Forget AMD: Amazon Declares War on Nvidia by Selling Its Own AI Chips
1 min read
The story
Amazon has moved beyond using its custom Trainium and Inferentia chips internally and is now offering them to external cloud customers, marking a direct commercial challenge to Nvidia's dominance in AI accelerator silicon. This is part of a broader hyperscaler trend — Google's TPUs and Microsoft's Maia are similarly in-house — but Amazon's AWS scale gives its silicon push real commercial weight. The headline framing ('declares war') overstates the near-term threat: Nvidia just posted $215.9B in revenue growing 65.5% YoY with 71% gross margins, numbers that reflect a supply-constrained market where demand for H100/H200/B200 far exceeds alternatives.
The second-order question is whether hyperscaler custom silicon gradually commoditizes the workloads at the margin — inference especially — reducing Nvidia's addressable TAM over a multi-year horizon even if training demand holds. Watch for AWS re:Invent announcements on Trainium2 adoption rates, any enterprise customer wins Amazon discloses, and Nvidia's next datacenter revenue print for signs of softening mix. The AMD angle (displaced by Amazon in the 'alternative to Nvidia' narrative) is the sleeper risk the headline hints at.
The case — both sides
AMZN's move into external AI chip sales diversifies AWS revenue streams and could structurally reduce its own Nvidia procurement costs while opening a new silicon-as-a-service revenue line — with $716.9B in revenue and 10.8% net margins, AWS infrastructure leverage makes this a credible long-term margin driver.
Nvidia's 71.1% gross margin and 65.5% YoY revenue growth reflect a CUDA software moat and enterprise lock-in that custom silicon from hyperscalers has repeatedly failed to dent at scale, and AMD's own purpose-built AI GPU effort remains a more direct competitive comparison than Amazon's nascent external chip business.
The house read
Two-sidedThe question for NVDA and AMD is whether Amazon's decision to sell Trainium/Inferentia externally marks the beginning of meaningful TAM erosion for third-party AI silicon, or whether Nvidia's entrenched CUDA ecosystem and explosive growth make custom-chip competition a distant multi-year risk.
Wrong ifNvidia's Blackwell demand remains severely supply-constrained; any analyst or management comment reaffirming backlog at the next earnings print collapses the short NVDA leg. Amazon's chip business could also remain a rounding error in AWS revenue for years, making the 'war' framing premature.
Published read · research, not advice