What chip is Anthropic developing with Samsung?

Details are scarce—Anthropic reportedly hasn't finalized the chip's purpose, server role, or performance targets. The conversations are exploratory at this stage.

How does this compare to OpenAI's custom chip?

OpenAI recently announced a custom inference chip called 'Jalapeño,' built in partnership with Broadcom, claiming better performance-per-watt than competing chips. Anthropic's effort appears to be earlier in development.

Why are AI companies building their own chips?

Custom silicon allows AI labs to optimize hardware for specific workloads like inference, potentially lowering cost-per-token and reducing reliance on Nvidia's expensive and supply-constrained GPUs.

Is Samsung a credible chip manufacturing partner for AI?

Samsung is already a key supplier in Nvidia's hardware ecosystem and is co-developing an AI chip factory in South Korea with Nvidia, though it faces stiff competition from TSMC in advanced node fabrication.

Home›AI Hardware›Article

AI Hardware

Anthropic Eyes Samsung for Custom AI Chip—But Doesn't Know What It Wants Yet

Anthropic is in early discussions with Samsung to co-develop a custom AI chip, but reportedly hasn't decided on the chip's purpose, power, or architecture yet. The move echoes a broader industry push to reduce dependence on Nvidia.

Dispatch desk·July 3, 2026·3 min read

TL;DR

Anthropic is talking to Samsung about a custom AI chip, but the spec sheet is apparently still blank—making this an early-stage hedge against Nvidia dependency rather than a concrete product announcement.

What changed

Anthropic opened chip development discussions with Samsung, following OpenAI's announcement of its own custom inference chip built with Broadcom.

Why it matters

As AI labs race to cut compute costs and reduce Nvidia leverage, custom silicon is becoming a strategic necessity—but the engineering complexity and multi-year timelines mean most announcements are promises, not products.

Editorial record

Drafted by autopilot. Reviewed when marked.

Source

There's a certain pattern to how AI companies talk about custom silicon. First comes the vague announcement that they're "exploring" chip development. Then comes the partner reveal. Then, eventually, the product—if the project survives the inevitable cost reckoning. Anthropic appears to be somewhere between step one and step two right now, which is either exciting or concerning depending on how many of these cycles you've watched play out.

What We Actually Know (Which Isn't Much)

According to a report from The Information, Anthropic has opened conversations with Samsung about manufacturing a custom AI chip. The catch? Anthropic reportedly hasn't nailed down what the chip will actually do, how it will slot into a server architecture, or how much raw compute it needs to deliver. That's... a lot of open questions for a partnership announcement.

To be fair, this is how chip development often starts—with a hardware partner conversation running parallel to the product spec work. But it also means we should pump the brakes on reading too much into this. "Talking to Samsung" is a long, expensive road away from "shipping a chip that meaningfully reduces inference costs."

When TechCrunch asked Anthropic about it directly, the company confirmed its existing hardware strategy—a diversified stack pulling from Google, Amazon, and Nvidia—and said nothing further about Samsung. Which is PR-speak for: yes, there's something happening here, but we're not ready to commit to it publicly.

Why Nvidia Independence Is the Real Subtext

The broader context here isn't really about Samsung. It's about the fact that every major AI lab on Earth is quietly (or not so quietly) trying to reduce its dependence on Nvidia. Nvidia's H100s and B200s are extraordinary pieces of hardware, but they come with extraordinary price tags, constrained supply, and a vendor relationship that gives Nvidia enormous leverage over the labs that need them most.

Custom silicon offers a way out—or at least a hedge. The promise is hardware purpose-built for specific workloads: inference rather than training, for example, or attention-heavy transformer architectures that don't need general-purpose GPU flexibility. When you stop paying for capabilities you don't use, cost-per-token can drop substantially. The tricky part is actually getting there.

Amazon's Trainium and Google's TPUs are the mature examples of this working at scale. Both took years and billions of dollars to become genuinely competitive alternatives. OpenAI just announced its own inference chip, called "Jalapeño," built in partnership with Broadcom—and OpenAI is claiming better performance-per-watt than competing chips. Whether that claim holds up outside of carefully controlled benchmarks remains to be seen.

The Samsung Angle Makes Sense—With Caveats

Samsung isn't a random pick here. The company is already deeply embedded in Nvidia's supply chain, producing memory and packaging components critical to Nvidia's AI accelerators. Samsung and Nvidia are also jointly developing an AI chip manufacturing facility in South Korea. Separately, Samsung has had exploratory conversations with Google about chip production partnerships.

So Samsung has the relationships, the fabrication capacity, and the motivation to land a deal with a major AI lab. The question is whether Samsung's foundry business—which has faced real yield and process node challenges compared to TSMC—can deliver the performance and reliability Anthropic would need for production inference workloads. That's not a knock on Samsung; it's just an honest accounting of where the foundry competition currently stands.

The Competitive Pressure Is Real

It's hard not to notice the timing here. OpenAI's Broadcom chip announcement landed last week. Anthropic's Samsung conversations surfaced this week. These companies watch each other obsessively, and custom silicon has clearly become a strategic checkbox—partly for genuine infrastructure reasons, and partly for the narrative value of not being purely dependent on third-party hardware.

The risk is that Anthropic ends up spending significant engineering resources on a chip program that takes three to five years to bear fruit, while its existing compute stack (Google TPUs via the Amazon and Google partnerships) is already reasonably competitive. Custom silicon is genuinely valuable at scale. But it's also a massive distraction if you don't have the chip design talent and supply chain expertise to execute.

Right now, Anthropic is asking the right questions—it's just not clear it has the answers yet. Check back in about two years, when the real story starts.