort, because the existing partnerships are the compute layer Anthropic depends on, while a custom chip is what it might run on in three to five years, if the talks mature into engineering and engineering into a working part.

The Bottom Line

A custom chip is not a break from Nvidia so much as a hedge against depending on it forever, and Anthropic is late enough to this game that not moving would have been the bigger risk.

The Samsung talks are early, unglamorous and might lead nowhere, but the Clive Chan hire tells you the intent is real. Watch whether these conversations turn into an actual design, because that is the moment this stops being a story about ambition and becomes one about execution.

Category: Competitive Analysis

"Act as a strategy analyst. I will give you a company and a big infrastructure or supply-chain move it is making, such as [company] building its own [component]. Break down the real motive behind it, separating what the company says publicly from what the move actually achieves.

Then map who it threatens, who it helps, and the single condition that would have to be true for the bet to pay off. Finish with the one metric I should track to know whether it is working."

For years the AI race was a contest over models, and the winner was whoever trained the smartest system. That contest is steadily becoming a contest over hardware, where the edge goes to whoever can run those systems most cheaply at scale.

A chip Anthropic has not even designed yet is a reminder that the fight is moving down the stack, closer to the physics, and the labs that own that layer will set the terms for everyone renting it.

Hit reply, I read every response.

See you tomorrow.

— Vivek

P.S. Know someone who tracks the business of AI as closely as the models themselves? Forward this to them. They can subscribe at https://savvymonk.beehiiv.com/