Anthropic in Talks With Samsung for First Custom AI Chip on 2nm

Per The Information, Anthropic has begun early-stage talks with Samsung Electronics to manufacture its first custom-built AI chip on Samsung’s 2-nanometer process. The project is still nascent, no detailed design work has begun, and the company may still walk away. A deal would hand Samsung its first marquee AI-lab foundry customer and mark the latest custom-silicon move aimed at chipping away at Nvidia’s grip on the accelerator market.

Anthropic Opens Early-Stage Talks With Samsung for a 2nm Chip

A Tuesday report by The Information, published as Anthropic’s Samsung 2nm chip talks on Yahoo Finance, said Anthropic has begun early-stage work on a custom AI chip and held talks with Samsung Electronics as a potential manufacturing partner. The specific pieces on the table are Samsung’s 2-nanometer manufacturing process and the Korean conglomerate’s advanced packaging facilities. The project is described as “still nascent,” with no detailed design or manufacturing work yet begun, and Anthropic itself may decline to proceed. Even at this stage, Anthropic has begun adding the engineering muscle any custom-silicon program needs.

Anthropic recently brought on Clive Chan, an early member of OpenAI’s own custom chip team, as part of what The Information called a “deliberate engineering buildout.” Chan is OpenAI’s second-ever hardware hire on the custom chip program, and he announced his move on X earlier in the month.

Anthropic’s growing scale makes the strategy easier to underwrite. Run-rate revenue crossed $47 billion earlier this year, per the company’s Series H announcement. The round raised $65 billion at a $965 billion post-money valuation, and it brought Samsung Electronics onto Anthropic’s cap table as a “strategic infrastructure partner.”

OpenAI Just Shipped Its Own Chip, and Anthropic Is Following the Map

The clearest proof point sits one lab over. OpenAI and Broadcom unveiled OpenAI’s Jalapeño launch announcement last month, the first inference processor designed from scratch for large-language-model inference. Jalapeño was co-developed from initial design to tape-out in nine months, what OpenAI calls the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. Engineering samples are running ML workloads at production target frequency and power, including on GPT-5.3-Codex-Spark.

Anthropic’s project runs on the same playbook, earlier in the timeline. As Reuters reported in April, Anthropic was already considering developing its own chips; Chan’s hire confirms the work has begun in earnest. The table below tracks where four of the frontline AI labs sit on custom silicon today. The shape across the four is consistent: a design partner (often Broadcom) or in-house team, a foundry partner (TSMC historically, with Samsung now competing), and a deployment target that slides as the silicon matures. None of the four has fully decoupled from Nvidia, and Anthropic’s path is unlikely to be either.

Clive Chan, OpenAI’s “Hardware Hire Number Two”

The 2.4-year tenure Chan logged at OpenAI gives Anthropic a hire who has lived through a full custom-chip build. By his own description, he was the second hardware hire on OpenAI’s custom-chip initiative, a team he called an “exceptional chip design organization.” Chan announced his move on X on June 6, in the form of the report on Clive Chan leaving OpenAI’s chip program.

Chan had joined OpenAI in January 2024 from Tesla’s Dojo supercomputer project, where he had worked on training-chip silicon. At OpenAI he helped build the program that produced Jalapeño alongside Broadcom. He tied his exit to a personal itch: to “climb a new mountain from the bottom again.”

I joined Anthropic this week because I was deeply impressed with the team’s talent, values, and ambition, and I’m already energized by the pace and intensity of the past few days. It’s time to build.

The quote is from Clive Chan, a hardware engineer, in his X post on June 6, 2026.

Chan said he cannot yet discuss many details about OpenAI’s chip program. He pointed instead to the previously announced timeline for OpenAI-designed AI accelerators to deploy in the second half of 2026.

Anthropic today runs on chips from Nvidia, Google, and Amazon, with no in-house silicon announced yet. The Information’s account of the company’s reply on a hardware partnership was a reaffirmation of its multi-vendor compute strategy. The Series H made the chip conversation a more public part of that diversification. Bringing a former colleague from the rival lab onto the team is one way to make the diversification look credible from the inside.

Can Nvidia’s 74% Lead Hold Against a Convergence of Lab Silicon?

The data points that complicate Nvidia’s position are stacking up. Nvidia holds an estimated 74% share of the AI chip market, per The Information’s own estimate, higher than it was before the inference-chip arms race began. Anthropic’s project, OpenAI’s Jalapeño, Google TPUs, Amazon Trainium, and the Meta silicon program all push in the same direction, every frontline lab engineering itself out of single-vendor dependence.

Anthropic’s Series H makes the financial glue visible. Samsung, SK Hynix, and Micron all joined the $65 billion round as “strategic infrastructure partners” whose products sit inside Anthropic’s stack. The compute footprint now spans AWS, Google Cloud, and Microsoft Azure, with AWS remaining the primary cloud and training partner.

OpenAI’s Jalapeño roll adds pressure on the same axis. Engineering samples running GPT-5.3-Codex-Spark in OpenAI’s labs are a concrete signal that the path from prototype to deployment is shorter than skeptics expected, with Broadcom targeting initial deployment by the end of 2026.

The world is moving to a compute-powered economy. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.

The quote is from Greg Brockman, President and Co-Founder of OpenAI, in the company’s Jalapeño release on June 24, 2026.

What an Anthropic Win Would Change for Samsung and TSMC

A signed Anthropic-Samsung manufacturing agreement would land at a moment Samsung has been building toward for years. South Korea’s president, Lee Jae Myung, joined the chairs of Samsung Electronics and SK Hynix in announcing a combined $518 billion chip-hub investment, the 800 trillion won buildout covered in PBS’s reporting on South Korea’s $518 billion chip hub plan. Each company is building two new fabrication plants in the country’s southwest region, expanding beyond the existing Gyeonggi Province complexes south of Seoul. Samsung’s new fabs are slated for the southwestern city of Gwangju. The broader government strategy is to build a nationwide semiconductor ecosystem, with chip packaging concentrated in the central Chungcheong region.

The technical case for Samsung as a foundry partner is also narrowing. Samsung’s SF2 2-nanometer process is set to expand from mobile mass production into high-performance computing and AI accelerators in 2026, per the company’s Foundry Forum roadmap. The jump is measured at a 25 percent power efficiency improvement at the same frequency and complexity and a 12 percent performance boost at the same power consumption and complexity, against Samsung’s second-generation 3GAP at 3 nanometers.

Samsung has reportedly secured an inaugural 2nm AI-accelerator customer in Japanese AI startup Preferred Networks. Google is also reportedly in talks with Samsung to manufacture part of a next-generation TPU, a second marquee AI customer on Samsung’s order book. TSMC still owns the higher-margin position with Apple’s 2nm order book, Nvidia’s leading-edge GPUs, and the longer client roster. Anthropic would not flip that on its own, though it would give Samsung a second credible AI reference design to compete for the next round of frontier-lab silicon business.

Anthropic’s Multi-Vendor Bet, Built Around the New Chip

Anthropic’s broader strategy, laid out in the Series H statement, treats any custom silicon as one more option rather than a replacement for the rest. The company runs on three clouds at once and draws compute from multiple chip families; an internal Anthropic chip would sit alongside, not displace, Nvidia GPUs, Google TPUs, and Amazon Trainium. The throughline is the same one OpenAI’s Brockman described: more of the stack under the lab’s own control, as laid out in our related coverage of Anthropic’s trillion-dollar IPO filing ahead of OpenAI.

1. Up to 5 gigawatts of new compute capacity with Amazon, with AWS remaining the primary cloud and training partner.
2. 5 gigawatts of next-generation TPU capacity with Google and Broadcom.
3. GPU capacity inside SpaceX’s Colossus 1 and Colossus 2 data-center complexes.
4. $15 billion of previously committed hyperscaler investment, including $5 billion from Amazon, layered into the Series H.

None of these agreements is anchored to the custom-silicon timeline in any public sense. They are the financial and capacity scaffolding that lets Anthropic pick its moment on a chip of its own. The custom chip, when it lands, would be one more lever inside that multi-vendor environment, not the thing the portfolio was waiting on.