Amazon wants your AI stack, your CPU fleet, and maybe your network link too

Andy Jassy’s annual shareholder letter is meant for investors. This year, it also reads like a broad challenge to the infrastructure market.

Amazon says it plans to spend $200 billion in capex in 2026, and Jassy uses the letter to defend that number with three points: AWS custom chips are already a huge business, demand is ahead of supply, and Amazon’s next expansion goes beyond data centers. It reaches into low Earth orbit too.

For developers and infrastructure teams, the interesting part is simple. AWS is trying to own more of the stack at once: training silicon, inference silicon, general-purpose CPUs, cluster networking, and eventually satellite connectivity. That has obvious implications for cost, portability, and how much of your software ends up tuned around one cloud’s hardware and software assumptions.

The numbers are hard to ignore

Jassy makes a few claims worth taking seriously.

He says Amazon’s custom silicon business, across Trainium, Inferentia, and Graviton, is already running at a $20 billion annual revenue rate. He also argues that if AWS sold chips the way a conventional semiconductor vendor does, it would look like a $50 billion ARR chip company.

That framing is convenient. AWS doesn’t sell chips in boxes, and cloud revenue isn’t chip revenue. Still, the point is clear enough. Amazon wants its silicon efforts treated as a major platform shift, not a side project in vertical integration.

The demand signals are aggressive too. According to Jassy, Trainium3 capacity is nearly sold out, and Trainium4, still roughly 18 months away, is already mostly committed. On the CPU side, he says 98% of the top 1,000 EC2 customers use Graviton, and that two enterprises asked to buy all available Graviton capacity for 2026.

That last anecdote feels polished. It also fits the larger trend. ARM in the cloud has moved past the trial phase. At AWS scale, it’s already standard for a lot of general compute.

Trainium looks real now, with limits

The sharpest claim in the letter is the one Nvidia will care about most: Amazon thinks the market is ready to move a meaningful slice of AI training off CUDA-first hardware.

Nvidia isn’t about to lose its grip overnight. But AWS clearly sees a large set of workloads where price-performance matters more than first-day access to every new kernel, attention variant, and framework optimization.

That’s where Trainium sits.

The technical pitch is familiar. Trainium is built for large-scale model training, with mixed-precision support such as bfloat16 and FP16, high-bandwidth interconnect inside the node, and AWS’s Elastic Fabric Adapter across nodes for collective operations like all-reduce. The software layer is Neuron, which compiles models from PyTorch, TensorFlow, and JAX, often through PyTorch XLA, and turns model graphs into something the hardware can execute efficiently.

When the workload fits, the economics can be compelling. If your model uses common operators, standard attention patterns, and fairly normal parallelism, Trainium can deliver strong throughput per dollar. AWS has pushed that line for years. Now it’s pointing to supply pressure as proof that customers are actually buying in.

The limitation hasn’t changed. It’s software.

CUDA still has years of ecosystem weight behind it. Custom kernels, unusual attention variants, experimental optimizers, low-level profiling tools, and third-party libraries usually land there first. On Trainium, you’re working inside the limits of the Neuron compiler and its operator support matrix. If your stack depends on unsupported ops or heavily tuned CUDA code, porting is real work. Sometimes a lot of it.

For some teams, that trade-off is fine. Especially teams running large, repeatable training pipelines where infrastructure cost matters more than chasing every fresh research trick.

Graviton has already taken a big chunk of the market

Trainium gets the attention because AI spending is ridiculous right now. Over time, Graviton may matter more.

AWS says 98% of its top 1,000 EC2 customers use Graviton. Even if that includes partial adoption rather than broad migration, the message is clear. x86 is no longer the default answer for general cloud compute.

That shift has been building for years, and the reason is pretty mundane: if most of your services are in Go, Java, Rust, Python, or modern Node runtimes, moving to arm64 is often manageable. Recompile, retest, then look at the bill. For a lot of teams, the savings are hard to dismiss.

AWS’s edge here isn’t just the CPU itself. It’s the system around it. Nitro offloads networking and storage functions, which cuts host overhead and improves consistency. That matters in production. Tail latency and noisy-neighbor problems are still what wreck the nice benchmark slide.

The catch is also familiar. Architecture migration tends to expose all the messy dependencies you forgot about: native extensions, x86-specific SIMD assumptions, vendor binaries, CI images, odd internal tools. Graviton adoption usually comes down to software hygiene as much as processor performance.

A lot of companies are finally in decent shape there.

The capex number is really about power, land, and supply

A $200 billion infrastructure spend sounds absurd until you stop thinking only about chips.

AI capacity is constrained by land, power, substations, transformers, water, and fiber almost as much as by accelerators. Jassy is effectively saying Amazon will spend at that level because demand is already there, and the bottleneck is physical buildout.

He points to a reported OpenAI commitment to spend $100 billion on AWS capacity as evidence that the spending is backed by customers. It’s fair to be skeptical about how much of that lands on schedule. Big AI infrastructure commitments tend to shift once deployment reality kicks in. But the exact figure matters less than Amazon’s posture. AWS is building on the assumption that hyperscale AI demand stays strong and supply stays tight.

That’s a large bet. It’s also a rational one if you believe customers will accept a more mixed hardware stack in exchange for capacity and lower cost.

The satellite part fits the rest of the strategy

Jassy also highlighted Amazon Leo, the company’s low Earth orbit network, slated for mid-2026. He says it already has contracts with Delta, AT&T, Vodafone, Australia’s NBN, and NASA.

That puts Amazon in direct competition with Starlink, especially in aviation and telecom backhaul. For AWS customers, the interesting part is the integration path into Amazon’s cloud and network services.

LEO systems generally offer much lower latency than geostationary satellites, often in the 30 to 70 ms round-trip range to ground. That’s workable for a lot of enterprise and mobility use cases, assuming routing, peering, and traffic management are predictable.

If Amazon ties Leo into AWS edge locations, CloudFront, Global Accelerator, and managed private connectivity, it can sell a much broader package: a controlled path from remote users or aircraft into AWS services, with one billing relationship and one account team.

Plenty of enterprises will like that. Anyone already worried about how much operational surface area one cloud provider can control probably won’t.

What this changes for builders

If you’re planning AI infrastructure for the next 12 to 24 months, assuming “Nvidia for training, x86 for everything else” is no longer the safe default. It may still be the right one. It’s no longer the only serious option.

A few questions now matter more:

• Can your training workloads fit within Neuron’s compiler and operator constraints?
• Do you need quick access to niche CUDA ecosystem features, or do you need lower training cost at scale?
• Is your application estate clean enough to move heavily onto arm64?
• How much vendor lock-in are you willing to accept for better economics and better capacity access?
• If you operate in remote, mobile, or telco-heavy environments, would integrated cloud-plus-satellite networking actually simplify anything important?

Amazon is betting that a lot of customers will answer those questions its way.

That doesn’t mean AWS replaces Nvidia, Intel, or Starlink across the board. It does mean Amazon is applying pressure to all three at once, and at enough scale that engineers should pay attention. When Trainium capacity is selling out before the next generation ships, and Graviton shows up across nearly every large EC2 customer, the market has already shifted.