Amazon hopes to challenge Nvidia more directly by selling its AI chips

Amazon says AWS is in talks to sell its Trainium AI chips to other companies, implying a potential $50 billion annual run rate if the chip business were standalone. The move would expand Amazon deeper into Nvidia’s AI hardware market, while AWS says current Trainium and Trainium4 capacity is already sold out. The article is strategically positive for Amazon’s chip ambitions but raises competitive pressure concerns for Nvidia.

Analysis

This is less a near-term displacement of Nvidia than a credible attempt to break the current monopoly economics at the margin. The key second-order effect is not unit share, but bargaining power: if AWS can sell Trainium externally, it creates an alternative procurement path for hyperscalers and enterprise buyers that currently have little leverage against Nvidia’s pricing, allocation, and software lock-in. The market should frame this as a medium-term constraint on NVDA multiple expansion rather than a near-term earnings hit, because the real inflection depends on whether AWS can prove third-party performance, software compatibility, and reliable supply at scale.

The more important risk for Amazon is self-cannibalization of its own cloud margin stack. Selling chips externally may optimize fab utilization, but it potentially sacrifices the highest-margin layer of the AWS model: once the chip leaves the walled garden, Amazon gives up the attach rates on storage, networking, security, and managed AI services that make the economics so attractive. That means the bear case for AMZN is actually not chip demand weakness; it is that management trades away some of the cloud ecosystem rent to build a lower-margin hardware franchise.

For the supply chain, the bottleneck is likely to be capacity allocation rather than design. Any meaningful third-party Trainium ramp would have to compete with internal AWS demand for wafers and advanced packaging, which keeps TSM in the center of the story even if the headline feels negative for NVDA. Consensus is probably underestimating how slowly this can matter: enterprise qualification cycles for AI infrastructure are measured in quarters, and software porting/integration usually takes 6-12 months. So the setup is more of a 6-18 month competitive overhang on Nvidia than an immediate revenue threat.

Key Decisions for Investors

• Maintain a tactical short NVDA / long AMZN pair for 3-6 months: thesis is multiple compression on NVDA as competition becomes more credible, while AMZN retains upside from optionality; risk is NVDA re-accelerates revenue or AWS signals no external supply.
• Buy NVDA downside structures into strength: use 6-9 month put spreads to express a capped downside view, since the near-term earnings base remains intact but sentiment may fade as Trainium gains mindshare.
• Stay long TSM on any weakness for 6-12 months: even if the narrative sounds anti-Nvidia, the practical outcome is more AI silicon demand and tighter advanced packaging utilization; risk/reward favors a buy-the-dip stance unless AWS shifts meaningful volume away from TSM.
• Do not chase INTC/AMD as immediate winners: they benefit only if AWS proves third-party distribution and software adoption, which is a slower catalyst; better expressed via optionality after evidence of customer wins.
• Watch for AWS supply commentary at the next earnings cycle; if Amazon confirms external rack sales plus incremental manufacturing capacity, add to the AMZN relative-long view, because that would validate a new revenue stream with limited near-term P&L dilution.