Amazon confirmed it will sell Trainium AI chips directly to external customers — moving from cloud-only to merchant silicon. With $225 billion in sales commitments from clients including OpenAI and Anthropic, and costs 80% lower than Nvidia’s H100, this is the most direct challenge to Nvidia’s monopoly yet.
Why This Changes the Stack
Until now, Amazon’s chips were captive — available only through AWS. Going merchant means anyone can buy Trainium directly, without an AWS contract. This changes the competitive dynamics at Layer 3 (Silicon) of the AI Supercycle.
THE NVIDIA CHALLENGE IS NOW DIRECT
Nvidia holds 74% of the inference chip market. Amazon is saying: we can do it 80% cheaper. That’s not a competitive claim — it’s a pricing wedge. At $41B/quarter in Nvidia inference revenue, even a 10% shift to Trainium = $4B/quarter market share transfer.
THE IRONY: AMAZON AS ANTHROPIC’S CHIP SUPPLIER AND SABOTEUR
Amazon sells chips to Anthropic. Amazon’s CEO triggered the government action that shut Anthropic’s best model. And now Amazon is selling chips directly to Anthropic’s competitors. The investor-competitor-supplier triangle gets more entangled every week.
The Supercycle read: This is Layer 3 (Silicon) bifurcating — exactly what the nine-layer stack predicted. The generalist (Nvidia) sets the software stack. The customs (Amazon Trainium, Google TPU) optimize for specific workloads at lower cost. The question: does CUDA lock-in hold, or does 80% savings break it?
The Bottom Line
Amazon just moved from cloud-only chips to merchant silicon. $225B in commitments. 80% cheaper than Nvidia. OpenAI and Anthropic already signed up. Nvidia’s 74% inference monopoly faces its most credible challenge — not from a startup, but from the company with the largest cloud, the deepest pockets, and a customer list that includes every frontier AI lab. The silicon layer just bifurcated.
Source: TechCrunch









