Amazon's Silicon Is Now a Business, Not a Project

The Number That Changes the Conversation

Amazon's chip business has crossed a threshold. The company now has an annual revenue run rate over $20 billion, growing at triple-digit percentages year over year, disclosed by CEO Andy Jassy in his April 2026 shareholder letter. This is not a footnote.

The portfolio spans three chip families: Graviton CPUs, Trainium AI accelerators, and Nitro networking controllers. Each targets a different layer of the stack. The vertical integration they represent cannot be replicated by independent chip vendors working from outside.

The $50B figure Jassy mentioned circulates widely but warrants skepticism. Amazon's hypothetical standalone chip valuation of approximately $50 billion assumes external third-party sales that have not yet materialized. The $20B run rate is the real metric. The $50B is speculative.

Why the Economics Work Now

Custom silicon was always theoretically sensible for hyperscalers. AI workloads made it irresistible at AWS scale.

Jassy argues Trainium will "save us tens of billions of capex dollars per year, and provide several hundred basis points of operating margin advantage versus relying on others' chips for inference." This matters within a $200 billion capital expenditure plan—the largest single-year spend among major technology companies. The chip program is not a hedge. It is the foundation.

Performance deltas tell the story. Trainium2 instances cost roughly half the price of comparable Nvidia H100 instances while delivering competitive performance for many workloads. At inference scale, that cost difference moves the needle materially. Early results from customers testing Trainium3 show AI training and inference costs reduced by up to 50%.

The roadmap is accelerating. Trainium3, launched in December 2025, is AWS's first 3nm AI chip, delivering 2.52 petaflops of FP8 compute per chip with 144 GB HBM3e memory and 4.9 TB/s bandwidth. Graviton5, with 192 cores at 3nm and a five-times larger L3 cache than its predecessor, now powers more than half of all new CPU capacity added to AWS for the third consecutive year. These numbers describe production chips, not ambitions.

The Nvidia Relationship

Amazon is not displacing Nvidia wholesale. Jassy acknowledges a "strong partnership with NVIDIA" and notes that "virtually all AI thus far has been done on NVIDIA chips, but a new shift has started." The language is deliberate—validating the installed base while reframing Nvidia as yesterday's default.

Nvidia enters 2026 from strength. The company reported revenue of $68.1 billion in Q4 2025, a 73% year-on-year increase, with dominant market share intact. Amazon is not contesting that merchant market. It is eroding Nvidia's position inside AWS, one workload at a time.

Trainium4 incorporates NVLink Fusion, allowing customers to pair Trainium accelerators with Nvidia GPUs within the same system. This is pragmatic. Amazon captures new workloads on Trainium while letting enterprises preserve existing Nvidia investments. No forced write-offs. No bridge-burning.

Anchor Customers and Scale

The Anthropic deal shows how custom silicon becomes a competitive moat. Project Rainier, activated in October 2025, deploys nearly 500,000 Trainium2 chips across a 1,200-acre Indiana facility dedicated exclusively to training Anthropic's Claude models, providing five times the compute power Anthropic used for previous Claude versions.

OpenAI has joined the same dynamic. OpenAI's expanded partnership includes consuming approximately two gigawatts of Trainium capacity through AWS infrastructure. Two of the world's most influential AI labs are now validating Amazon's silicon at production scale—a reference account that startups cannot assemble.

Amazon Bedrock processed more tokens in Q1 2026 than in all prior years combined, with customer spending growing 170% quarter over quarter. Each token processed increasingly runs on Amazon silicon.

For Infrastructure Teams

Choosing AWS now means choosing into a silicon roadmap Amazon controls entirely: architecture, pricing, upgrade cycles, and the Neuron SDK software layer.

The AWS Neuron SDK enables standard ML frameworks to run on Trainium and Inferentia, with 2025 releases substantially improving developer experience. Workload portability was already difficult; tighter SDK integration with PyTorch and Amazon's own models makes it harder still.

For inference-heavy workloads, the economics favor migration. Above $100,000 per month in compute spend, the 40-50% cost reduction is material. Below that, the switching cost likely outweighs the savings. Know which side your workloads fall on before your next budget cycle.

Google has TPUs. Microsoft has Maia and Cobalt. Neither has disclosed a $20B run rate. If both eventually do, custom silicon becomes mandatory across the industry. If Amazon stands alone at this scale, it points to an execution gap—and a narrowing window for enterprises that want to preserve multi-cloud optionality on AI workloads.

What Matters Next

1. Google and Microsoft disclosures: Watch Q2 and Q3 2026 earnings for custom silicon revenue figures. Comparable numbers would signal industry convergence; silence would suggest Amazon's lead is widening.

2. Third-party Trainium sales: External Trainium chip sales are identified as a key Amazon opportunity. The first announcement of customers buying chips outside AWS would mark a new competitive phase.

3. Trainium in enterprise RFPs: When companies stop treating Trainium compatibility as a nice-to-have and start requiring it, workload lock-in has become structural.

4. AI chip startup trajectories: Hyperscalers at $20B scales shrink the addressable market for independent accelerator companies. Track whether Series B and C funding in this space slows or pivots toward acquisition.

5. Margin attribution: AWS generated $14.2 billion in operating income on $37.6 billion in revenue in Q1 2026. Watch whether Amazon explicitly credits custom silicon for margin gains—that disclosure would let operators quantify the advantage precisely.