Who makes hyperscaler arm CPUs for AI data centers

Every large cloud provider now designs at least one Arm server CPU for its own fleet, and none of them sell it. Graviton, Axion and Cobalt exist because the host CPU in an AI cluster does work the accelerator cannot — it runs the scheduler, the storage client, the network stack and the data loader that keeps the GPUs fed — and that work is predictable enough to design silicon around. Owning the design lets the operator tune core count, memory bandwidth and power for its own instance shapes instead of paying for a general- purpose part. The part sits inside the server next to the accelerator, on the same board or on a paired head node, and it competes with the merchant x86 and Arm CPUs in the parent category rather than adding to them. Its effect on the supply chain is indirect but real: each of these designs is fabricated and packaged by the same foundries and OSATs in this layer, so a hyperscaler CPU program is a claim on the same wafer and substrate capacity as everything else here. What to look for is disclosure, because there is very little of it. These parts have no price, no ticker and no unit volumes; the only public signal is what the operator says about the share of its own fleet running on them, and how quickly a new generation appears. Treat any placement here as an in-house program, not a merchant supplier relationship.