Samsung’s OpenAI Custom Chip Stalls as Anthropic Steps In

The custom processor Samsung Electronics had been quietly designing for OpenAI has stalled, and the company best placed to benefit is the one that just overtook OpenAI in value. Talks over the bespoke Samsung OpenAI custom AI chip have cooled over what both sides describe as strategic differences, according to a report by Korean outlet Greened, even as Samsung’s foundry arm courts a fresh client in Anthropic.

Samsung’s ambitions in custom silicon have not collapsed with those talks. In the same stretch of weeks, the company signed on as a backer of Anthropic, the maker of the Claude chatbot, in a deal that explicitly names logic chips as a future area of work.

Samsung’s OpenAI Inference Chip Stalled Over Strategy

The blueprint was ambitious. Samsung’s custom chip division had been building a dedicated inference processor for OpenAI, a data-center neural processing unit (dNPU, a chip tuned to run trained models rather than train them) designed on Arm’s power-efficient architecture.

Engineers on both sides had made real headway before the project went quiet. A report by the Korean outlet Greened, surfaced in English by SamMobile, says the work cooled indefinitely over what the parties called strategic differences about the chip’s direction. Samsung has not confirmed a halt, and people close to the project say the design has not been scrapped and could resume.

That ambiguity matters because Samsung built this team on purpose. The company recently stood up a dedicated unit for custom system-on-chip (SoC) work, chasing a slice of the cloud-processor market that Broadcom and Marvell have largely owned.

An inference chip for the most-used chatbot on the planet would have been a marquee reference win. Losing momentum on it, even for a while, stings a division that has spent two years trying to prove it can compete at the leading edge.

Anthropic Steps Into the Seat OpenAI Left Open

Samsung did not have to wait long for a second shot. On May 28, Anthropic closed one of the largest private funding rounds in tech history, and Samsung’s name sat on the list of backers.

The round, detailed in Anthropic’s Series H funding announcement, named Samsung, SK Hynix and Micron as strategic infrastructure partners and pointed to their role in supplying memory, storage and logic chips. That last phrase is the tell. Among the three, only Samsung runs its own foundry; the other two build memory, not processors. The number that grabbed most attention, of course, was the valuation behind Anthropic’s leap past OpenAI in valuation.

• $65 billion raised in Anthropic’s Series H round
• $965 billion post-money valuation as of May 28
• 3 named strategic infrastructure partners: Samsung, SK Hynix and Micron

Anthropic still leans heavily on third-party cloud infrastructure to train and serve Claude. Designing its own silicon, and having Samsung fabricate it, would buy the kind of independence that every large model builder now wants. Industry watchers read the logic-chips line as a signal that Samsung’s plants, not its peers’, are the ones in play.

Samsung’s Foundry Needs an Anchor AI Client

For Samsung, an Anthropic order would land somewhere far more consequential than a product roadmap. It would land on the factory floor.

The Taylor Bet Still Has to Pay Off

Samsung’s roughly $44 billion fab in Taylor, Texas is the centre of this story. The site began risk production on its 2-nanometer SF2 process in February, held an equipment installation ceremony on April 24, and is building toward full mass production targeted for 2027, with capacity plans that climb toward 50,000 wafers a month. A fab that size cannot run profitably on a thin order book.

Filling 2nm Lines Is the Whole Game

The clearest anchor so far is Tesla, which awarded Samsung a $16.5 billion contract for its AI5 and AI6 driving chips, with the AI6 set to run on the SF2 node. Other names are circling: AMD has been sampling Samsung’s 2nm process and Google’s chip team has reportedly visited Taylor, as covered in TrendForce’s read on Tesla’s AI5 and AI6 production split across foundries. Each name that converts pulls the foundry unit closer to break-even.

What Samsung and OpenAI Still Build Together

None of this means Samsung and OpenAI have stopped working together. The custom chip is one lane; several others stay open.

• A co-development effort using Samsung SDS resources to build physical data centers for OpenAI
• Supply of high-bandwidth memory (HBM, the stacked DRAM that feeds AI accelerators) and other high-performance memory modules
• The broader, wide-ranging cooperation agreements the two firms signed before the chip talks cooled

A Samsung representative says cooperation with OpenAI continues smoothly behind the scenes, which fits a relationship that was always wider than a single processor.

OpenAI and Anthropic Take Their Fight to the Chip Supply

The bigger frame here is that the AI race has spilled into the supply chain. Anthropic’s $965 billion mark put it ahead of OpenAI, which carried a roughly $852 billion valuation in March, and the two are now competing for memory, foundry capacity and engineering talent at the same time.

OpenAI is not standing still while its rival courts the chipmakers. The company has been pushing back on multiple fronts, including the competitive moves behind OpenAI’s competitive push against Anthropic.

For Samsung, the timing could hardly be better. With TSMC’s leading-edge lines packed with orders from Apple and Nvidia, there is genuine demand spilling toward a credible second source, a point TrendForce made in its analysis of Samsung’s opening in the 2nm foundry race.

If the Anthropic talks turn into a signed foundry contract, Samsung gets the AI anchor client its Texas lines need and the OpenAI stall becomes a footnote. If they settle where the OpenAI project now sits, preliminary and frozen, Samsung is left courting customers it cannot yet count.