The headline: Amazon Web Services' custom silicon business — Graviton, Trainium, and Nitro chips — hit an annual revenue run rate above $20bn in Q1 2026. That is triple-digit year-on-year growth. Their latest chip, Trainium3, is reportedly nearly sold out. The customers who have locked in capacity include OpenAI (two gigawatts of Trainium), Anthropic (up to five gigawatts), Meta, and Uber. For UK businesses that use ChatGPT, Claude, or any major AI platform, you are already customers of this infrastructure — indirectly. Understanding why it matters helps you make better decisions about which AI tools to rely on.
Why Amazon is building its own chips
Nvidia produces the H100 and H200 GPU chips that currently power most frontier AI training and inference. They are in chronic short supply and command premium prices. Amazon's response — like Google's (TPU) and Microsoft's (Maia) — is to build custom silicon that reduces dependence on Nvidia and brings the cost and availability of compute under internal control. Amazon's CEO Andy Jassy has stated that custom silicon could save tens of billions annually in capital expenditure over time. That saving eventually flows toward the AI tools built on top of AWS.
Trainium is specifically designed for AI model training — the computationally intensive process of building and improving AI models. Graviton handles general computing workloads. Together they give Amazon a vertically integrated stack: data centres, networking, compute, and the AI applications running on top (including AWS Bedrock, which serves Claude, Llama, Titan, and other models to enterprise customers).
What "nearly sold out" means in practice
When Trainium3 capacity is described as nearly sold out, it does not mean ordinary businesses cannot access AWS AI services tomorrow. It means the large-scale reservation contracts — gigawatt-scale capacity commitments — are already committed. The supply crunch is at the infrastructure layer, not the application layer that most UK businesses interact with.
What it does mean, practically:
- AI pricing for end-users will remain elevated in the near term. When the inputs to AI infrastructure are supply-constrained, the per-token and per-API-call costs for AI services do not drop as fast as they would in an oversupply environment.
- Providers who secured capacity early have a structural advantage. OpenAI and Anthropic have locked in gigawatts of Trainium capacity. That means their ability to serve users at scale is secured. Smaller providers who did not secure similar agreements may face constraints as demand grows.
- AWS Bedrock customers benefit from this supply security. If your business uses AWS-hosted AI services (including Claude via Bedrock), Amazon's own chip security means your supply chain is more protected than if you were depending on third-party Nvidia-only providers.
The strategic signal for UK businesses
When Amazon commits $200bn in 2026 capital expenditure and a meaningful portion goes to custom AI silicon, it is a statement about where management believes the long-term value sits. It also means that AI infrastructure is shifting from a commodity (buy Nvidia chips, build data centre, serve AI) to a strategic capability (design your own chips to reduce external dependency and improve economics at scale). The businesses best positioned to offer AI tools at sustainable prices over the next five years are the ones building this vertical integration now.
For UK small business owners, the practical implication is not about which chip to buy. It is about which AI platform to commit to. The providers who have secured their own compute supply chains — AWS (with Trainium), Google (with TPU), Microsoft (with Maia) — are structurally better positioned to maintain service levels and pricing stability as demand continues to grow. Pure API resellers who depend entirely on Nvidia availability are exposed to the supply chain in ways those providers are not.
The open-source dimension
One underreported implication of the AWS chip scale is what it means for open-source AI. Amazon has signalled interest in potentially selling Trainium chips to external companies — positioning it as a standalone business rather than just an internal cost reduction. If that happens, it creates a path for open-source AI projects and smaller AI companies to access competitive compute infrastructure outside the Nvidia ecosystem. That broadens competition, which over time reduces prices for everyone — including UK small businesses running AI automation on third-party platforms.
What UK operators should do with this information
Choose platforms with supply security: When evaluating AI tools for your business, consider whether the platform has its own infrastructure or is entirely dependent on spot Nvidia compute. AWS Bedrock (Claude, Llama), Google Vertex AI, and Azure OpenAI are structurally more stable than platforms without their own compute agreements.
Expect pricing stability, not significant near-term drops: The chip supply crunch means AI API costs will not fall as fast as some predictions suggest. Build your AI business case on current pricing, not a projected 10× cost reduction in the next 12 months.
Watch for open-source AI maturation: If Amazon sells Trainium chips externally, it creates a path for high-quality, lower-cost open-source AI alternatives. This is 12-24 months away from mattering at SMB scale, but worth monitoring.
The infrastructure signal confirms the adoption signal: Demand for AI compute growing at triple-digit pace is the most independent validation that AI adoption is real and accelerating. It is not a bubble — it is backed by hardware commitments worth hundreds of billions of pounds.