Anthropic to Develop Its Own Chips for Claude

Anthropic is putting together a “custom silicon team” to develop the chips its models will run on. The news came to light after Business Insider noticed a job listing for a senior engineer with experience bringing semiconductor products to market — and the company's careers portal currently has openings for a silicon engineer and a technical program manager as well. An Anthropic spokesperson confirmed the plans in comments to Business Insider and TechCrunch.

At the same time, the company stresses that its approach will remain “multi-chip”: alongside its own designs, Anthropic will continue to use hardware from other manufacturers as it scales. Rumors to this effect had circulated before — according to The Information, the company had been considering Samsung as a manufacturing partner.

A trend, not an exception

Anthropic isn't alone on this path. Its rival OpenAI recently unveiled Jalapeño, a custom chip for inference of large language models in data centers, developed jointly with Broadcom. Google has long been running its models on its own hardware, Meta is designing and deploying its own chips too, and Mistral, by some reports, is exploring the option.

There are several reasons for this race. First, a large part of the industry is heavily dependent on Nvidia, and that supplier's enduring dominance is a potential strategic vulnerability — especially at a time when demand for computing power outstrips supply. Second, designing chips around specific models (and vice versa) promises performance gains. If OpenAI can reap the benefits of such vertical integration, every other developer of frontier models will want the same.

That's precisely why Anthropic says its teams will design new hardware and models side by side. The company has co-developed hardware with partners before, but now it intends to bring more silicon expertise in-house.

Anthropic may be counting on this to give its flagship models an added edge — at a time when developers and other users are increasingly experimenting with running cheaper, smaller, or open models on their own hardware and on edge devices.

Still, it's too early to get one's hopes up: the team's key hires are only just being made, so the first fruits — for the company itself and for its users alike — are still a long way off.

Source: Ars Technica