Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers - TechCrunch

This is a collective revolt against the "CUDA tax." Nvidia's 15-year moat was never silicon—it was 4 million developers trapped in an ecosystem. Google's TPU v5p + JAX and Amazon's Trainium2 + SageMaker are twin blades aimed at that soft underbelly. The real threat isn't FLOPS; it's developer migration cost falling below the switching threshold. Technically, when hyperscalers push inference below $0.05/token on proprietary silicon, CUDA's gravitational pull decays. Upstream, CoWoS advanced-packaging allocation logic shifts—Nvidia's 70%+ AI accelerator share faces structural dilution. Downstream, PyTorch's multi-backend strategy becomes the decisive battleground. This isn't export-control-driven substitution. It's margin-capture-driven de-risking: hyperscalers are internalizing the 40-60% gross margin Nvidia extracts per GPU. Nvidia's counter is predictable: NIM microservices + DGX Cloud for lock-in, H200/B200 to maintain a 2-3 generation perf gap, and selective M&A at the framework layer. AMD gets a window but its software ecosystem remains the fatal gap. 12-24 month tail: AI compute fragments into 3-4 viable stacks. "CUDA-compatible" degrades from technical fact to marketing copy. The true losers are mid-tier cloud providers who can't afford dual-silicon strategies.