Anthropic, the company behind Claude, entered preliminary discussions with Samsung Electronics on Thursday to manufacture a custom AI chip — a move that positions the $965 billion AI company as the latest major lab to break away from near-total dependence on Nvidia, and that could hand Samsung the marquee foundry client it needs to challenge TSMC at the leading edge of chip production.

The talks, first reported by The Information, are exploratory. Anthropic has not yet decided what the chip will be optimized for, how powerful it should be, or how it will slot into a server rack. Three sources familiar with the matter told The Information that Anthropic is still defining the processor's specifications, power requirements, and cluster configurations. No physical prototypes have been built and no manufacturing timeline exists.

What has already happened is more telling than what hasn't. Anthropic hired Clive Chan — the second engineer ever to join OpenAI's dedicated custom chip team — in early June, before The Information's report surfaced. Chan spent two and a half years at OpenAI building the Broadcom-designed inference accelerator that OpenAI unveiled as "Jalapeño" on June 24, and arrived at Anthropic having learned exactly what it takes to design an AI accelerator from the software layer up.

Why Samsung and Why Now

Samsung's appeal to Anthropic has two distinct dimensions: financial alignment and manufacturing capability.

On the financial side, Samsung participated in Anthropic's $65 billion Series H round in May 2026 — alongside SK Hynix and Micron — as a strategic infrastructure partner. That relationship created a natural opening for deeper collaboration. But Samsung stands apart from the other two participants in one critical way: it is the only investor in that round that actually operates a foundry. SK Hynix and Micron manufacture memory chips; Samsung's foundry division manufactures other companies' chip designs in its own fabrication plants. That capability makes it a plausible manufacturing partner in a way the others are not.

On the manufacturing side, Anthropic's discussions center on Samsung's 2nm process node and its advanced packaging facilities. The 2nm designation — formally called Samsung's SF2 process — refers to a generation of transistor architecture that uses Gate-All-Around (GAA) nanosheet transistors, a structural shift from the FinFET transistors that dominated the previous decade. GAA transistors allow the gate to fully surround the channel from all four sides, providing tighter electrical control and enabling either 15% better performance at the same power or significant power savings compared with the prior generation. Samsung entered volume production of its SF2 process in late 2025.

Samsung's packaging division is separately relevant. Modern high-performance AI chips rarely consist of a single die. They combine multiple semiconductor components — the core logic, memory interfaces, and networking silicon — in a single package using advanced techniques such as 2.5D and 3D stacking. Samsung's packaging business specializes in precisely this kind of heterogeneous integration, and the spec discussions Anthropic is conducting reportedly include both the manufacturing process and the packaging architecture.

Why Inference, Not Training, Is the Real Prize

The technical framing matters here. Anthropic's chip conversations are not primarily about training — the resource-intensive process of teaching a model on massive datasets, where Nvidia's GPUs remain dominant and no credible near-term alternative exists at scale. They are about inference: the continuous, real-time process of serving Claude's responses to millions of users per day.

Inference and training chips have fundamentally different engineering priorities. A training chip must sustain extremely high throughput across enormous datasets, with precision requirements that favor GPU-style flexible architectures. An inference chip must deliver fast responses at low latency and low cost per query, which means eliminating the overhead of general-purpose computation in favor of hardware that runs specific transformer operations as close as possible to theoretical peak efficiency.

OpenAI's Jalapeño, built with Broadcom and unveiled June 24, was purpose-designed for inference. Broadcom CEO Hock Tan told Bloomberg the early testing showed cost savings of roughly 50% compared with typical AI GPUs for inference workloads. That figure — if it holds at production scale — illustrates why the economics of custom inference silicon are so compelling at frontier AI scale. Anthropic, running Claude for millions of users across enterprise and consumer products, faces the same inference cost structure. A chip optimized for Claude's specific model architecture and serving patterns could unlock the same kind of savings.

This is the inference chip's core logic: when a company's primary workload is a single class of operations — transformer-based language model serving — a custom ASIC can eliminate all the general-purpose overhead a GPU carries and dedicate every transistor to that one task. The efficiency gain is not marginal.

Geopolitics Woven Into Silicon

There is a structural reason beyond cost and capability that makes Samsung an attractive partner specifically. A chip designed in-house and manufactured at Samsung's fabs in South Korea — including its large-scale facility in Taylor, Texas, which is expected to begin 2nm production in 2027 — sits outside some of the most acute US-China semiconductor conflict pressure points. Samsung's fabs operate under South Korean and US jurisdiction, not subject to Chinese data-sharing obligations or the technology-transfer concerns that shadow other potential partners. For an AI company handling enterprise data at the scale Anthropic operates, that jurisdictional clarity carries weight.

Samsung is also actively expanding its semiconductor footprint. Samsung Group and SK Group announced a combined $518 billion, decade-long investment to build four new chip manufacturing facilities in South Korea — a commitment that gives Samsung credible capacity expansion to support large anchor customers.

The OpenAI Opening

Timing matters in ways that go beyond Anthropic's own strategic calendar. Samsung had been developing a custom AI chip for OpenAI — an ARM-based inference neural processing unit — before those talks stalled in early June 2026 due to what Korean media described as strategic differences between the two companies. OpenAI CEO Sam Altman subsequently cancelled a planned visit to Seoul that had been expected to advance the relationship.

If Samsung redirects the engineering attention and 2nm capacity that had been pointed at the OpenAI project toward Anthropic instead, the competitive implications compound. Anthropic would gain a foundry partner with recent, directly applicable experience in AI inference chip design. Samsung would gain a marquee client at a critical moment in its foundry business's effort to close the gap with TSMC.

The Yield Question Samsung Cannot Avoid

Samsung's 2nm ambitions have a documented constraint: yield. Yield is the percentage of chips on a wafer that pass quality testing; low yields mean high per-chip costs and supply unpredictability. Samsung's first-generation SF2 process was reported at yields in the 50-60% range through much of 2025, well below the 70-80% benchmark analysts consider economically viable for high-volume production. TSMC's rival N2 process has reportedly reached yields near 65-80% and entered volume production with anchor customers including Apple and Nvidia.

Samsung's SF2P node — the performance-optimized second iteration of its 2nm roadmap — has made progress, with some reports placing yields near 70% as of early 2026. Whether that improvement is stable at high volume is unproven. For Anthropic, which cannot afford chips that don't arrive consistently at a predictable cost, that uncertainty is a material consideration.

The discussions with Samsung are not exclusive. Anthropic is also in conversations about using chips from Microsoft and from UK startup Fractile. Google is separately in talks with Samsung about contributing to future Tensor Processing Unit production.

Industry Context: Nvidia Holds, but the Walls Are Moving

The strategic context Anthropic is navigating is an AI chip market still dominated by Nvidia. The Information's own estimate places Nvidia's share of the AI chip market at 74% — higher than it was before the custom silicon race intensified, because demand for AI training infrastructure has grown faster than alternatives have matured. Nvidia is not losing ground in absolute terms; its competitors are simply growing faster than Nvidia's growth alone can supply.

The infrastructure toll of running AI at this scale is visible in sustainability data. Google's total carbon emissions rose 25% and Amazon's rose 16% in 2025, driven significantly by data center expansion and hardware manufacturing for AI workloads, according to sustainability reports both companies released Thursday. Chips optimized for AI inference — whether Jalapeño at OpenAI or a future Claude chip at Anthropic — represent one path toward serving more users on less power, a goal that is becoming as much a regulatory and environmental imperative as an economic one.

Anthropic declined to comment on the Samsung discussions beyond a statement to The Information: "Amazon Web Services's Trainium chip, Google Tensor Processing Units and Nvidia graphic processors will remain central to how the company scales its compute strategy." Samsung also declined to comment.

The statement is technically compatible with a parallel custom chip program — Anthropic's existing partnerships represent the compute layer the company runs on today. A custom chip is what it might run on in three to five years, if the Samsung conversations mature into engineering, and engineering into silicon.

Frequently Asked Questions

What is Anthropic's custom AI chip for, and what stage is it at?

Anthropic's chip discussions with Samsung are at an early exploratory stage. The company has not yet determined what the chip will do, how powerful it will be, or how it will fit into a server. What is known is that Anthropic is evaluating Samsung's 2nm manufacturing process and its advanced packaging facilities — both relevant to a high-performance custom inference accelerator — and that it hired Clive Chan from OpenAI's chip team in early June 2026, the second hardware engineer to join that team.

Why would a custom AI chip reduce Anthropic's costs?

Custom inference chips — designed specifically for the mathematical operations that power large language model responses — can eliminate the general-purpose overhead that Nvidia GPUs carry for workloads beyond AI. OpenAI's Jalapeño inference chip, unveiled June 24, 2026, reportedly showed roughly 50% cost savings compared with standard GPU inference in early testing. A chip tuned to Claude's specific model architecture and serving patterns could deliver comparable efficiency gains for Anthropic.

What is Samsung's 2nm process, and is it ready?

Samsung's 2nm process, called SF2, uses Gate-All-Around nanosheet transistors — a major architectural advance from the FinFET transistors of the prior generation — that provide better power and performance characteristics. Samsung entered volume production on SF2 in late 2025, but its first-generation node has faced yield challenges that TSMC's rival N2 process has reportedly navigated better. The SF2P node, the performance-optimized second iteration, is reportedly approaching 70% yield as of early 2026, but high-volume production stability at that rate remains to be demonstrated at commercial scale.

Will Anthropic stop using Nvidia, Google, and Amazon chips?

No, based on available information. Anthropic told The Information that AWS Trainium, Google Tensor Processing Units, and Nvidia GPUs will remain central to its compute strategy. A custom chip, if it materializes, would add a new layer to that stack rather than replace it — likely handling specific Claude inference workloads at scale while the existing partnerships cover training and other workloads where Nvidia's ecosystem remains unmatched.