Anthropic may start making its own chips for the same reason Nvidia started making its own models.

If you're not familiar with the term "extreme co-design" yet, then it's time to start learning. Nvidia executives used the phrase repeatedly at the company’s GTC event in March, and I expect it's a big part of the strategy behind the new Reuters report that Anthropic is exploring the opportunity to design its own AI chips.

Extreme co-design allows you to architect the data center, chips, software stack, and LLMs in tandem, resulting in far better optimization and integration. The result can reduce the cost per token, lower latency, and minimize power consumption.

That's why Nvidia is now making its own open-source frontier models: By building its own models, Nvidia learns how to better optimize ALL models for its AI chips.

At GTC in March, Kari Briski, VP of generative AI software at Nvidia told The Deep View that the process allows the company "to break some rules, to see around the corner… When we say extreme co-design, it's literally about sitting down side by side, day by day, showing each other reports… because we're all racing at the speed of light to the next thing... [And] you have to be building it right now. It can't be a year from now. So we talk about extreme co-design literally in lockstep every single day, understanding each other's pains."

But beyond co-design, there are also a couple of other reasons why Anthropic is likely considering the move:

• Remove the constraint: Now that Anthropic has proven it can consistently build some of the world's best and most powerful AI models, one of the company's biggest constraints is simply getting enough compute to run its models, serve its chatbot and coding tools to new and existing customers, and train future models. That's why it's been signing new deals with companies like Google/Broadcom and CoreWeave. But chips themselves are one of the biggest constraints for hyperscalers and neoclouds (and leads me to ponder if Anthropic might also consider acquiring a neocloud at some point).
• Save money in the long run: Spinning up a chips program would be incredibly expensive and it likely wouldn't cover the company’s entire demand for chips. Anthropic would still end up buying chips from other vendors for the foreseeable future. However, in the perspective of decades, it could eventually mean Anthropic would control one of the most important drivers of its technology, and could save money by avoiding padding the profit margins of other companies (like Nvidia).

Our Deeper View

Let's keep in mind that this report indicates Anthropic is looking to design its own chips, not manufacture them. It would likely still rely on TSMC or Samsung to actually make the chips. However, even that could drastically reduce the time it takes for Anthropic to get the chips it needs, because it wouldn't have to stand in line with OpenAI, Microsoft, Meta, and basically everyone else building AI to get the brains it needs to run the operation. If you consider all the pieces of the extreme co-design process, you can quickly make a list of long-term acquisition targets for Anthropic, OpenAI, and other AI giants.