Anthropic reportedly in talks with Samsung to manufacture custom AI chip

Anthropic PBC is reportedly in talks with Samsung Electronics Co. to produce a custom artificial intelligence chip.

The Information today cited sources as saying that the discussions are in an early stage. According to the publication, Anthropic has not yet finalized key details such as what workloads the processor will run and its performance.

Some AI accelerators, including Nvidia Corp.’s flagship Rubin graphics processing unit, can both train AI models and perform inference. Other chips have a narrower focus. Etched Inc., a startup that closed a $800 million funding round this week, has developed an accelerator designed solely for inference.

Some AI processors are optimized for even more granular tasks. Earlier this year, Nvidia debuted a chip called the LPU 30 that is designed to perform a specific subset of the calculations involved in inference workflows. The chip is based on Samsung’s four-nanometer process.

Today’s report didn’t specify which manufacturing technology Anthropic will use to make its AI accelerator. Samsung offers its four-nanometer node alongside a more efficient two-nanometer process called SF2P. The latter technology, which is set to enter production later this year, is specifically optimized for data center chips.

A transistor is built around a structure called a channel that shuffles electrons between two points. The flow of electricity is coordinated by a so-called gate. In transistors made using Samsung’s SF2P process, the gate surrounds the channel to reduce power leaks. Additionally, the tiny wires that link together SF2P circuits feature optimizations designed to boost performance.

Samsung manufactures not only logic circuits but also HBM memory, which many AI chips use to store the data they process. It’s possible that Anthropic’s upcoming accelerator will also feature the high-speed RAM variety.

The report about the company’s chipmaking push comes a few days after OpenAI Group PBC debuted its first custom processor. Jalapeño is an inference accelerator that the AI provider designed in collaboration with Broadcom Inc. Earlier, the latter company helped Google LLC develop its competing TPU chips. It’s possible that Anthropic will likewise team up with an external chip design partner to accelerate its semiconductor roadmap.

The company will presumably use the processor to power its upcoming data center network. Last year, Anthropic announced a $50 billion initiative to build AI facilities in the US. It’s pursuing the project in partnership with Fluidstack Ltd., a startup that helps customers design and operate data centers.

Anthropic told TechCrunch today that it plans to continue using chips from Amazon Web Services Inc., Nvidia and Google. In April, the AI developer committed to purchasing more than $100 billion worth of AWS infrastructure over the next decade.

Image: Anthropic