Here's a thing that happens in tech: a company builds its business on someone else's chips, then decides it could do a better job itself. It's happening right now in the world of artificial intelligence, and it's reshaping who makes the brains inside the biggest data centers on Earth.
According to a recent note from Counterpoint Research, the AI infrastructure market is undergoing what they call a "structural shift." The giants of the cloud—Alphabet Inc.'s Google (GOOGL), Amazon.com Inc.'s AWS (AMZN), Microsoft Corp. (MSFT), and Meta Platforms Inc. (META)—are increasingly moving away from the legacy x86 central processing units (CPUs) made by Intel Corp (INTC) and Advanced Micro Devices Inc. (AMD). Their new destination? Proprietary silicon built on designs from Arm Holdings plc (ARM).
The goal is straightforward: optimize for cost, efficiency, and control. But the move is a big deal because it rewrites a long-standing rulebook.
Why Hyperscalers Are Driving the Shift
For years, these companies relied on Intel and AMD. The x86 architecture was the standard, the software was compatible, and the infrastructure was built around it. It was the safe, established choice. But the explosive rise of AI has changed the calculus.
Running modern AI workloads at scale is incredibly power-hungry and expensive. When you're operating thousands of servers, even a small efficiency gain per chip adds up to massive savings on electricity and cooling. More importantly, the rise of custom AI accelerators—the specialized chips that actually do the heavy lifting of AI training and inference—has created a need for new, "heterogeneous" architectures. The old general-purpose CPU isn't enough; you need a CPU that's perfectly tuned to work with your specific AI accelerator.
That's where Arm comes in. Its architecture, particularly the Neoverse cores designed for data centers, is proving to be a better fit. Counterpoint notes that Arm delivers "significantly better performance per watt than traditional x86 systems." In a world where data centers are hitting power limits, that's not just a nice-to-have; it's a critical advantage.
This shift reflects a broader strategy you see across Big Tech: vertical integration. By designing their own silicon in-house, hyperscalers reduce reliance on external vendors, improve their own profit margins, and gain more control over their technology roadmap. They're not just buying chips off the shelf anymore; they're architects of their own destiny.
From Theory to Reality: Arm in Action
This isn't a future prediction; it's happening now. Companies are already deploying Arm-based CPUs across their AI infrastructure, and they're designing them to be inseparable from their AI systems.
Google is scaling its custom Arm-based CPU, called Axion, to work with its next-generation Tensor Processing Unit (TPU) systems. Amazon Web Services is expanding the use of its Graviton processors (built on Arm) to work alongside its custom Trainium AI chips. Microsoft didn't just add an Arm chip later; it integrated its Azure Cobalt Arm CPU with its Maia AI accelerators from the very beginning, embedding Arm into its AI stack from the outset.
The message is clear: Arm is no longer just for low-power mobile devices or general-purpose cloud workloads. It's becoming the central host CPU in the design of AI servers.
Reshaping the Competitive Landscape
The transition is unfolding step by step, generation by generation. Hyperscalers are aligning their CPU design cycles closely with the development of their proprietary AI accelerators. A prime example is Meta, which has selected Arm as a "strategic partner" for its next-generation Meta Training and Inference Accelerator (MTIA) infrastructure. Meta is even serving as the launch customer for Arm's new Artificial General Intelligence (AGI) CPU platform.
This coordinated shift is expected to accelerate dramatically starting in the second half of 2026, driven by the broader deployment of these in-house Arm CPUs. The projections from Counterpoint are striking: they believe Arm-based CPUs could account for around 90% of host CPU deployments in custom AI Application-Specific Integrated Circuit (ASIC) servers by 2029. For context, that figure is only about 25% in 2025.
Think about that for a second. In less than five years, the dominant architecture in the most advanced AI servers could flip from x86 to Arm. That's a seismic change for the semiconductor industry.
The ripple effects are extending across the entire supply chain. The demand for advanced manufacturing—the cutting-edge fabs that etch these incredibly complex chips—is rising to support both the new wave of AI accelerators and the Arm-based CPUs that host them. It's a whole new game, and the players who defined the old one are watching their turf get redrawn.