Anthropic has joined the race to build its own brains for its models. On August 5, the maker of Claude confirmed to TechCrunch that it is assembling an in-house "custom silicon team" to design AI chips tailored to its own workloads — the clearest signal yet that the frontier labs no longer believe they can rent their way to enough compute.
The company said it plans to co-design hardware and models so its technology can "run faster and more efficiently," according to TechCrunch, which confirmed the effort after Business Insider first reported it. It is a modest description for what amounts to one of the most capital-intensive bets a software company can make: designing a leading-edge accelerator can cost hundreds of millions of dollars before a single chip ships.
What Anthropic confirmed
The disclosure is the first time Anthropic has publicly acknowledged a chip program. A company job listing seeks engineers across the full hardware-software stack — front-end design, pre-silicon verification, physical design, design-for-test, analog and mixed-signal, foundry and technology, design infrastructure, and advanced packaging with signal and power integrity. Salaries for the roles run from roughly $320,000 to $485,000, a sign of how fiercely the labs are competing for a small pool of experienced silicon engineers. Anthropic is reportedly hiring from rivals; among recent additions is Clive Chan, who joined from OpenAI's chip effort.
Crucially, Anthropic framed this as a multi-chip strategy rather than a break with its cloud partners. The company will keep buying and using AWS's Trainium, Google's TPUs, and GPUs from Nvidia and AMD even as it develops its own silicon. As TechCrunch put it, Anthropic "has inked deals with AWS, Google, Nvidia, and AMD to access AI computing hardware. But to really scale to meet the level of demand, relying on others clearly isn't enough."
The manufacturing question remains open. Last month The Information reported that Anthropic was scouting Samsung as a potential fabrication partner, with subsequent coverage pointing to Samsung's 2nm process and advanced packaging facilities as the target. Those talks are described as early and exploratory: the chip's purpose and specifications have not been finalized, no prototypes have been built, and there is no public manufacturing timeline.
The compute-cost squeeze
Anthropic's move only makes sense against the backdrop of its staggering infrastructure bills. The company runs Claude across a multi-cloud "AI factory," and the commitments behind it have ballooned over the past year. In October 2025, Anthropic and Google announced an expanded cloud deal worth "tens of billions of dollars," giving Anthropic access to up to 1 million TPUs and more than a gigawatt of capacity coming online in 2026. In April 2026, Anthropic, Google, and Broadcom expanded that arrangement with a multi-gigawatt deal for up to 5 gigawatts of next-generation TPU capacity beginning in 2027. On the Amazon side, Anthropic has committed to spending well into the tens of billions on AWS technologies — a relationship anchored by Amazon's roughly $33 billion-plus in total investment and multi-gigawatt Trainium buildouts.
When a company is signing compute contracts measured in gigawatts and tens of billions of dollars, even single-digit percentage gains in performance-per-watt translate into enormous savings. That is the core logic of vertical integration: a chip co-designed around Claude's specific attention patterns and inference profile can, in principle, do more work per dollar and per watt than general-purpose hardware bought off a partner's price list. Industry coverage has floated the possibility that co-designed silicon could meaningfully cut Claude's inference costs, though Anthropic has not put a number on any such target.
Custom silicon as table stakes
Anthropic is late to a party that its peers have been throwing for years. Google has run its models on Alphabet's in-house TPUs for the better part of a decade. Meta has been shipping its MTIA accelerators for internal AI workloads. And in June 2026, OpenAI unveiled its first custom chip, the Broadcom-built "Jalapeño," designed specifically for inference. Amazon (Trainium and Inferentia) and Microsoft (Maia) round out a field in which owning at least part of the silicon roadmap has shifted from an edge to a baseline expectation for any lab operating at the frontier.
The strategic reasoning is consistent across all of them. Frontier training and inference are now bottlenecked as much by chip supply, energy, and cost as by algorithms. Depending entirely on Nvidia — or on any single partner — means accepting that partner's pricing, allocation, and roadmap. Designing your own chip is a hedge against all three, and a way to capture margin that would otherwise flow to a supplier. For Anthropic, which has positioned itself as a safety-focused lab now scaling commercially at breakneck speed, custom silicon is also a way to keep unit economics from swallowing the business as usage climbs.
The risks are just as real. Chip design is slow, expensive, and unforgiving; first silicon can take years to reach production, and a misstep can burn hundreds of millions with nothing to show for it. Anthropic will be running its own program while simultaneously spending record sums with the very vendors it hopes to eventually supplement.
What to watch
The next signals to track: whether the Samsung talks harden into a formal foundry agreement and at which process node; whom else Anthropic poaches to staff the team, and how quickly it grows; and whether the company clarifies whether its first chip targets training, inference, or both. Any concrete efficiency or cost targets — and a realistic timeline for tape-out — would tell us how seriously to take this beyond a hiring page. For now, the message from Anthropic is unambiguous: in the compute era of AI, the companies building the models increasingly intend to build the chips that run them.
"Anthropic has inked deals with AWS, Google, Nvidia, and AMD to access AI computing hardware. But to really scale to meet the level of demand, relying on others clearly isn't enough."- Rebecca Bellan, Reporter, TechCrunch