Arm just stopped being Switzerland and the entire chip industry is pretending this is fine.
The Summary
- Arm announced it will start selling its own chips, shifting from pure licensing to direct competition with its customers
- Investors like Liontrust's Clare Pleydell-Bouverie are cheering the move, betting Arm can capture more AI chip margin
- The pivot puts Arm in direct tension with licensees like Qualcomm, Apple, and Samsung—companies that built empires on Arm's neutral stance
The Signal
For thirty years, Arm made money by staying out of the arena. They designed the instruction set, licensed it to everyone, and watched the money roll in while chip makers beat each other bloody. That neutrality was the entire value proposition. You could be Samsung or Apple or Qualcomm, all sworn enemies, and still use Arm's designs because Arm wasn't competing for the same margin dollars.
Now they're jumping in. The investor enthusiasm from firms like Liontrust suggests the market believes Arm can have it both ways: keep licensing revenue while capturing the fat margins of actual chip sales. That's the dream. The reality is messier.
“You can’t be the referee and play forward at the same time without someone calling foul.”
Here's what makes this interesting for the agent economy:
- AI inference workloads are moving to edge devices—phones, cars, robots—where Arm architecture dominates - Nvidia and AMD own datacenter training, but inference is fragmenting across billions of devices - If Arm can ship custom silicon optimized for on-device AI agents, they're positioned at the chokepoint of Web4 infrastructure
The risk? Alienating the licensees who made Arm ubiquitous. Apple designs its own chips. Qualcomm just spent years fighting Arm in court over licensing terms. Samsung has foundry ambitions. None of these companies will be thrilled to see Arm become a direct competitor, especially in AI where the margin pools are largest.
The Implication
Watch how Arm's biggest licensees respond over the next two quarters. If Apple or Qualcomm start exploring alternative architectures more seriously—RISC-V keeps looking better—that's your signal that Arm misjudged the trade-off. If they keep licensing and Arm's chip sales grow without licensee defection, then Pleydell-Bouverie and the bulls were right: Arm can thread the needle.
For builders in the agent space, this matters because chip architecture determines what's possible at the edge. Faster, cheaper inference on-device means agents that run locally, not in the cloud. That's the difference between agents that work everywhere and agents that only work when you have Wi-Fi.