Anthropic is in talks with Samsung to explore building a custom AI chip, according to TechCrunch AI, which cited a Thursday report from The Information. The move signals that a company best known for its Claude models is now serious about designing its own silicon. Anthropic first floated the idea back in April, when Reuters reported it was weighing homegrown chips as a hedge against ongoing shortages. This is the clearest sign yet that the idea has legs.
What stands out is how early this still is. Anthropic hasn’t decided what the chip would actually do, how it would fit into a server, or how powerful it would be, per the report. In other words, this is an exploration, not a finished product.
What Anthropic is saying
When TechCrunch AI reached out, Anthropic kept its cards close. The company said a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will stay central to its compute strategy. On the Samsung talks specifically, it had nothing further to add.
That non-denial is worth reading carefully. Anthropic isn’t confirming a deal, but it isn’t waving off the reporting either. The message it wants out there: no single vendor owns its future.
Why this matters
Custom silicon has become the new arms race in AI, and the reasons are practical:
- Independence from Nvidia. Nvidia is still the undisputed leader in AI chips, which means everyone else is at the mercy of its supply, pricing, and roadmap. Building your own chip loosens that grip.
- Purpose-built performance. A chip designed for one specific job, like running models rather than training them, can be tuned for efficiency in ways general-purpose hardware can’t match.
- Supply security. Shortages have been a recurring headache. Owning more of the stack means fewer surprises.
Anthropic wouldn’t be a pioneer here. Amazon and Google both already offer custom-built TPUs through their cloud businesses, and both are Anthropic investors and compute partners. So Anthropic is surrounded by companies that have already done this. Designing its own chip is the logical next step.
The OpenAI factor
The timing is hard to ignore. Just last week, OpenAI teamed up with Broadcom to unveil its own custom inference processor, nicknamed “Jalapeño.” OpenAI says the chip delivers better performance-per-watt than competing hardware, a direct pitch on efficiency.
When your closest rival announces custom silicon, staying silent isn’t a great look. Anthropic’s Samsung talks read, at least in part, as a response. The two leading model labs are now racing on hardware as well as software, and that competition tends to speed everyone up.
Why Samsung
Samsung is an interesting choice, because it’s already deep in the AI supply chain. It’s a major Nvidia partner, manufacturing chips that Nvidia needs to train and run its models. In a nice twist, Samsung uses Nvidia’s software to make those chips, and the two are building an AI chip factory in South Korea together.
Samsung has also discussed chip-making work with Google. So it brings real manufacturing muscle and existing AI relationships to any Anthropic collaboration. For a company designing its first chip, a partner who already knows the terrain is a sensible bet.
What to watch next
A few things worth tracking as this develops:
- Confirmation. Neither company has confirmed a formal partnership. Watch for an official announcement, or a pointed lack of one.
- Inference vs. training. OpenAI’s Jalapeño targets inference, the cheaper, higher-volume side of running models. If Anthropic’s chip goes the same way, that tells you where the cost pressure really is.
- The Nvidia balance. Anthropic insists Nvidia stays pivotal. How it threads custom silicon into a stack that still leans on Nvidia, Google, and Amazon will say a lot about where the industry heads.
For now, treat this as a serious signal rather than a done deal. The direction is clear: the biggest AI labs want more control over the hardware their models run on, and they’re willing to build it themselves to get it. You can find the full details in the original TechCrunch AI report.