Anthropic’s Samsung Chip Talks Expose the Foundry Race Below Nvidia

Anthropic is in early talks with Samsung Electronics to manufacture a custom AI chip, according to a July 2 report from The Information, the latest in a string of frontier labs that have decided to design their own silicon.

Amazon’s Trainium chips, Google tensor processing units, and Nvidia graphics processors will stay central to Anthropic’s compute strategy, the company told The Information, while declining to comment on the chip plans. Samsung also declined to comment. The Information reported that Anthropic is weighing Samsung’s 2-nanometer manufacturing process and its advanced packaging facilities, and that no detailed design or manufacturing work has begun. Anthropic may still choose not to proceed.

Anthropic Opens Early Talks With Samsung on a Custom AI Chip

The custom AI chip Anthropic is exploring would sit alongside, not replace, the compute strategy the company has been building. Anthropic runs Claude on Google TPUs, Amazon Trainium accelerators, and Nvidia GPUs in parallel, and expanded that footprint through 2025. A purpose-built Anthropic chip would slot into that stack as a fourth platform, designed around the model’s specific kernels and serving patterns. The diversification runs across more than chips, too: Anthropic’s earlier Microsoft Maia talks showed the same instinct to spread suppliers across AWS, Google, and now Microsoft. None of those moves dented the relationship with Nvidia, which remains central to training Claude’s largest models.

The Information noted one staffing move that signals Anthropic’s intent. Anthropic recently hired Clive Chan, an early member of OpenAI’s custom chip team, as part of what The Information described as a deliberate engineering buildout. Chan comes from the team that produced OpenAI’s first inference chip, Jalapeño. Anthropic’s chip plans remain at an early stage, per The Information.

No detailed design or manufacturing work has begun, and the company may not proceed. The direction of travel is what is firm: Anthropic is publicly shopping for a foundry partner. Samsung is on the short list.

The Three-Platform Strategy a Custom Chip Would Join

Anthropic’s existing silicon footprint is one of the most diversified in the industry. In October 2025 the company said it would expand its use of Google Cloud technologies to include up to one million TPUs, an expansion Anthropic described as worth tens of billions of dollars and expected to bring well over a gigawatt of capacity online in 2026. Anthropic remains committed to Amazon as its primary training partner and cloud provider, and continues to work on Project Rainier, a compute cluster with hundreds of thousands of AI chips spread across multiple U.S. data centers. Nvidia GPUs round out the third platform, per Anthropic’s three-platform compute strategy announcement.

That three-platform footprint is broader than the typical one- or two-supplier configuration at frontier labs. A custom Anthropic chip would slot into the same stack as an additional tool, designed around Claude’s specific kernels and serving patterns. The most recent proof point for that playbook sits at OpenAI, where a chip built around the same logic was unveiled on June 24.

Why Samsung, and Why Now

Three things shifted at Samsung over the past year that put the Korean conglomerate on Anthropic’s short list. None of them on its own is a marquee customer moment; the cumulative effect is what tipped the foundry onto the menu. Samsung is now read alongside TSMC, not in place of it.

First, yields. Samsung’s 2-nanometer gate-all-around process yield has climbed into the 55% to 60% range, per Samsung Foundry’s 2-nanometer process details in Chosun, after years of yield problems that left the foundry well behind TSMC’s leading-edge nodes. Yield is the metric that decides whether a foundry can take a marquee AI client, and Samsung has now crossed the threshold frontier labs will tolerate. That single change opened doors that were closed for most of 2024 and 2025.

• 55% to 60%: Samsung Foundry’s 2nm yield, per Chosun
• 74%: Nvidia’s share of the AI chip market, per The Information
• $8.4 billion: Broadcom AI semiconductor revenue in Q1 FY2026
• 106%: Broadcom AI revenue growth year-over-year
• $100 billion: Broadcom’s guided AI chip revenue by 2027

Second, customers. Tesla awarded Samsung a 23 trillion Korean won contract to produce its AI chips in July 2025, Qualcomm and AMD have held 2-nanometer discussions, and Samsung is also manufacturing the language processing unit for Nvidia’s next-generation Vera Rubin accelerator. The customer set now spans hyperscalers, automotive, and the incumbent AI leader.

Third, alignment. Samsung, SK Hynix, and Micron all participated in Anthropic’s $65 billion Series H funding round in late May, per Anthropic’s $65 billion Series H announcement, a round that valued Anthropic at $965 billion post-money. Samsung is now both an investor in Anthropic and a candidate foundry, a combination that shortens the distance between a phone call and a tape-out. The alignment does not guarantee a deal, but it moves Samsung up the list ahead of foundries that have only a transactional relationship.

The Design Question, and Broadcom’s Likely Role

The design half of the equation is separate from the foundry half. On that side, Broadcom is the firm with the most reach on the design side.

• AI semiconductor revenue of $8.4 billion in Q1 FY2026, up 106% year-over-year
• Fiscal 2025 AI chip sales of $20 billion, up 65%
• Guided AI chip revenue above $100 billion by 2027, with six anchor customers including Google, Meta, OpenAI, and Anthropic
• April 2026: Anthropic expanded its partnership with Broadcom for additional compute

Anthropic’s chip plans, as described by The Information, do not specify who would design the silicon. The precedent is unusually clean: OpenAI’s Jalapeño inference chip, unveiled on June 24 with Broadcom as its silicon implementation partner, was designed from initial concept to manufacturing tape-out in nine months. OpenAI and Broadcom described the cycle as the fastest ASIC development program they believe has been achieved in high-performance advanced semiconductors, and Broadcom’s Tomahawk networking silicon and packaging expertise are also part of the Jalapeño stack.

Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems.

Greg Brockman, President and Co-Founder of OpenAI, made the comment in OpenAI’s Jalapeño inference chip unveiling on June 24. The same full-stack logic is now guiding Anthropic’s chip exploration. Anthropic’s hire of Clive Chan from OpenAI’s chip program points to a similar buildout over time. For a first chip, the most likely path is the Broadcom model OpenAI already validated.

The Nvidia Question the Headlines Skip

The default read on Anthropic’s chip move is that it is aimed at Nvidia. The Information’s own numbers say otherwise. The Information’s reporting estimates Nvidia’s AI chip market share at 74%, a figure higher than it was before the inference-chip arms race began.

NVDA shares were up 0.7% in morning trade the day the report surfaced. AVGO, the design partner OpenAI chose for Jalapeño, was also up 0.7%. TSM, the foundry leader Nvidia depends on for its flagship GPUs, gained 3.5% in the same session. The market is expanding fast enough that even a smaller share of a much larger pie produces more dollars, and Anthropic’s compute buildout is a case study in that expansion.

How OpenAI Built Jalapeño, and What Anthropic Can Reuse

OpenAI’s path from idea to chip offers the closest template for what an Anthropic-built accelerator would look like. OpenAI designed Jalapeño around its deep understanding of LLM fundamentals, with Broadcom handling silicon implementation and Celestica handling board and rack integration. Engineering samples were already running machine learning workloads at production target frequency and power, including GPT-5.3-Codex-Spark. OpenAI plans initial deployment by the end of 2026, with a multi-year ramp and gigawatt-scale rollout alongside Microsoft and other data center partners. The architecture reduces data movement and balances compute, memory, and networking resources to achieve realized utilization closer to theoretical peak performance.

The Information noted one underappreciated wrinkle. Samsung Foundry is looking toward Anthropic partly because preliminary OpenAI custom-chip efforts are losing momentum. Custom silicon is genuinely hard to ship on time, and the people who can do it are scarce.

Anthropic has started early, has hired at least one OpenAI chip veteran in Clive Chan, and has the financial cushion of a $65 billion Series H round to absorb a multi-year design cycle. The piece still missing is a timeline: The Information reported no design work has begun and that the project may not proceed at all. What has begun is the conversation. And in custom silicon, the conversation is the part that takes longest to start.

Disclaimer: This article references specific public companies and market figures. It is informational and does not constitute investment advice. Chip stocks and AI valuations move quickly, and figures cited are accurate as of publication.