Anthropic is exploring a custom AI chip and wants Samsung to make it

Anthropic is creating its own AI chip and is in talks with Samsung Electronics (KS:005930) regarding its production.

This has further increased pressure on NVIDIA (NASDAQ:NVDA), which currently holds around 74 percent share in the AI chip market.

The discussions are concerning Samsung’s 2-nanometer manufacturing process as well as the company’s packaging segment, which is concerned with connecting different components of a processor together. Anthropic has also hired Clive Chan, who was part of OpenAI’s early custom-chip team.

In fact, OpenAI had decided to work with Broadcom (NASDAQ:AVGO) to design their silicon in 2024. This collaboration resulted in Jalapeño, an inference chip allegedly designed to operate large language models more efficiently.

Anthropic builds a chip team as AI companies cut their dependence on Nvidia

Alphabet (NASDAQ:GOOGL) has its tensor processing units. Amazon (NASDAQ:AMZN) has Trainium. Meta Platforms (NASDAQ:META) and Microsoft (NASDAQ:MSFT) have also been working on their own chips.

Broadcom is already involved in customized AI silicon with OpenAI. Taiwan Semi-Conductor Manufacturing (NYSE: TSM) is still the name to remember when it comes to advanced chips.

Samsung has been struggling in its attempts to rival TSMC in terms of yields at leading edge due to TSMC advancing toward its own N2 process.

Nvidia (NASDAQ: NVDA) surged by 0.7% on Thursday by close, and Broadcom (NASDAQ: AVGO) added 0.7%, while TSMC (NYSE: TSM) increased by 3.5%.

Samsung is also being considered by Google for part of a future tensor processing unit. That would be another large AI manufacturing order if it happens. Samsung, SK Hynix, and Micron Technology (NASDAQ:MU) also joined Anthropic’s $65 billion May fundraising round. Earlier this week, Samsung Group and SK Group said they plan to invest $518 billion over ten years to build four memory-chip plants in South Korea.

Anthropic keeps Nvidia, Amazon, and Google in its compute mix

Anthropic allegedly told The Information that “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.” Anthropic did not give more details about its roadmap.

The company is also talking about using chips from Microsoft (NASDAQ:MSFT) and UK startup Fractile. Anthropic wants options:- Nvidia GPUs, Google TPUs, Amazon Trainium chips, possible Microsoft hardware, possible Fractile chips, and maybe one day its own Samsung-made processor.

At the same time, Anthropic just came out of a fight with the US government over access to its top models. The US Department of Commerce lifted export controls on Anthropic’s Mythos and Fable models on Tuesday evening.

That ended a weeks-long standoff between the company and the Trump administration. The decision lets Anthropic release Fable 5 to the public again. The government had already allowed Mythos 5 to return last week, but only for about 100 pre-vetted partners. Mythos 5 has fewer safeguards and is built for enterprise customers.

Commerce Secretary Howard Lutnick told Anthropic co-founder Tom Brown in a letter that he was removing the ban after the company had “agreed to proactively detect and address security risks associated with the models.”

A person close to Anthropic said the company had “implemented a new safeguard” that answered the government’s concern. That fix was tested and approved by the Center for AI Standards and Innovation.

Howard initially imposed the restrictions following officials’ discovery of a jailbreak that could circumvent model safety measures on June 12. Subsequently, Anthropic decided to withdraw its most capable models from users worldwide.

The standoff sparked frustrations among sections of Silicon Valley and foreign governments. Industry leaders accused Washington of adopting an abrupt approach to AI regulation. While some foreign officials accused the US of betraying its partners through such restrictions.

The ban followed the Trump administration’s issuance of a voluntary monitoring program for advanced AI models. The administration had assured that it would not enforce a comprehensive regulatory regime on top AI firms. However, several US government officials have warned about the dangers of frontier models due to their capability of helping attackers exploit cybersecurity weaknesses in key industries without sufficient expert review.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.