Skip to content
All posts
ai-briefingcustom-siliconchipsanthropic

Anthropic Is Building Its Own Chips Now. Every Frontier Lab Is a Silicon Company.

Anthropic confirmed this week it is staffing a custom silicon team to co-design chips with Claude, the clearest sign yet that renting Nvidia is no longer a complete strategy.

The last pure-play AI lab just stopped being pure-play.

On August 5, Anthropic confirmed to TechCrunch that it is assembling a "custom silicon team" to design its own AI chips, co-designed alongside future Claude models. Business Insider first spotted the move in job listings; one posting for chip engineers offered compensation up to $485,000. Anthropic's stated goal is to cut per-token inference costs by tuning hardware directly to Claude's computation patterns.

The confirmation lands on top of reporting from The Information in July that Anthropic has held exploratory talks with Samsung about manufacturing on its 2nm process with advanced packaging. Those talks are early: no agreement, no finalized design, and per The Information's reporting the company has not even settled what the chip should do. Anthropic also hired Clive Chan, an early member of OpenAI's custom chip team, which suggests the intent predates the job listings. Samsung is not a stranger here. It participated in Anthropic's $65 billion Series H round in May, alongside SK Hynix and Micron.

Context matters. OpenAI unveiled Jalapeño, a Broadcom-built inference ASIC, on June 24, with a nine-month development cycle that people in the industry described as the fastest high-performance ASIC program they had seen. Google has run TPUs for a decade and is deploying roughly 36,000 Ironwood racks this year, per industry coverage of its 2026 buildout. Meta has MTIA. Amazon has Trainium. Anthropic was the last frontier lab without a silicon story. Now it has one.

For practitioners, the significance is not that another chip exists. It is that the compute procurement playbook is changing underneath us. Anthropic already buys across four silicon families: Nvidia GPUs, Google TPUs (a commitment reported at around one million chips), AWS Trainium, and AMD. Its own statement on the news was careful: Trainium, TPUs, and Nvidia GPUs "will remain central to our compute strategy." Custom silicon, in this framing, is a fifth option and a negotiating lever, not a replacement.

The case for

The strongest argument is arithmetic, and Anthropic made it itself: inference cost per token is the line item that decides whether an AI business scales profitably, and a chip shaped around one model family's compute graph should beat a general-purpose GPU on that metric. Google proved the pattern. A decade of TPU investment means Alphabet now trains and serves frontier models without paying Nvidia's 75-percent gross margins on every FLOP, and Google's willingness to rent TPUs to Anthropic at scale shows the economics work for tenants too.

The market data supports the shift. One industry tracker, Presenc AI, estimates Nvidia's share of data center AI accelerator units has fallen from roughly 95 percent in 2022 to somewhere between 62 and 66 percent in 2026, even as Nvidia keeps most of the revenue dollars. Custom ASIC shipments are projected by industry analysts to grow about 44.6 percent in 2026 against 16.1 percent for GPUs. The unit migration is real.

Capital is not the constraint either. Quartz reported this month that Blackstone is sounding out investors on a second mega debt package, one proposal set at $36 billion, to finance more of Anthropic's Google TPU usage, structured like the earlier $35 billion Apollo and Blackstone deal where a special-purpose vehicle buys chips and leases them back. When Wall Street will pre-fund your silicon appetite twice before your IPO, the build-versus-rent question answers itself.

The case against

Start with the clock. Custom silicon typically takes years from architecture to volume production, and OpenAI's nine-month Jalapeño cycle was the exception that leaned on Broadcom's deep ASIC bench. Anthropic has a handful of job postings and early Samsung talks. Samsung's foundry, meanwhile, has struggled with yields at advanced nodes compared to TSMC, a gap analysts have flagged repeatedly in coverage of the talks. Betting your cost curve on a partner still proving its 2nm ramp is not a free option.

Then there is the moat argument. Writing for TensorFeed in July, analyst Marcus Chen framed it bluntly: the chip is the easy part now, and Nvidia's real moat is CUDA and the roughly four million developers on top of it. Anthropic itself conceded the point in May, when it signed across four silicon families precisely because no single ASIC covers every workload.

The sharpest data point comes from Nvidia's side. At a Morgan Stanley roadshow on July 10, Jensen Huang disclosed that a representative frontier model customer that had primarily used ASICs had raised Nvidia's share of its usage to nearly 50 percent. Market consensus, per reporting of the disclosure, identifies that customer as Anthropic. Read that again. The lab most committed to diversification has been drifting back toward Nvidia, because software maturity beats spec-sheet efficiency when you are serving production traffic.

There is also timing risk. Anthropic has confidentially filed for a US IPO targeting as early as October, per Quartz. Launching a multi-year, multi-billion-dollar silicon program while asking public markets to underwrite the story is bold. Public investors have been kind to AI capex so far. So far.

What to watch

Three things. First, whether the Samsung talks convert into a signed manufacturing agreement, or whether Anthropic follows OpenAI to Broadcom, the safer pair of hands. Second, the S-1: if the IPO lands in October, the filing will put real numbers on compute commitments and show how much silicon ambition the market is being asked to fund. Third, hiring velocity on the custom silicon team. Job postings are cheap. A taped-out chip is not. The distance between the two is where this story will be decided.

Sources