Anthropic wants to own the denominator

Anthropic's early talks with Samsung Foundry are not, primarily, an engineering story. They are a margin story dressed for an IPO roadshow. The company is trying to own the denominator of its cost-per-token equation before public investors price that denominator in for it.

The reporting, briefly. The Information reported on 2 July that Anthropic has entered preliminary discussions with Samsung Electronics to co-develop a custom AI accelerator on Samsung Foundry's 2-nanometre process, with advanced packaging. No signed agreement, no confirmed architecture, no timeline. TechCrunch carried the same story the same day. Anthropic filed its confidential S-1 with the SEC a month earlier, on 1 June.

Those two facts belong in the same paragraph, and the market has mostly been reading them in separate ones.

The pattern is not new. Google shipped its first TPU (Tensor Processing Unit, Google's in-house AI accelerator) in 2016. Amazon's Trainium is in production for AWS customers and, notably, is part of the commercial furniture of Anthropic's own $4bn relationship with Amazon. Meta has its MTIA line. The frontier labs and hyperscalers that could afford to build silicon have built it. What has changed is the price of not building it.

Why this matters for a company about to list. Public market investors will value Anthropic on some multiple of forward revenue, but they will discount hard for gross margin trajectory. Every token Anthropic serves on rented Nvidia silicon is a token whose margin is set by Jensen Huang's pricing committee. That is a fine arrangement for a private company burning venture capital. It is a much less fine arrangement for a company asking public investors to underwrite a decade of compounding inference volume.

Google and Amazon did not build custom accelerators because their engineers were bored. They built them because GPU rental economics were compressing cloud margins in a way that showed up on quarterly earnings calls. Anthropic is now looking at the same maths, one balance sheet removed.

Samsung's side of the table is the more interesting half. Samsung Foundry trails TSMC on advanced-node yield in a way that is well-understood by everyone who buys leading-edge silicon; Apple, Nvidia and AMD route their most demanding designs through TSMC for good reason. Samsung needs a marquee AI logic customer on 2nm to validate the process to the rest of the market. Anthropic needs optionality against both Nvidia and its own Amazon-Trainium dependency.

Each party is, to some extent, using the other as a proof point. That is not a criticism. It is how these deals get done. It is also why "early talks" should be read as exactly that.

The contrarian read deserves airtime, because it is probably half-right. Preliminary discussions between a pre-IPO lab and a foundry hungry for customers are not, by themselves, evidence of a coherent silicon strategy. TPU took Google roughly a decade of iteration to reach the point where it materially displaced Nvidia inside Google's own workloads. Trainium has taken Amazon multiple chip generations to get to production reliability. Anthropic has no disclosed silicon engineering organisation of that depth.

The realistic near-term output of talks like these is often a preferred-customer or co-design arrangement, not a ground-up bespoke accelerator. The gap between the strategic framing ("Anthropic goes custom") and the operational reality ("Anthropic secures capacity and IP optionality on an advanced node") is wide enough to drive an S-1 through.

The sovereignty layer changes the calculation, though, and this is where I part company with the pure margin read. The Trump administration's export restrictions on Anthropic's Fable 5 weights in 2026 turned the dependency chain from an accounting question into a policy one. If a government can restrict your weights, the composition of your compute stack is no longer just a make-versus-buy problem. It is a question of which jurisdictions can, in principle, turn you off.

Samsung is a South Korean company with US CHIPS Act-aligned facilities and an increasingly serious presence in Texas. TSMC is Taiwan-based with the geopolitical exposure that entails. Reading the Samsung talks purely as a cost play misses that the choice of foundry is now itself part of the risk disclosure Anthropic's underwriters will have to write.

What the talks do not do. They do not, on any credible timeline, pressure Nvidia's 2027 pricing. Custom accelerators, when they arrive, tend to eat inference workloads first and leave training on Nvidia — the most capital-intensive part of the stack. Google and Amazon still buy Nvidia in volume alongside their own silicon. Anthropic will too, for the foreseeable future.

The aggregate signal matters more than any single deal. Every lab that enters foundry talks is a lab that is not committing to more H200s at list price in the out-years. That is a slow-moving pricing pressure, not a step change, and it is priced into Nvidia's stock only in the loosest sense.

What to watch. Three things. Whether the talks produce a named chip programme with a disclosed tape-out target, or remain in the "sources familiar with the matter" register through the S-1 amendment cycle. Whether Anthropic's S-1 risk factors mention supply concentration or foundry dependency in language that reads as prepared, rather than boilerplate. And whether Amazon's posture shifts — a lead investor watching its portfolio company court a rival's foundry partner is not a neutral observer, and the tension between the Trainium relationship and the Samsung talks is real.

The chip may or may not exist in three years. The pressure that made Anthropic pick up the phone to Samsung exists now, and it is the more durable fact.

Footnotes