Amazon's Durability | Stratechery by Ben Thompson

ELI5/TLDR

Amazon just announced it will sell its entire logistics stack — planes, trucks, last-mile — as a service to other companies, the way it once sold its servers as AWS. Ben Thompson predicted this exact move a decade ago. The pattern is always the same: Amazon builds infrastructure for itself, becomes its own best customer, then rents the leftovers to everyone else. The same playbook explains why AWS, once feared to be losing the AI race, is now well-positioned for inference. And why Amazon Leo (satellites) and drone delivery are next on the same conveyor belt.

The Full Story

The Amazon Tax, paid in full

The trigger is small but telling. Amazon Supply Chain Services bundles freight, trucking, and last-mile delivery into a single offering for outside businesses. P&G and 3M are already using it. FedEx and UPS shares dropped on the news. Thompson called this in 2015 in a piece titled “The Amazon Tax” — that logistics would follow AWS and the marketplace down the same well-worn path: build for yourself, scale to absurdity, then sell access.

The pattern, stated plainly: convert marginal costs into capital costs, then earn leverage on those capital costs by selling them to other businesses. It takes a decade to play out. Amazon is one of the few companies that actually thinks in decades.

How AWS almost got written off

Rewind to 2023. SemiAnalysis published a piece arguing AWS would lose the future of computing. The argument was sharp. Amazon insisted on its own networking (Nitro, EFA) instead of Nvidia’s. It insisted on its own chips (Graviton, Trainium) instead of buying Nvidia’s best. Nvidia, sitting on a monopoly, had no reason to ship its scarce GPUs to a customer trying to replace it. So Nvidia would starve AWS and feed the “me-too” clouds — Oracle, the neo-clouds — who played along.

This was correct at the time. Training large models needs thousands of GPUs networked tightly together to update weights in lockstep. AWS’s architecture wasn’t built for that. Microsoft and Oracle’s were.

Then the workload changed

Training is no longer the biggest AI market. Inference is. And inference is now layered: a chatbot answers, then a reasoning model thinks before answering, then an agent triggers reasoning models on a loop to finish a task. Thompson cites a framework: ChatGPT moment, o1 moment, Opus 4.5 moment. Each one squares the token demand.

Inference doesn’t need thousands of GPUs wired together. A model often fits on a single server. Memory matters more than horizontal networking. Agentic workloads are CPU-heavy and benefit from disaggregating CPU and GPU resources — which is exactly what Nitro has been doing for years. The thing that looked like a weakness in 2023 turned into a fit in 2026.

Jensen’s counter is “tokens per watt”: power is the bottleneck, so pay up for the best chips. Thompson’s reply: if you can afford to buy that many Nvidia chips, you can afford to buy more power. AWS has been investing in power upstream so it can spend less downstream. Electricity is more commoditizable than logic. And agentic workloads make perfect GPU utilization much harder, which weakens the case for the most expensive chip.

The compounding trick

Amazon bought Annapurna Labs in 2015. First AI chip in 2019, mediocre. Trainium 3 in 2026, decent and on a good trajectory. Seven years of patient iteration nobody noticed until it mattered.

There is a second, subtler advantage. Because AWS threw off so much cash, Amazon could write a multi-billion-dollar check to Anthropic. Nvidia couldn’t. Jensen admitted in an interview this was his mistake — he didn’t realize foundation labs needed their compute supplier to also be their investor. Amazon and Google did. Now Anthropic, the only frontier lab available on all major clouds, runs partly on Trainium. Customers using Bedrock are using Amazon silicon without ever choosing it. Graviton playbook, replayed.

Why Amazon can afford to be neutral on models

Microsoft has to own a frontier model — software is its core business and AI threatens it directly. Earlier this year Microsoft missed Azure growth targets because it diverted compute to its own AI workloads. Cloud demand is eternal; the AI threat is existential. Google is the same story. Both are aggregators competing for attention one click away from substitution.

Amazon’s core businesses are physical: shipping goods, building data centers. The AI threat there is far enough away that it can hand most of its chips to customers without panicking. Same for Apple. The further your business sits from the model layer, the less you need to win the model wars.

What’s already on the conveyor belt

Amazon Leo, the satellite constellation, looks duplicative of Starlink. But Andy Jassy described it on the last earnings call in language that should sound familiar by now — capital-intensive upfront, asset-leveraged over long periods, AWS-like free cash flow profile. The first best customer is Amazon itself.

Then look at drones. Amazon started talking about drone delivery 13 years ago. The endgame is delivery costs that are just depreciation on drone fleets. Those drones need reliable satellite coverage. If Amazon doesn’t want to depend on Jensen for chips, why would it want to depend on Elon for connectivity? Leo is the chip story applied to the sky.

Key Takeaways

The Amazon formula in one sentence: build infrastructure for yourself, become your own best customer, rent the rest. AWS, marketplace, logistics, satellites, drones — same pattern, different decade.
Marginal costs become capital costs become leverage. This is the move.
The 2023 case against AWS in AI was right about training, wrong about what would matter. Inference and agents reward Amazon’s disaggregated, custom-silicon architecture.
Owning a chip program is what let Amazon fund Anthropic, which is what gave AWS frontier-model parity. The flywheel has more spokes than it looks.
The further your core business is from the model layer, the more comfortable you can be not owning a frontier model. Amazon and Apple sleep fine. Google and Meta cannot.
Long-term investments are a strategy in themselves: by the time threats arrive, you’ve already built the response.

Claude’s Take

Thompson is doing his best work here — the kind where a 2015 prediction becomes a 2026 victory lap without smugness, and the framework holds up because it was never about Amazon specifically, it was about how capital-intensive infrastructure compounds.

The one piece worth pushing on: “AWS is fine now” is a much weaker claim than “AWS will win.” The argument that inference suits AWS’s architecture is real, but Nvidia is still the scarce resource, and Oracle, CoreWeave, and the neo-clouds still get preferential allocation. AWS is back from “losing the future” to “credible second” — which is enough to keep the cash machine running, but not the same as winning AI.

The deeper point — that companies rooted in the physical world have natural patience while aggregators have to sprint — feels right and is the kind of frame you can carry into other arguments. Worth keeping.

8/10. A clean essay on a real pattern from someone who has earned the right to say “I told you so” and mostly chose not to.