Amazon’s Durability (Stratechery Article 5-5-2026)

Author: Ben Thompson

1. Key Themes

Theme 1: The "Primitives" Playbook — Amazon Repeatedly Converts Internal Infrastructure Into External Revenue

Amazon's core strategic pattern is to build infrastructure that serves itself first, then monetize it externally. This has worked with AWS, logistics, and is now playing out in AI and satellites.

"In both cases Amazon built 'primitives' that had Amazon itself as their first, best customer, justifying and driving initial development, but in both cases the ultimate play was to sell those primitives to other companies."

The logistics announcement (Amazon Supply Chain Services) is the decade-delayed fulfillment of a 2016 prediction — air/ocean freight, trucking, and last-mile delivery now sold as a unified suite to companies like P&G and 3M. FedEx and UPS shares dropped on the news.

"Amazon, more than any other company, actually operates with decade-long timeframes, consistently making real-world investments at massive scale that (1) convert their marginal costs into capital costs and (2) gain leverage on those capital costs by selling them to other businesses."

Theme 2: The AI Compute Market Has Shifted From Training to Inference — and That Shift Favors AWS

The narrative that AWS was "behind" in AI was predicated on training workloads dominating. That is no longer the case. Thompson identifies three AI inflection points that have changed the compute landscape:

"The first inflection point was the emergence of LLMs — call this the ChatGPT moment… The second inflection point was the emergence of reasoning models — call this the o1 moment… The third inflection point was the emergence of functional agents — call this the Opus 4.5 moment… This isn't just an exponential increase in the addressable market for tokens, it's two exponential increases squared."

Inference workloads — especially agentic ones — don't require networking thousands of chips together (which was AWS's weakness for training). Single-server model storage, disaggregated CPU/GPU routing, and Amazon's Nitro architecture are now structural advantages:

"Achieving maximum utilization of heterogeneous compute means unbundling CPUs and GPUs and routing workloads between resources, which is exactly the sort of disaggregated-resource abstraction that Amazon has been building with Nitro."

Theme 3: Custom Silicon Beats Nvidia Margins in the Long Run — If You're Patient Enough

Amazon's 2015 acquisition of Annapurna Labs and decade of chip investment is now paying dividends. The bet was never about beating Nvidia on day one — it was about achieving structural cost advantages over time through vertical integration, mirroring the Graviton ARM chip playbook.

"Amazon bought Annapurna Labs, which makes their chips, in 2015, and launched their first AI-focused chip in 2019. No, it wasn't very good, but critically, that was seven years ago: now Trainium 3 is decent and the trajectory is even better. AWS is positioned to have a sustainable cost advantage for inference going forward."

The mechanism for embedding custom chips is through abstraction layers — customers using Bedrock use Trainium without knowing it:

"Trainium chips help undergird Bedrock, its AI platform, which is to say that users are using Trainium chips even if they didn't explicitly choose to do so."

Theme 4: Physical-World Anchoring Creates Durable AI Investment Incentives

Thompson introduces a framework for assessing which companies have the strongest long-term incentives to invest in AI infrastructure: the more rooted in the physical world, the less existentially threatened by AI disruption, and therefore the more capacity to invest rationally rather than defensively.

"I increasingly suspect that long-term vulnerability to AI — or, to put it more positively, long-term incentives to invest in AI — are very strongly correlated with the degree to which a company interacts with the physical world, and secondarily, the degree to which companies feel secure in their control of distribution."

Microsoft's Azure growth miss is the cautionary tale — a software-first company diverted compute from cloud customers to protect its own AI products:

"The company missed Azure growth projections because they devoted more compute to their internal workloads. It was an understandable decision: cloud demand is eternal, but the risk from AI for existing software businesses is existential."

2. Contrarian Perspectives

Contrarian 1: Jensen Huang's "Tokens-Per-Watt" Argument for Nvidia Has Real Holes

The consensus view is that Nvidia's GPU dominance is cemented by power constraints — if you can't add watts, you need the most efficient chips possible. Thompson pushes back with three counter-arguments:

"First, if you have the money to buy that many Nvidia chips, you also have the money to spend on getting more power — which is exactly what AWS has been focused on… Second, in the long term, electricity is more of a commodity than logic is… Third, the nature of inference workloads — particularly agentic ones — is such that perfect accelerator utilization is going to be a much harder problem to solve than when it comes to training."

This implies that at scale, the power constraint argument actually favors companies who can buy their way out of it, and that custom silicon ROI improves as utilization challenges grow.

Contrarian 2: AWS Was Never Actually "Behind" in AI — The Market Simply Hadn't Arrived at the Workload It Was Built For

The prevailing 2023 narrative (captured by SemiAnalysis) was that AWS would "lose the future of computing." Thompson argues this analysis was correct but time-bound — it described a training-dominated world that has since been superseded:

"These concerns were well-founded in the 2023 time-period when that Article was written: that was a time when AI, thanks to ChatGPT, had hit the mainstream, but the largest share of compute still went to training… Both the shift to inference and the shift in the nature of inference have been positives for AWS' approach."

The implication: apparent competitive weakness in a rapidly shifting technology market may simply mean a company is optimized for the next phase rather than the current one.

Contrarian 3: Amazon's Strategic Neutrality Is a Competitive Advantage Over Microsoft's Cloud

Conventional wisdom treats Microsoft Azure + OpenAI as the dominant enterprise AI cloud stack. Thompson argues Amazon is structurally better positioned as a neutral provider precisely because it lacks an existential AI threat to its core business:

"You can also make the case that Amazon is the best choice for frontier model access in a world of limited compute: Microsoft's core business is software, which is to say that the company faces massive pressure to invest in their own AI capabilities, even at the cost of de-prioritizing cloud customers."

Enterprise buyers who need reliable compute allocation have reason to prefer a provider whose business incentives are aligned with customer access over self-preservation.

3. Companies Identified

4. People Identified

5. Operating Insights

Insight 1: Build Internal Infrastructure to Justify Investment, Then Externalize It

Amazon's durable advantage comes from a specific capital allocation discipline: build large-scale infrastructure where you are the anchor customer (ensuring economic viability), then open it to third parties to generate leverage on fixed costs. This converts capital expenditure into a compounding moat.

"Amazon itself would be this logistics network's first-and-best customer, just as was the case with AWS. This justifies the massive expenditure necessary to build out a logistics network that competes with UPS, FedEx, et al… I expect the company to offer its logistics network to third parties, which will increase the returns to scale, and, by extension, deepen Amazon's eventual moat."

Tactical application: When building expensive internal tools or infrastructure, design them from day one to be externally sellable. The internal use case funds development; the external market funds profitability.

Insight 2: Abstract Your Cost Advantages Away From Customers — They Don't Need to Know

Amazon embeds its cheapest silicon (Graviton, Trainium) into managed services (RDS, Bedrock) so customers consume the cost advantage without needing to opt in or even know it exists. This is a deliberate strategy for monetizing infrastructure investment without customer friction.

"PaaS lets Amazon double-dip in terms of profitability: first, AWS could sell PaaS products at a higher margin than IaaS products, and second, the company could leverage its own cheaper silicon to serve those products, reducing their costs."

Tactical application: When you have a cost or technology advantage, the highest-leverage way to monetize it is often through an abstracted service layer — not by forcing customers to choose your underlying technology explicitly.

Insight 3: Patience on Chip/Technology Investment Has Asymmetric Returns

Amazon's Trainium chips were "not very good" at launch in 2019. Seven years later, Trainium 3 is competitive. The implication is that companies who can tolerate long payback periods on silicon investment gain structural cost advantages that are very difficult to replicate quickly.

"Amazon bought Annapurna Labs, which makes their chips, in 2015, and launched their first AI-focused chip in 2019. No, it wasn't very good, but critically, that was seven years ago."

6. Overlooked Insights

Overlooked Insight 1: Amazon Leo + Drone Delivery Is a Multi-Decade Convergence Bet — Not a Me-Too Starlink Play

The article briefly frames Amazon Leo (its satellite internet constellation) as a potential infrastructure layer for autonomous drone delivery, not primarily a consumer broadband product. This reframes its entire competitive logic and timeline. The satellite service provides the communications backbone for a drone fleet — eliminating marginal delivery labor costs entirely and replacing them with depreciation on drone assets.

"It's increasingly plausible to imagine a future where delivery costs are a matter of depreciation on drone assets, and what would such a future require? How about reliable widespread satellite coverage for communicating with and guiding those drones?"

If this convergence plays out, Leo is not competing with Starlink for broadband subscribers — it is an internal logistics infrastructure investment following the exact AWS/ASCS playbook, with third-party revenue as a secondary benefit.

Overlooked Insight 2: Nvidia's GPU Allocation Was Deliberately Tilted Away From Amazon During the Training Era

This is stated plainly but easy to miss: Nvidia actively de-prioritized Amazon for H100 shipments because Amazon was simultaneously building competing chips and not adopting Nvidia's networking stack. This was a deliberate strategic choice by Nvidia to favor more dependent cloud customers — and it may have inadvertently accelerated Amazon's custom silicon roadmap by forcing self-reliance.

"Why would Nvidia prioritize Amazon for these GPUs, when they know Amazon will move to their in-house chips as quickly as they can… Nvidia prioritizes the me-too clouds. Amazon does get meaningful volume, but nowhere close to where demand is."

The implication for investors: companies that threaten Nvidia's stack by building alternatives may face supply constraints during critical windows — but those that survive emerge with more durable cost structures.