Why AI companies are starting to design their own computer chips

Reports indicate that Anthropic, the company that created the AI assistant Claude, is in early discussions with Samsung to manufacture custom-built computer chips. While Anthropic has not finalized what these chips will actually do or how they will be used within their server systems, the search for a new partner signals a clear shift. Anthropic currently relies on hardware from industry giants like Nvidia, Google, and Amazon to power its systems. By working with Samsung, the company appears to be moving toward a strategy where they design their own specialized processing units.

Modern AI models are essentially giant engines for prediction. To get them to answer questions or write code, they require massive amounts of power to perform calculations at high speeds. Most companies achieve this by using general-purpose processors designed by firms like Nvidia. These chips are powerful and reliable, but they are built for many different tasks, not just one specific AI. Designing a custom chip is like moving from driving a mass-produced sedan to engineering a high-performance race car tailored for a specific track. By building a custom chip, a company can optimize every single circuit to handle the specific type of math that their AI needs. This increases performance per watt, meaning the system can do more work without consuming as much electricity or generating as much heat.

The computer chip industry is currently dominated by a few massive companies. This creates a bottleneck: if everyone is using the same hardware, it becomes harder for an AI company to differentiate its product or protect itself from supply shortages. By designing their own chips, companies like Anthropic and its rival, OpenAI, are trying to reduce that reliance. This trend suggests that the next generation of AI will not just be faster because of better software, but because the hardware underneath is finally being built with one single purpose in mind. For the average person, this doesn't change how you use an AI today, but it ensures that the services you use remain stable and become more efficient over time.