Nvidia commands 81% of the AI chip market, a grip forged over three and a half years according to IDC data. Billions pour into its GPUs. Hyperscalers like Microsoft and Amazon snap them up. Yet cracks appear. Alphabet looms largest among challengers—not AMD or Intel, but Google’s decade-long push into custom silicon. Its tensor processing units, or TPUs, started in 2015 for internal workloads. Now they scale massively. Ironwood, the seventh-generation TPU unveiled last November, delivers four times the performance per chip for training and inference over prior models. The Motley Fool spotlighted this as Nvidia’s top long-term risk.

Short punch. Google’s Axion CPUs join the fray. Arm-based, they promise twice the price-performance of Intel and AMD’s x86 chips. Ironwood pods link 9,216 chips, yielding 42.5 exaFLOPS—more than 24 times El Capitan, today’s fastest supercomputer. Bandwidth hits 9.6 Tb/s per chip. Nvidia’s Blackwell matches on raw specs like 4.5 petaFLOPS FP8 compute and 192 GB HBM.…