Arm’s quiet coup: 90% of AI servers by 2029?

A report from Omdia drops a number that should unsettle Intel and AMD: by 2029, 90% of AI servers running custom silicon will use Arm’s instruction set. That’s not a guess about market share—it’s a projection of where hyperscalers like Amazon, Google, and Microsoft are already heading as they build their own chips. These companies aren’t just swapping architectures for fun; they’re chasing 2x the performance-per-watt Arm claims over x86 in AI workloads, and the control that comes with designing silicon in-house.

The shift isn’t theoretical. Amazon’s Graviton chips already power 40% of its cloud instances, while Google’s Tensor Processing Units pair with Arm-based CPUs for efficiency. Even Microsoft’s Azure Cobalt chips, built on Arm, signal the same trend: hyperscalers are optimizing for AI first, and legacy x86 compatibility is now a nice-to-have, not a requirement.

This isn’t just about raw performance. It’s about the total cost of ownership—power bills, cooling, and the ability to tweak hardware for specific AI models. When a single data center can cost $1 million per month to run, even a 10% efficiency gain justifies a full architectural overhaul. The question isn’t whether Arm wins, but how fast x86 gets pushed to the margins.

The losers here aren’t just Intel and AMD. RISC-V, despite its open-source promise, lacks the mature ecosystem Arm offers—compilers, debuggers, and decades of server-grade optimization. Hyperscalers tried RISC-V in prototypes but hit walls with software support. Arm, meanwhile, has spent years courting cloud providers with Neoverse cores designed for scale, plus a licensing model that lets companies customize without reinventing the wheel.

For end users, the change will be invisible at first. Your AI-generated images or chatbot responses won’t suddenly feel faster—but the companies serving them will pay less per query, and that savings might trickle down to pricing. The bigger impact is on developers. Arm’s dominance means optimizing for SVE2 vector extensions and memory tagging becomes table stakes, while x86-specific tweaks become legacy debt. Small cloud providers without custom silicon budgets? They’ll be stuck renting Arm instances from the hyperscalers or paying a premium for x86.

The wild card is regulation. If Arm’s IPO struggles or Nvidia’s acquisition ghost haunt its neutrality, hyperscalers might hesitate to bet everything on one ISA. But for now, the math is simple: Arm offers the efficiency gains AI demands, and the hyperscalers hold the checkbooks.