Amazon just told Wall Street it might start shipping AI chips to customers who aren't AWS users, which is roughly equivalent to Netflix announcing it's going to sell DVDs.
The Summary
- Amazon disclosed $225 billion in revenue commitments for its Trainium AI chips and says it could start selling physical chip racks to external customers within two years, a first-ever timeline for the initiative.
- CEO Andy Jassy confirmed this business model shift during the Q1 earnings call, moving beyond his vague "in the future" language from the April shareholder letter.
- AWS revenue estimates hit $36.68 billion for Q1, with Amazon planning $200 billion in capex this year, almost entirely aimed at AI infrastructure.
- The move puts Amazon in direct competition with Nvidia, which currently supplies GPUs for AWS cloud services, creating what one might call a commercially awkward relationship.
The Signal
Amazon is about to do something it has never done before: sell computing hardware as a standalone product, not wrapped in a cloud service. The company revealed during its Q1 earnings call that Trainium AI chips now carry $225 billion in revenue commitments, and within two years, those chips could ship as physical rack-mounted systems to customers outside AWS. This isn't cloud credits or virtual instances. This is boxes of silicon that leave Amazon facilities and get bolted into someone else's data center.
The shift matters because it rewrites Amazon's core business logic. Since AWS launched in 2006, Amazon has monetized infrastructure by renting it, not selling it. You don't buy an EC2 server. You rent compute time. That model built a $100 billion annual revenue business. Now Amazon is saying the AI chip demand is so extreme that the rental model isn't enough.
"We have such demand from various companies who will consume as much as we make," Jassy said on the call.
The $225 billion commitment figure tells you how serious the demand is. That's not speculative interest. That's contracted revenue, likely spanning multiple years, representing companies who've bet their AI roadmaps on Amazon's chips instead of Nvidia's H100s or upcoming Blackwells. Amazon is planning $200 billion in capital expenditures this year, nearly all of it aimed at AI infrastructure. For context, that's more than Meta and Microsoft's AI infrastructure budgets combined for 2025.
The Nvidia angle is the real story beneath the story. Amazon still relies on Nvidia GPUs to power much of AWS's AI workload today. Customers rent Nvidia-powered instances when they need raw GPU horsepower. But Jassy took a rare public shot at Nvidia in his April shareholder letter, positioning Trainium as the smarter, more cost-efficient choice. Now Amazon is saying it might sell chips directly, which puts it in competition with Nvidia's data center business while still being Nvidia's customer for AWS.
That's not sustainable long-term. If Trainium takes off, Amazon either stops buying Nvidia chips or keeps buying them to serve a shrinking cohort of customers who specifically want CUDA compatibility. Either way, the relationship shifts from partnership to competition. Nvidia, for its part, has spent two years building software moats (CUDA, NIM, cuDNN) precisely to prevent cloud providers from commoditizing AI chips. Amazon is testing whether $225 billion in commitments is enough leverage to break those moats.
Here's what the chip sale timeline means:
- Amazon moves from cloud-only to hardware vendor within 24 months
- Direct competition with Nvidia's data center GPU sales, not just cloud services
- Potential margin expansion if Trainium costs less to produce than buying Nvidia chips
Wall Street analysts are watching AWS revenue growth closely, with Q1 estimates at $36.68 billion and expected year-over-year growth of 25.7%. That growth rate holds steady even as Amazon pours capital into Trainium development, which suggests AWS isn't cannibalizing existing revenue to fund the chip pivot. It's expanding the total addressable market by offering a new product category.
The Implication
If you're building AI infrastructure, watch what Amazon does with Trainium pricing and availability in 2027. If they undercut Nvidia on price-per-FLOP for training workloads, the entire AI chip market reprices. If they can't, Trainium becomes a niche product for AWS-native workloads, and Nvidia's dominance extends another hardware generation.
For enterprises already committed to AWS, the question is whether to wait for Trainium racks or lock in Nvidia GPU capacity now. $225 billion in commitments suggests a lot of companies have already made that bet. If you haven't, the next 18 months will clarify whether Amazon can actually deliver on that timeline and at what cost. The business model shift is real. The execution risk is still unknown.