Amazon revealed in its annual shareholder letter that it could begin selling custom AI chips to third parties, a move that would pit the company directly against Nvidia and AMD in the infrastructure market.
Jassy wrote that demand for Amazon's chips is so strong that the company could potentially "sell racks of them to future third parties." AWS already has an AI revenue run rate exceeding $15 billion as of Q1 2026. Amazon's broader chip business runs at $20 billion annually and is growing at triple-digit percentages year over year.
The significance lies not in the chips themselves but in what they represent: Amazon eliminating middlemen in its infrastructure stack.
The capacity constraint
Amazon faces a fundamental problem. Two large customers requested all of its Graviton CPU capacity in 2026 but could not obtain it because AWS must serve other customers. Trainium2 chips are sold out. Trainium3 is nearly fully subscribed after starting shipments at the start of 2026.
The company added 3.9 gigawatts of AI infrastructure capacity in 2025 and plans to double that by 2027. Even at that pace, Amazon cannot meet internal demand, let alone sell to external customers.
But the constraint is temporary. It is a capacity problem, not a capability problem. Amazon has the capital, the engineering talent and the manufacturing partnerships to scale. The question is not whether Amazon can build more chips. It is whether it will.
Why this threatens Nvidia
Nvidia's market position rests on being the only realistic option for AI infrastructure. Companies buy Nvidia GPUs because they have no alternative. Amazon offering custom silicon at better economics undermines that monopoly.
Jassy claimed Amazon's Trainium chips save "tens of billions of capex dollars per year" compared with relying on others' processors for inference. If that claim is accurate, large enterprises will eventually demand Amazon's chips. Not because they are superior in raw performance, but because they are cheaper.
The broader strategy
Selling chips to third parties is not Amazon's main goal. The main goal is reducing AWS costs and improving margins. Selling to external customers is secondary, a way to amortize development costs and maintain utilization when AWS demand fluctuates.
But it signals Amazon's long-term strategy: control every layer of the stack. Custom chips. Custom satellites. Custom robots. Custom everything. Vertical integration at the scale only Amazon can execute.
Nvidia built its position by being indispensable. Amazon is now building a position by being unnecessary. The company is constructing an alternative to every critical infrastructure component Nvidia dominates.
The chip market remains Nvidia's to lose. But Amazon has both the capital and the incentive to try
