Why AI companies are racing to design their own chips

Until recently, the AI industry's hardware story has largely revolved around a single company: Nvidia. Nearly all the graphics processing units used to train and run large language models come from that one supplier. But according to Ars Technica, a leading AI company is now trying to change that picture, building an in-house chip team to design its own custom hardware.

The move is the latest example of an increasingly common strategy. Google has developed its own Tensor Processing Units for years; Amazon and Microsoft have invested in similar custom chip programs. Adding another name to that list signals just how far control over AI infrastructure is being pushed down to the most fundamental layer of the supply chain.

The logic behind the move is both strategic and economic. Demand for Nvidia's chips is so high that even the largest AI companies sometimes face supply delays stretching months. A company that designs its own hardware can become largely independent of that bottleneck.

Cost is also a major factor. Nvidia's chips command a substantial profit margin due to high demand and limited competition. A company that can produce its own hardware could significantly lower the long-term cost of large-scale AI operations, especially as model training and inference have become some of the biggest line items in AI budgets.

Another advantage of custom chip design is performance optimization. A general-purpose GPU is built for a wide range of tasks, but a chip designed specifically for the operations a company's own models actually need can deliver much higher efficiency for the same energy and hardware cost.

Still, this is far from an easy path. Chip design is a highly specialized field, and only a handful of companies worldwide can do it at scale. Building that capability from scratch requires years of investment and a significant amount of specialized engineering talent.

Manufacturing adds another layer of difficulty. Designing a chip is one thing; physically producing it is another — a process that typically depends on a small number of advanced semiconductor fabricators such as TSMC. So even a company that designs its own chip still needs an outside manufacturing partner.

The development can be read as a sign of the AI industry maturing. In its early years, companies focused on software and simply bought hardware off the shelf. Now competition is playing out both in model quality and in the cost and speed of running those models — and hardware has become a central part of that competition.

Experts say custom chip efforts like this one won't pay off in the short term, but if successful, they could reshape the balance of power in the industry over the long run. Nvidia remains the clear market leader, but its largest customers building their own alternatives could put pressure on its future market share.

Ultimately, the move shows that the AI race isn't just about producing better algorithms — increasingly, what hardware runs those algorithms, at what speed, and at what cost is becoming just as decisive.