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

Amazon said it will take delivery of 1 million Nvidia GPUs by the end of 2027, a deal that should generate tens of billions of dollars for Nvidia over the next two years. However, CEO Andy Jassy also said AWS is bringing in more Trainium than Nvidia chips, with a $225 billion Trainium backlog and $364 billion total AWS backlog, signaling a growing shift toward custom AI hardware. The message is negative for Nvidia investors because major cloud customers are increasingly diversifying away from its GPUs, even as near-term demand remains strong.

Analysis

The key second-order shift is not that Amazon is still buying Nvidia; it’s that the largest buyers of AI infrastructure are becoming their own demand suppressors for Nvidia’s unit economics. Once a hyperscaler proves it can meet a majority of internal workload growth with in-house silicon, the negotiating leverage flips: Nvidia can still win premium workloads, but pricing power on the broader volume mix should erode over the next 4-8 quarters as cloud providers standardize around lower-cost internal accelerators.

This matters more than the headline GPU order because the market has been underwriting a long runway of hyperscaler capex translating into near-deterministic Nvidia revenue. The new setup introduces a split regime: near-term revenue remains supported by backlog and supply commitments, but medium-term growth becomes more dependent on enterprise AI adoption and inference intensity, where custom chips and model abstraction layers reduce Nvidia’s attach rate. That should compress the multiple even if revenue stays strong, because the market will start discounting lower terminal share and lower mix quality.

The biggest beneficiary is AWS if Trainium continues to improve price-performance faster than expected; the economic payoff is not just lower capex, but lower variable cost per inference, which can be reinvested into price cuts or margin expansion. Alphabet and Microsoft are validating the same playbook, so the risk is industry-wide: hyperscalers are turning Nvidia from a structural monopolist into a premium component supplier for only the hardest workloads. The consensus is still too focused on backlog visibility and not enough on the fact that backlog can coexist with margin compression and a lower long-duration growth profile.

Key Decisions for Investors

• Trim NVDA on strength over the next 1-3 weeks; the stock can still trade on headline order flow, but the risk/reward deteriorates as the market prices in lower 2026-2027 share of hyperscaler capex.
• Relative-value pair: long AMZN / short NVDA for 3-6 months. Thesis: AWS captures more economics per dollar of AI capex as Trainium scales, while Nvidia’s growth quality deteriorates despite continued absolute revenue growth.
• Add GOOGL on 6-12 month horizon versus NVDA. Alphabet’s TPU monetization is earlier than consensus, and the market is underestimating how quickly custom silicon can shift inference economics in cloud.
• For hedging, buy NVDA downside via 6-9 month put spreads rather than outright shorts. The catalyst path is slower than the headline suggests, but multiple compression can hit before visible revenue deceleration.
• Avoid chasing MSFT on this theme; its custom-chip strategy is more strategic than monetizable in the next two quarters, so it is a weaker direct expression than AMZN/GOOGL.