Published by Emerging Technologies Laboratory · via ETL Newswire
Technology· 

Anthropic Opens Custom Chip Talks With Samsung as Compute Costs Climb

The Claude developer is in early discussions with Samsung's foundry unit about a bespoke AI accelerator, its first move into custom silicon after years of renting capacity from Nvidia, Google, and Amazon.

By Theo Okafor, Staff Reporter · Technology Desk

Anthropic has begun preliminary discussions with Samsung Electronics about manufacturing a custom AI chip, according to reporting by The Information. The talks put the Claude developer inside a race it sat out while OpenAI, Google, Amazon, and Meta each built proprietary silicon, a gap that is getting harder to ignore as inference costs eat into the company's margins.

The conversations are early, and the missing components are significant. According to The Information, Anthropic has not yet decided what workloads the chip will run, how powerful it needs to be, or how it would fit into a server rack. That's not a minor detail list, those are the three decisions that define a chip program. What the company appears to have settled on is the manufacturing target: Samsung's 2-nanometer process node and its advanced packaging facilities, per The Information.

The hire that signals this is real: Anthropic brought on Clive Chan in June. Chan was among the first hardware engineers on OpenAI's custom-chip program and worked on that project from its earliest stages, according to reporting reviewed by UPI. You don't recruit that profile to run a thought experiment.

The financial pressure behind the move is concrete. Anthropic is spending an estimated $1.25 billion per month on compute, according to the AI tools tracker AIToolsRecap. The company currently relies on Amazon's Trainium chips, Google's Tensor Processing Units, and Nvidia's GPUs. Anthropic confirmed to TechCrunch that a diversified hardware stack drawing on those three suppliers will remain central to its compute strategy, but declined to comment on the Samsung discussions directly.

Custom silicon is the standard answer to that kind of bill. When a lab designs a chip around its own model architecture, it can strip out general-purpose compute it doesn't use and optimize the memory bandwidth and interconnect topology it does. OpenAI went this route with Broadcom, unveiling an inference chip called JalapeƱo on June 24. Google has run multiple generations of its Tensor Processing Units. Amazon built Trainium. The pattern is established; Anthropic is arriving later than most.

Samsung is not a neutral pick. The Korean conglomerate was one of three memory chipmakers, alongside SK Hynix and Micron, that participated in Anthropic's $65 billion fundraising round in May, per reporting reviewed by The Street. Samsung is also the only one of those three investors that operates its own chip foundries, which means the manufacturing relationship and the investor relationship would sit inside the same corporate entity. That is either a useful alignment of interests or a governance question, depending on how the deal is eventually structured.

For Samsung's foundry business, the strategic value is clear. The unit has been trying to close the gap with TSMC, which holds a near-monopoly on leading-edge AI chip manufacturing. TrendForce projects that server shipments using custom application-specific integrated circuits will grow 44.6 percent in 2026, compared with 16.1 percent for general-purpose GPU-based servers, per data cited by UPI. A marquee AI lab customer would give Samsung's foundry a reference win it badly needs.

Anthropics's talks with Samsung aren't exclusive. The company is also in discussions with Microsoft and U.K.-based chip startup Fractile about potential processors, according to The Information. That looks like a competitive evaluation process, not a done deal with one partner.

Nvidia, for its part, hasn't lost the business yet. The Information estimates the company still holds 74 percent of the AI chip market, higher than before the custom-silicon arms race started. Custom chips take years to design, tape out, validate, and integrate. Anthropic's compute strategy for 2025 and most of 2026 runs through Nvidia regardless of what Samsung eventually builds.

Sources cited:
- The Information (https://www.theinformation.com/articles/anthropic-talks-samsung-manufacture-custom-ai-chip)
- TechCrunch (https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/)
- UPI / Asia Today (https://www.upi.com/Top_News/World-News/2026/07/03/Anthropic-Samsung-Electronics/7811783128641/)
- The Street (https://www.thestreet.com/technology/anthropic-samsung-custom-chip-talks)
- AIToolsRecap (https://aitoolsrecap.com/Blog/AINewsJuly2026.aspx)
- SiliconANGLE (https://siliconangle.com/2026/07/02/anthropic-reportedly-talks-samsung-manufacture-custom-ai-chip/)

Reporting by Theo Okafor, Staff Reporter, for the Technology desk · ETL Newswire staff
Read more at the source

This release was originally distributed via ETL Newswire. Visit The Information for the full story, related releases, and contact information.

Visit The Information →