Kings Research has published an analysis asking why hyperscalers design their own AI chips, and the question answers itself: the largest cloud operators no longer wish to rent their competitive edge from a vendor.

Google, Amazon, Microsoft and their peers have spent years building custom accelerators for their data centers. The report frames this as a structural shift in who designs silicon and who merely consumes it. The commercial logic is straightforward. A hyperscaler that designs its own chip controls its own roadmap.

What drives the in-house design push?

The analysis identifies several pressures behind the trend. Cost sits near the top. Merchant AI accelerators carry margins that hyperscalers would rather keep for themselves, and procurement at cloud scale turns every percentage point of vendor margin into a line item worth arguing about.

Supply matters just as much. Demand for AI training hardware has repeatedly run ahead of available supply, and a buyer dependent on a single merchant vendor absorbs the shortage first. An internal design program, run alongside external purchases, gives a hyperscaler a second source of its own making.

Power efficiency is the third lever. Hyperscalers can tailor a chip to their exact workloads rather than accept a general-purpose design. When a data center's ceiling is its electricity bill, a part tuned for one job beats a part built for every job.

How does this reshape the vendor landscape?

The report's framing implies a divided market. Nvidia and other merchant suppliers retain the customers who cannot afford a design team, while the largest cloud operators become their own preferred suppliers. That split does not collapse overnight. Custom silicon takes years and billions to reach volume, and even the hyperscalers continue to buy merchant chips alongside their own.

Still, the direction is set. Each in-house accelerator that reaches production volume displaces merchant silicon at the customer with the deepest pockets. For vendors, the loss is quiet but compounding: not a lost contract, but a customer that gradually stops needing one.

The report treats the shift as durable rather than cyclical. As AI workloads grow, the argument for owning the silicon that runs them grows with them. Expect hyperscaler silicon roadmaps to widen, and expect merchant vendors to defend their position on the one ground that remains: building what everyone else cannot.