Dario Amodei's Samsung chip talks put Anthropic deeper into the compute business

Why it matters

Anthropic's Samsung talks show the compute race moving below the model layer. If Claude demand keeps rising, frontier AI labs will need leverage over chips, memory, power and server design, not just better software.

Dario Amodei and Daniela Amodei are pushing Anthropic closer to custom silicon, with the Claude maker in talks with Samsung Electronics about manufacturing a custom AI processor, according to Bloomberg Technology, which cited The Information and people familiar with the plan.

The talks, reported July 2, remain early. TechCrunch also summarized The Information's reporting and said Anthropic has not decided what the chip would be used for, how it would fit into a server, or how powerful it should be. Anthropic told TechCrunch that a diversified hardware stack including Google, Amazon and Nvidia chips will remain central to its compute strategy.

That makes the Samsung discussion a sourcing and architecture move rather than a product launch. There is no disclosed contract, production timeline, process node, server design, performance target, order volume or contract value. The useful read is narrower and more important: Anthropic is studying whether the next stage of model competition requires it to shape the hardware beneath Claude, rather than simply buy whatever accelerators the cloud market can make available.

That is a natural turn for the Amodeis. Dario Amodei, a former OpenAI vice president of research who helped lead work on GPT-2 and GPT-3 after time at Google Brain, built Anthropic around reliable, interpretable and steerable AI systems. Daniela Amodei, Anthropic's co-founder and president, came through politics, Stripe and OpenAI, where she became vice president of safety and policy. Their company has always framed control as a safety question. The Samsung talks put a physical version of that same idea on the table: control over supply, cost and performance at the chip and server layer.

Claude demand is forcing a hardware conversation

Anthropic's compute needs have moved far beyond the scale implied by its 2021 origin story. The company announced a $124 million Series A in May 2021 to build more reliable general AI systems. In February 2026, it said it had raised a $30 billion Series G at a $380 billion post-money valuation. In May 2026, it announced a $65 billion Series H at a $965 billion post-money valuation, led by Altimeter, Dragoneer, Greenoaks and Sequoia.

Those company-reported numbers describe the pressure behind the chip talks. Anthropic said in May that run-rate revenue had crossed $47 billion earlier that month, up from the $14 billion run-rate revenue it reported in February. It also said Claude Code run-rate revenue was over $2.5 billion, more than 500 customers were each spending over $1 million annualized, and eight of the Fortune 10 were Claude customers. The figures are Anthropic's own, and the company did not publish the underlying customer contracts or gross margin profile behind them.

The May funding announcement also showed how much of Anthropic's growth plan already depends on hardware partners. Anthropic said the Series H included $15 billion of previously committed hyperscaler investments, including $5 billion from Amazon. It said it had signed agreements with Amazon for up to five gigawatts of new capacity, with Google and Broadcom for five gigawatts of next-generation TPU capacity, and with SpaceX for access to GPU capacity in Colossus 1 and Colossus 2. Anthropic also named Micron, Samsung and SK hynix as strategic infrastructure partners whose memory, storage and logic technologies matter to its compute buildout.

A Samsung-manufactured custom chip would fit into that pattern if it advances beyond talks. It would give Anthropic another path through a market where frontier labs compete for the same Nvidia GPUs, hyperscaler capacity, power contracts, memory supply and packaging capacity. Anthropic's own position, as relayed to TechCrunch, is that Google TPUs, AWS Trainium and Nvidia GPUs remain part of the plan. The company is looking for optionality, not declaring independence from existing suppliers.

Samsung would get a customer it can talk about

Samsung has its own reason to pursue this kind of work. Its semiconductor leadership page lists Young Hyun Jun as vice chairman and CEO, head of the Device Solutions division and head of the Memory Business, with responsibility across memory, System LSI and foundry units. Jinman Han, president and head of Samsung Foundry, is described by Samsung as focused on process technology and key customer relationships.

A marquee frontier-model customer would be strategically useful for Samsung Foundry. Advanced AI accelerator business has tended to concentrate around a small number of manufacturing and packaging paths, and any Anthropic project would give Samsung a chance to attach its foundry pitch to one of the most heavily financed AI labs in the market. That said, the current reporting does not establish that Anthropic has selected Samsung, committed to a Samsung node, or designed a chip around Samsung's manufacturing stack.

The same caveat matters for Anthropic. A custom chip program can absorb years of engineering time before it delivers usable economics. The chip still needs a workload target. Training and inference make different demands. Server integration can decide whether a chip is useful at data center scale. Memory bandwidth, networking and packaging can matter as much as raw compute. The absence of those details in the current reporting is the story's main constraint.

The founder bet is control

Anthropic's public posture has been unusually explicit about safety, policy and controlled deployment, even while it races OpenAI, Google DeepMind, Meta, xAI, Mistral and other model developers for enterprise adoption. That tension has sharpened in recent weeks. RuntimeWire reported on July 2 that Dario Amodei's Pentagon emails showed why Anthropic's military AI fight keeps escalating, with court records tracing disputes over Claude use in autonomous weapons and domestic surveillance. The same day, RuntimeWire reported that Anthropic's Fable 5 redeployment faced a same-day cyber safety test after access was restored on July 1.

Those fights sit on one side of Anthropic's operating problem: how to make frontier systems available without losing control over use. The Samsung talks sit on the other side: how to supply those systems without relying entirely on other companies' roadmaps and capacity decisions. Both problems get harder as Claude usage grows.

Anthropic's November 2025 infrastructure announcement made the scale clear. The company said it would invest $50 billion in American AI infrastructure. It also said it served more than 300,000 business customers.

That kind of demand turns chip strategy into founder strategy. The Amodeis can keep saying Anthropic is an AI safety and research company, and Anthropic's company page still says it builds reliable, interpretable and steerable systems. But the business they now run has the infrastructure profile of a hyperscaler customer, a model lab, an enterprise software vendor and a national industrial project at once.

Samsung is only a possible manufacturing partner at this stage. The chip may never reach production. The use case may change before it leaves planning. Still, the talks mark another step in Anthropic's shift from renting frontier compute to designing around it. For a company whose founding thesis was that powerful AI should be steered, silicon is becoming part of the steering wheel.