The Race to Beat Nvidia: Does Google or Amazon Have the Better In-House Silicon

This is less a near-term displacement of merchant GPUs than a re-division of economics inside the AI stack. In the next 1-3 months, the market is likely to keep rewarding hyperscalers that can turn capex into proprietary operating leverage, while still paying up for the infrastructure vendors that sit one layer lower in the build cycle. The first-order losers are not necessarily semiconductor names broadly; the more exposed segment is GPU pricing power and top-end inference/training mix, not unit demand.

The key second-order effect is that custom silicon does not reduce AI spending; it redirects it. That means HBM, advanced packaging, interconnect, and foundry capacity should remain tight even if some accelerator share migrates away from NVDA, which limits how fast a bearish NVDA thesis can work. The sharper medium-term risk is margin compression at the cloud owners if memory and supply chain costs rise faster than utilization, so the trade is really about who captures the spread between compute demand and delivered FLOPs.

Contrarian view: the market may be underestimating how long NVDA’s ecosystem moat can coexist with TPU/Trainium adoption because custom silicon rollouts are usually slower than investor narratives. The more realistic 6-18 month setup is relative share shift at the margin, not outright substitution. That argues for expressing the view through relative value rather than an outright semiconductor short, with GOOGL the cleaner near-term beneficiary and AMZN the longer-dated one if rack-scale distribution monetization keeps compounding.