Amazon seeks to offload about $8 billion of Nvidia chips - proactiveinvestors.com

Amazon's $8 billion Nvidia GPU liquidation is not a balance-sheet cleanup—it is the first visible crack in the AI compute arms race. Hyperscalers stockpiled H100s and B200s on the premise of insatiable training demand, but inference workloads are growing far slower than projected, pushing fleet utilization below the 60% breakeven threshold. The ripple effect hits upstream: when the single largest buyer starts dumping inventory, TSMC's CoWoS advanced-packaging expansion cadence in Taiwan, China faces a forced re-evaluation, and the 2025 capacity roadmap carries meaningful overbuild risk. Strategically, this move is a quiet declaration that Trainium2 has crossed the viability threshold. Nvidia's pricing monopoly, sustained by scarcity and FOMO, encounters its first credible buyer-coalition pressure. Expect Microsoft and Google to accelerate TPU and Maia internalization, while AMD seizes the 20-to-30% secondary-market price dip to poach CoreWeave-class GPU cloud tenants. Over the next 18 months, the industry pivot from 'more cards equals more power' to 'cost-per-inference-token' will compress Nvidia's data-center revenue growth from triple digits toward the 40-to-50% band. Custom ASICs crossing 25% market share is no longer a scenario—it is the base case.