Amazon guided to roughly $105B in 2025 capital expenditure — the largest of any company on earth — and the majority of it goes to AWS AI infrastructure.
That's the short answer. The longer answer is more interesting, because Amazon is not just buying GPUs and pouring concrete. It is running a two-track bet: rent Nvidia's best chips to keep customers happy today, while designing its own Trainium and Inferentia silicon to own the cost curve tomorrow. The capex number is the headline. The chip strategy is the actual story.
Amazon AWS AI capex 2025: the numbers
Amazon's AWS AI capex in 2025 totals roughly $105B in companywide capital expenditure, the highest of any hyperscaler, with the large majority directed at AWS data centers, networking, and AI compute. That is up from about $83B in 2024 and roughly $48B in 2023 — a near doubling in two years, driven almost entirely by generative AI demand rather than retail fulfillment, which was the historical driver of Amazon's capex line.
| Year | Total Capex | YoY Change | Primary Driver |
|---|---|---|---|
| 2021 | ~$55B | — | Retail fulfillment + logistics |
| 2022 | ~$59B | +7% | Fulfillment, early AWS buildout |
| 2023 | ~$48B | −19% | Capex discipline, retail pullback |
| 2024 | ~$83B | +73% | AWS + generative AI infrastructure |
| 2025E | ~$105B | +27% | AWS AI: chips, data centers, power |
| 2026E | $115B+ | +10% | Continued AI capacity expansion |
Figures are approximate, blending reported actuals and management guidance. The shape matters more than the decimal: capex roughly doubled from 2023 to 2025, and the marginal dollar is now AI, not fulfillment. Track the full hyperscaler picture on the AI Spending Dashboard.
Where the ~$80B in AI capex actually goes
When Amazon says "the majority" of $105B supports AWS, the AI-attributable slice is roughly $75–85B. It breaks down across four buckets, and only one of them is the GPUs everyone talks about.
AI compute (GPUs + custom silicon)
~45%Nvidia H100/H200 plus Amazon's own Trainium2 and Inferentia2 — the single largest line
Data center shells + power
~30%New regions, multi-gigawatt sites, and long-lead electrical infrastructure
Networking + interconnect
~15%High-bandwidth fabric to cluster hundreds of thousands of accelerators
Storage + general AWS capacity
~10%Object storage, databases, and non-AI cloud growth
The reason power keeps showing up in capex calls is that GPUs are useless without electricity to run them and water or air to cool them. A single modern AI data center campus can draw 500MW to 1GW — enough for a mid-size city. Amazon has signed nuclear, solar, and grid deals to lock in that power, including a multi-year agreement tied to a Pennsylvania nuclear plant. The constraint on AI in 2026 is increasingly megawatts, not chips.
The custom chip strategy: why Amazon builds Trainium and Inferentia
Here is the part of Amazon's AWS AI capex that separates it from a company simply buying Nvidia. Amazon designs two families of custom silicon: Trainium for training and Inferentia for inference. The logic is brutal and simple — Nvidia's data center GPUs carry gross margins north of 70%, so every dollar AWS spends on an H200 is roughly 70 cents of Nvidia profit. Designing your own chip claws that margin back.
| Chip | Purpose | Status (2026) | Claimed advantage |
|---|---|---|---|
| Trainium2 | Model training | Volume production | 30–40% better price-performance vs comparable GPU |
| Trainium3 | Model training | Announced / ramping | ~2x performance, ~40% more efficient than Trn2 |
| Inferentia2 | Inference serving | Generally available | Up to 40% better price-performance for inference |
| Graviton4 | General CPU compute | Generally available | Powers non-GPU AWS workloads at lower cost |
| Nvidia H200 | Training + inference | Purchased at scale | Industry-standard, CUDA ecosystem |
| Nvidia GB200 | Frontier training | Deployed in clusters | Highest raw performance, premium price |
The strategic point is not that Trainium beats Nvidia on raw performance — it usually doesn't. The point is total cost of ownership. If AWS can serve a customer's inference workload on Inferentia at 60% of the cost of an Nvidia instance, AWS either keeps that margin or passes the savings on to win the deal. Amazon doesn't need Trainium to be the fastest chip; it needs it to be good enough at a structurally lower cost, and to give AWS leverage in every Nvidia negotiation.
The Anthropic deal and Project Rainier: Amazon's anchor customer
A custom chip is worthless without a flagship customer willing to bet a frontier model on it. That customer is Anthropic. Amazon has committed roughly $8B to Anthropic across two tranches, making AWS Anthropic's primary training and cloud partner. In exchange, Anthropic trains on Trainium at massive scale through Project Rainier — a cluster reported to involve several hundred thousand Trainium2 chips, one of the largest AI training deployments in the world.
This is a circular-economy bet that the whole industry now runs on: Amazon invests in Anthropic, Anthropic spends that capital on AWS compute, and AWS validates its custom silicon against a real frontier workload. Critics call it round-tripping revenue. Amazon calls it anchoring demand. Both are true. What matters is that Trainium now has a customer that forces it to actually work — and Anthropic's Claude models, available through Amazon Bedrock, give AWS enterprise customers a top-tier model without building one in-house.
For more on how the model layer is priced, see our breakdown of AI company valuations and how the largest private AI labs justify their numbers.
Amazon AWS AI capex vs Microsoft, Google, and Meta
Amazon's ~$105B leads the hyperscaler pack in 2025, but the gap is narrower than the headline suggests because the others are catching up fast. Here is how the four largest spenders compare.
| Company | 2025 Capex (est.) | Custom AI silicon | Flagship AI customer/product |
|---|---|---|---|
| Amazon (AWS) | ~$105B | Trainium, Inferentia | Anthropic / Bedrock |
| Alphabet (Google) | ~$85B | TPU v6/v7 | Gemini / Vertex AI |
| Microsoft | ~$80B+ | Maia 100 | OpenAI / Copilot |
| Meta | ~$66–72B | MTIA | Llama (internal + open) |
| Oracle | ~$25B+ | None (Nvidia-heavy) | OCI / OpenAI compute |
| Combined top 4 | $330B+ | — | — |
The common thread: every hyperscaler now designs its own AI chip. Google has TPUs (the most mature, now in their sixth-plus generation), Microsoft has Maia, Meta has MTIA, and Amazon has Trainium. Nvidia still wins the raw-performance crown and the developer ecosystem, but the buyers are systematically building escape hatches. See the side-by-side in our Big Tech Earnings Dashboard.
The bull and bear case on Amazon's AI capex
Bull Case
- ✓ AWS runs a ~30–38% operating margin and funds capex from cash flow
- ✓ Trainium/Inferentia claw back Nvidia's 70%+ margins on custom workloads
- ✓ The $8B Anthropic deal anchors real frontier-model demand
- ✓ AWS is the #1 cloud, so AI demand lands on existing distribution
- ✓ Owning the chip stack lowers long-run cost per token structurally
Bear Case
- ✕ $105B in capex means rising depreciation pressuring AWS margins by 2027
- ✕ Trainium adoption outside Anthropic is still unproven at scale
- ✕ Nvidia's CUDA lock-in makes custom silicon a hard sell to customers
- ✕ The Anthropic loop risks looking like circular revenue if growth slows
- ✕ Power and grid constraints could strand capacity that's already paid for
The $105B capex number is the headline. The chip is the strategy.
Amazon isn't just renting the AI boom from Nvidia — it's spending $105B to own the layer underneath it, one Trainium chip at a time.
Track real-time AI infrastructure and capex data on the AI Spending Dashboard, AI Valuations, and Big Tech Earnings Dashboard at Value Add VC. See also: Microsoft $80B AI Capex and The SoftBank–OpenAI Stargate Deal.