Amazon just dropped some wild specs on their Trainium 3 setup. These UltraServers can pack 144 chips into one box—that's insane density. But here's where it gets crazy: their EC2 UltraClusters 3.0 can link up thousands of these servers, supporting up to a million Trainium chips total. That's literally 10X what the last generation could handle.

This kind of scaling matters beyond just cloud computing. With AI training and crypto mining both hungry for compute power, we're watching infrastructure wars heat up. Companies like Broadcom, NVIDIA, Marvell, AMD, and Google are all racing in this space. Anyone building on-chain AI or running heavy computation might want to keep tabs on how accessible this level of horsepower becomes.