Amazon Could be On the Cusp of Reshaping the Artificial Intelligence (AI) Chip Market. Should Nvidia Investors be Worried?

Amazon’s AWS has used its own Trainium AI chips since 2022 and has moved to Trainium2 (2024) and Trainium3 (2025), with reports suggesting it could sell Trainium to third parties—potentially pressuring Nvidia over time. The article argues rack-level performance can be matched by stacking 144 Trainium3 chips (UltraServers) at lower cost, but also notes Nvidia’s CUDA/software moat and customer lock-in mean near-term sales may be less affected. Overall, it’s a competitive risk signal for Nvidia’s data center franchise rather than an immediate earnings shock.

Analysis

The medium-term winner is not just the hyperscalers' cloud businesses but their procurement leverage: if large buyers can source acceptable performance at lower total cost, the first-order benefit accrues to AWS/Azure/Google Cloud margins and to customers that want cheaper inference without full NVIDIA dependency. That matters most in regulated or sovereignty-driven deployments, where the buyer values control over peak performance; that is a structural tailwind for AMZN, MSFT, and GOOG/GOOGL over 6-18 months even if unit volumes are modest at first.

The second-order loser is NVIDIA's pricing power, not its near-term revenue. Today the bottleneck remains supply and CUDA migration costs, so this is a share-of-wallet story rather than a demand-collapse story. The risk to NVDA emerges only if custom silicon shifts the procurement conversation from "best GPU" to "good-enough rack economics," which could compress gross margin over multiple refresh cycles; until then, the market may be over-discounting a threat that is still mostly theoretical.

The contrarian point: consensus is likely missing how slow enterprise workload migration is. Most training stacks are sticky, but inference economics are more elastic, so the first real damage to NVDA is likely to show up in cloud gross margins and capex allocation, not headline GPU unit sales. Falsifiers for the bearish NVDA thesis are simple: sustained guide-up in data center revenue, no slowdown in order backlog, and no evidence of custom-chip adoption translating into lower NVIDIA attach rates over the next 2-3 quarters.

Key Decisions for Investors

• Keep a 6-18 month overweight basket in AMZN/MSFT/GOOG over NVDA: the trade is that custom silicon improves cloud economics and bargaining power faster than it erodes NVIDIA's top line. Best entry is on any NVDA-led pullback in semis, with a stop if NVDA re-accelerates gross margin and raises data center guidance for two consecutive quarters.
• Avoid shorting NVDA on this headline alone; use it as a watch item. The thesis only becomes actionable if we see evidence of customer migration, pricing concessions, or cloud partners explicitly displacing NVIDIA in production inference workloads.
• Relative-value idea: long AMZN / short NVDA for 3-6 months only if AWS discloses external Trainium wins or capex efficiency improves. Risk/reward is attractive because the upside for AMZN is margin expansion, while NVDA downside is capped until supply/demand normalizes.
• For tactical expression, consider a small NVDA call-spread hedge rather than outright short if already underweight semis; this preserves upside in case the market keeps rewarding CUDA scarcity while limiting bleed if the custom-chip narrative gains traction.
• Set an alert for the next three earnings cycles: any evidence that cloud providers' own chips are lowering cost per token or reducing third-party GPU purchases would be the first real catalyst to rotate out of NVDA beta and into hyperscaler beneficiaries.