Arm announced their own CPU this week. Not a design to license - an actual chip you can buy. For a company that's spent 34 years licensing IP to other manufacturers, this is a structural shift, not a product line extension.
Ben Thompson's analysis at Stratechery cuts through the obvious competitive narrative. Yes, Arm's clients are building their own designs. But the deeper driver is how computing itself is changing. The old model - license the architecture, let others manufacture - worked brilliantly for mobile. It's less obvious that it works for what's coming next.
Why the Licensing Model Had Limits
Arm's business was elegant. They designed processor architectures. Companies like Apple, Qualcomm, and Samsung licensed those designs, customised them, and built chips. Arm took royalties on every unit sold. Low capital risk, high margins, massive scale.
That model assumed clients wanted general-purpose designs they could adapt. But AI workloads aren't general-purpose. They're highly specific - optimised for matrix multiplication, low-precision arithmetic, massive parallelism. The architecture matters less than the system integration. Memory bandwidth, interconnects, power delivery - the chip is part of a larger system, not the product itself.
When your clients are building systems, not just chips, licensing an architecture becomes less valuable. They need control over the full stack. Arm's royalty model doesn't capture that value. So Arm's moving up the stack too - selling complete solutions, not just blueprints.
What Arm's Actually Selling
Arm's new CPU isn't competing with Intel or AMD in the traditional sense. It's targeting cloud infrastructure - data centres running AI inference at scale. The pitch is efficiency. Arm's architecture uses less power per operation than x86. In a data centre, power is the largest operational cost after hardware. Small efficiency gains multiply across millions of servers.
But selling chips directly changes Arm's relationship with their ecosystem. Apple and Qualcomm are now potential competitors, not just customers. That's risky. Arm's betting the market's big enough for both models - licensing for some use cases, direct sales for others. Whether that's true depends on how their clients react.
Thompson's insight is that Arm's constraints forced this move. Licensing revenue is capped by the number of chips clients build. Direct sales let Arm capture more value per unit, but require manufacturing partnerships, supply chain management, and customer support. Different business, different risks.
The Bigger Pattern - Integration Over Modularity
Arm's shift reflects a broader trend. For decades, computing got more modular. Separate companies handled design, manufacturing, software, integration. That modularity enabled innovation - you could swap components without rebuilding the whole system.
AI workloads are reversing that trend. Performance comes from tight integration between silicon, software, and system design. Apple's M-series chips work because Apple controls the hardware, the OS, and the applications. NVIDIA's dominance in AI comes from their integrated stack - chips, drivers, CUDA, libraries. You can't just drop in a competitor's GPU and expect the same results.
Arm's realising that selling architecture in isolation isn't enough. They need to sell systems. That's a harder business - more capital intensive, more competitive, more exposed to manufacturing risk. But it's where the value is moving.
What This Means for Builders
If you're building on cloud infrastructure, Arm-based instances are already available from AWS, Google, and Azure. They're cheaper to run for certain workloads - web servers, containerised apps, inference tasks. The cost difference is real, but the tooling and ecosystem are still maturing. You'll hit more edge cases than with x86.
For hardware companies, Arm's move is a signal. The era of clean separation between design and manufacturing is eroding. Vertical integration is coming back, and it's being driven by the economics of AI workloads. If your business model assumes modularity, that assumption is getting riskier.
Arm's not abandoning licensing. They're adding a second model. But the fact they felt they had to tells you something about where the market's heading. Computing is consolidating again. The question is who controls the stack.