Amazon CEO Andy Jassy Has Good News and Bad News for Nvidia Investors

Amazon agreed to buy 1 million Nvidia GPUs by the end of 2027, a deal that could generate tens of billions of dollars in Nvidia revenue. However, CEO Andy Jassy said Amazon is bringing in more Trainium chips than Nvidia chips, with a $225 billion Trainium backlog and a broader shift toward custom AI accelerators that could pressure Nvidia's long-term growth. The article frames this as a growing competitive threat to Nvidia's market position as hyperscalers increasingly favor in-house silicon.

Analysis

The market is starting to price a structural shift from a single-vendor AI stack to a multi-architecture procurement model. That matters because the biggest buyers are also the best distribution channels: once AWS, Google Cloud, and Azure normalize custom accelerators for mainstream workloads, Nvidia loses not just unit volume but the default-choice status that has supported premium pricing and attach rates. The second-order effect is that the margin pool migrates from merchant silicon to vertically integrated cloud platforms, which can monetize lower inference costs through higher utilization, stickier workloads, and better gross margin per token.

The near-term winner is AMZN, not just because it buys fewer dollars of external GPUs over time, but because lower internal compute cost expands the addressable market for inference-heavy customers and improves AWS pricing flexibility. That creates a flywheel: cheaper compute drives more usage, which drives more Trainium deployment, which improves software support and chip learning curves. For NVDA, the risk is less about one customer and more about the cumulative purchasing behavior of all hyperscalers; if the buying mix keeps shifting toward in-house silicon over the next 4-8 quarters, consensus estimates for datacenter growth may prove too high even if absolute AI capex stays strong.

The contrarian view is that Nvidia's moat may be shrinking at the margin while still remaining dominant at the frontier. Custom chips are likely to take share first in inference and standardized training, but bleeding-edge model training still favors Nvidia's ecosystem, networking, and developer tooling. That implies the stock may be vulnerable less to a crash than to a de-rating: revenue can keep growing while the multiple compresses as investors realize the TAM is being shared with customers rather than captured exclusively by Nvidia.

The most important catalyst to watch is procurement mix over the next two earnings cycles. If Trainium/TPU/Maia adoption keeps compounding faster than external GPU spending, the market will likely shift from debating 'share gain' to modeling 'share stabilization,' which is a very different outcome for NVDA's forward multiple. Conversely, any evidence that custom chips are constrained by software maturity, model portability, or yield issues would quickly re-open upside for Nvidia into 2026.

Key Decisions for Investors

• Reduce NVDA exposure on strength over the next 2-4 weeks; the risk/reward is now skewed to multiple compression if hyperscaler capex keeps shifting to internal silicon rather than external GPUs.
• Go long AMZN vs short NVDA in equal dollar terms for a 3-6 month horizon; AWS benefits from lower cost of compute and the market is underestimating how much custom silicon can widen cloud margin structure.
• Add GOOGL on pullbacks as a relative winner in custom-accelerator adoption; upside comes from TPU monetization and differentiated cloud economics, while downside is partially cushioned by the broader search/ads franchise.
• Use MSFT as a lower-beta hedge within the AI stack: long MSFT against short NVDA for 6-9 months captures the shift toward internally optimized compute while reducing single-name volatility.
• If you want convexity, buy NVDA put spreads 6-9 months out rather than outright puts; the thesis is slower share erosion, not an immediate collapse, so defined-risk downside is better than paying up for near-dated gamma.