Email me

Economics · · 10 min read · Matt Busel

By , who built the apps, crates and game he writes about · every number links to its source and every dataset is downloadable (methods) · corrections · featured in Kirino, gamebiz, This Week in Rust, New Mobile Life, By-Press

Filed under Opinion AI industry Economics

AWS is a landlord: 39% margins and a $496B lease on your stack

AWS is a landlord: 39% margins and a $496B lease on your stack
On this page · 11 sections
  1. Key numbers
  2. The cloud earns like a landlord, not a utility
  3. All three landlords are raising the take
  4. The lease book just doubled, and the leases got longer
  5. Follow the money: the rent buys more buildings
  6. Same GPU, different landlord: the 3.5x premium
  7. The exit fee is the lock on the door
  8. The steelman, and why it fails for most teams
  9. The twist: the landlord's best tenants can least afford the rent
  10. What this means for you
  11. Sources

Short answer: The cloud is not a utility, it is a landlord. AWS kept 39.4 cents of operating profit from every dollar of revenue in Q2 2026, Microsoft's Intelligent Cloud kept 41.3% for its fiscal year, and Amazon's book of signed long-term cloud contracts doubled in six months to $496 billion. Those are rent numbers, and the builders paying them are the ones who can least afford it.

TL;DR: AWS was 18% of Amazon's 2025 sales and 57% of its operating profit, at a 35.4% margin, while Amazon's own North America retail business earned 6.9%. The same Nvidia H100 rents for $3.14 an hour at Massed Compute and $11.06 at Google Cloud and Azure, a 3.5x gap for identical silicon, per our GPU price tracker on Oct 25 2026. 37signals walked out, cut its cloud bill from $3.2 million a year to $1.3 million, and paid off about $700,000 of servers inside a year: the rent was always optional for steady workloads, and the people who notice early keep the difference.

Key numbers

WhatNumberSource, checked Oct 25 2026
AWS operating margin, Q2 202639.4% ($16.6B on $42.2B)Amazon 10-Q, Q2 2026
AWS share of Amazon operating income, 202557.0% ($45.6B of $80.0B)Amazon 10-K, FY2025
Amazon North America retail margin, 20256.9%Amazon 10-K, FY2025
Microsoft Intelligent Cloud margin, FY to June 202641.3% ($57.0B on $137.8B)Microsoft 10-K, FY2026
Google Cloud margin, Q2 2026 vs Q2 202535.6% vs 20.7%Alphabet 10-Q, Q2 2026
Amazon long-term contract backlog (mostly AWS)$80.4B (Dec 2021) to $496B (Jun 2026)Amazon 10-K and 10-Q filings
Average remaining term of those contracts4.1 years (Dec 2025) to 6.4 years (Jun 2026)Amazon 10-Q, Q2 2026
H100 SXM on-demand, cheapest vs Google Cloud$3.14 vs $11.06 per GPU-hourTensorust GPU price tracker
AWS internet egress, first 10 TB$0.09 per GB ($90 per TB)AWS Price List API
37signals cloud bill after leaving$3.2M a year to $1.3MDavid Heinemeier Hansson, Oct 17 2024
Animated bars: AWS operating margin 28.4% in 2018, 29.8% in 2020, 28.5% in 2022, 27.1% in 2023, 37.0% in 2024, 35.4% in 2025 and 39.4% in Q2 2026
AWS operating margin by year, from Amazon 10-K and 10-Q segment tables, pulled Oct 25 2026

The cloud earns like a landlord, not a utility

The industry sold the cloud as electricity: a utility, a commodity, a meter on the wall. Utilities earn thin, regulated margins. The cloud does not.

Amazon's own segment table is the cleanest proof, because it puts a landlord and a shopkeeper side by side in one filing. In 2025, Amazon's North America store sold $426.3 billion of goods and kept $29.6 billion, a 6.9% margin. AWS sold $128.7 billion of compute and storage and kept $45.6 billion, a 35.4% margin.

So AWS was 18% of Amazon's sales and 57% of its operating profit. The famous retailer is, by the numbers, a property company with a shop on the ground floor.

And the margin is climbing, not compressing the way a commodity's should. AWS ran between 26% and 30% from 2018 to 2023, jumped to 37.0% in 2024, and hit 39.4% in Q2 2026. Commodities get cheaper as they scale. Rent goes up when the tenants cannot leave.

Bar chart of operating margins: Microsoft Intelligent Cloud FY2026 41.3%, AWS Q2 2026 39.4%, Google Cloud Q2 2026 35.6%, AWS FY2025 35.4%, Google Cloud FY2025 23.7%, Amazon North America retail FY2025 6.9%
Cloud segments earn five to six times the margin of Amazon retail. SEC filings, pulled Oct 25 2026

All three landlords are raising the take

This is not an Amazon quirk. Microsoft's Intelligent Cloud segment, which holds Azure, earned $57.0 billion on $137.8 billion of revenue in the fiscal year to June 2026: 41.3%.

Google Cloud is the tell. It earned 5.2% in 2023, 14.1% in 2024 and 23.7% in 2025, and in Q2 2026 it kept $8.8 billion of $24.8 billion, a 35.6% margin, up from 20.7% a year earlier. The "underdog" cloud did not use its scale to undercut the others. It used it to catch up to their margin.

Add the most recent full years together and the three biggest clouds booked about $116 billion of operating profit (AWS 2025, Microsoft Intelligent Cloud fiscal 2026, Google Cloud 2025; our sum across their different fiscal calendars). That is the yearly rent roll on the internet's servers.

The lease book just doubled, and the leases got longer

Every landlord's real asset is the signed lease. Amazon discloses its version in the revenue note of every 10-Q: "commitments in customer contracts for future services," primarily AWS, not yet recognized as revenue.

That number was $80.4 billion at the end of 2021. It reached $244 billion at the end of 2025. Then it went to $364 billion in March 2026 and $496 billion in June 2026. It doubled in six months.

The length matters even more than the size. The weighted average remaining life of those contracts sat between 3.5 and 4.1 years for four straight years. By June 2026 it was 6.4 years. Customers are not just renting more. They are signing longer leases.

Line chart: Amazon long-term contract commitments not yet recognized rose from $80.4B in Dec 2021 to $244B in Dec 2025, $364B in Mar 2026 and $496B in Jun 2026
Amazon contracted backlog, mostly AWS, from 10-K and 10-Q revenue notes, pulled Oct 25 2026

Amazon's filing does not name who signed. Our reading: the AI build-out turned the cloud from month-to-month rentals into multi-year leases, and the tenants locking in for six-plus years are the AI companies that believe they cannot get compute any other way. That belief is worth half a trillion dollars of booked rent to Amazon alone.

Follow the money: the rent buys more buildings

Here is the loop nobody draws. The rent funds the next round of buildings, which locks in the next round of tenants.

Amazon, Microsoft and Alphabet together spent $30.1 billion on property and equipment in Q1 2022. In Q2 2026 they spent $134.9 billion in a single quarter: $54.2 billion at Amazon, $44.9 billion at Alphabet and $35.8 billion at Microsoft, per their XBRL filings.

Line chart of quarterly capital spending: Amazon $15.0B, Alphabet $9.8B, Microsoft $5.3B in Q1 2022, rising to Amazon $54.2B, Alphabet $44.9B, Microsoft $35.8B in Q2 2026
Hyperscaler property and equipment spending by quarter, SEC EDGAR XBRL, pulled Oct 25 2026

Who wins: the three landlords, whose capex turns into an owned asset with a rent stream attached, and Nvidia, which sells them the GPUs. Who loses: every startup and mid-size company paying list price, because it is their margin that funds the landlord's next data center.

The "nobody ever got fired for buying AWS" consensus is the control play. It turns a pricing decision into a career-safety decision, and career safety never asks what the rent is.

Same GPU, different landlord: the 3.5x premium

If the cloud were a commodity, the same chip would cost about the same everywhere. Our GPU price tracker pulls on-demand prices from 17+ providers every day. On Oct 25 2026, the cheapest on-demand H100 SXM 80GB listing per provider ran from $3.14 an hour at Massed Compute and $3.20 at Hyperstack, through $3.99 at Lambda, to $6.88 at AWS and $11.06 at Google Cloud and Microsoft Azure.

Same silicon. A 3.5x spread between the cheapest listing and the two biggest brand names.

Bar chart of H100 SXM on-demand price per GPU-hour: Massed Compute $3.14, Hyperstack $3.20, Lambda $3.99, CoreWeave $6.16, AWS $6.88, Oracle $10.00, Azure $11.06, Google Cloud $11.06
Same H100, up to 3.5x the price. Tensorust GPU price tracker, Oct 25 2026

Run one GPU around the clock for a year and the difference stops being abstract. At Google's $11.06 that is about $96,900. At Massed Compute's $3.14 it is about $27,500. The roughly $69,000 gap per GPU per year is the brand premium, and for a team running eight GPUs it is more than half a million dollars a year (our arithmetic on the tracker's prices).

The neoclouds are not charities. They are proof that the hyperscaler price is a choice, not a cost floor. More on that market in our pieces on Nvidia backstopping the neoclouds and how Big Tech depreciates its GPUs over six years.

The exit fee is the lock on the door

Landlords make it cheap to move in and expensive to move out. In the cloud, that is egress: the fee to send your own data out to the internet.

AWS's own price list, published Sep 16 2026, charges $0.09 per GB for the first 10 TB a month from us-east-1, or $90 per terabyte, after a 100 GB monthly free tier. Azure's East US list price is $0.087 per GB on Microsoft's network. Cloudflare's R2 storage charges $0 for egress, because it can.

Bar chart of list price to send 1 TB to the internet: AWS us-east-1 $90, Azure Microsoft network $87, Azure internet routing $80, Cloudflare R2 $0
Egress list prices from the AWS and Azure price APIs and Cloudflare R2 pricing page, Oct 25 2026

Cloudflare's founders did the math in 2021: over the previous decade, wholesale bandwidth prices fell 93%, while AWS's egress fees fell only 25%. The cost of the pipe collapsed. The toll on the pipe barely moved. Google went the other way: its pricing page still notes that starting Feb 1 2024, it raised prices on some Premium Tier internet egress.

When did the landlords blink? On Mar 5 2024, AWS announced it would waive egress charges for customers moving out, and its own post says the waiver "follows the direction set by the European Data Act." Read that twice. The exit fee came down for leavers when a regulator pointed at it, not when competition did, and it still applies to everyone who stays.

The steelman, and why it fails for most teams

The best case for renting is real. You can spin up 100 machines in minutes. You skip buying hardware, staffing a data center and guessing demand a year out. For a company that might not exist in 24 months, or whose traffic spikes 50x on launch day, renting is the right call.

David Heinemeier Hansson, who led 37signals out of AWS, says exactly that: the cloud fits companies early enough that the spend is immaterial, or with wildly unpredictable loads. Then he names the trap: "Just be careful that you don't look at those lavish cloud credits as a gift! It's a hook."

The numbers show what happens once a workload is steady. 37signals' cloud budget was $3.2 million in 2022, "incredibly optimized, with long service commitments." After moving to its own Dell servers, it got the bill down to $1.3 million in 2024, all of it S3 storage it was locked into until summer 2025. The roughly $700,000 of hardware paid for itself during 2023, and the ops team stayed the same size. He projects more than $10 million saved over five years.

Stat card: 37signals cut its yearly cloud bill from $3.2M in 2022 to $1.3M in 2024 after buying about $700K of servers
37signals cloud exit, from David Heinemeier Hansson posts of Jun 2023 and Oct 2024

And "the cloud is more reliable" is the weakest line in the brochure. On Oct 19 and 20 2025, AWS's us-east-1 region had a disruption that ran from 11:48 PM to 2:20 PM PDT, about 14.5 hours, triggered by what AWS's own post-event summary calls "a latent race condition" in DynamoDB's automated DNS management. When every tenant is in the same building, one faulty wire takes out the whole block. We wrote about that kind of shared-dependency failure in the quiet cascade.

The twist: the landlord's best tenants can least afford the rent

Here is the part that should make every founder angry. The customers with the least margin to spare are the ones the hyperscalers court hardest: startups handed free credits on day one, built on proprietary managed services, and then billed at list price right as they start to grow.

A big enterprise negotiates. A public company has a procurement team. A three-person AI startup gets $11 GPU-hours, $90-a-terabyte egress and a 6-year lease if it wants guaranteed capacity. The rent comes out of the exact runway that was supposed to buy product-market fit.

Our bet, stated plainly: within three years, the default for any steady workload, AI inference included, is owned hardware or a neocloud, and the hyperscalers become what they should have been all along, the place you burst to on a busy day. Open-weight models make it real: a model you can run on your own GPU is a model nobody can raise the rent on. Nobody has proven the timeline yet. The margins say the pressure is already there.

What this means for you

Do the math DHH told everyone to do. Pull your last three cloud invoices and split them into three lines: steady compute, spiky compute and egress. Price the steady line at a neocloud or on owned hardware, using live numbers from our GPU price tracker. Avoid proprietary managed services for anything you would want to move, and treat free credits as a loan against your future margin.

If your bill is AI-heavy, price the model side too: API tokens versus running an open-weight model on a rented or owned GPU. Our $7 LLM API Cost Calculator turns your real request volume into a per-request, monthly and yearly cost across 26 models, so you can see the rent before you sign the lease. If it does not give you a clear monthly cost, we keep fixing it at no extra cost until it does what it promises.

More from our opinion desk at /news/opinion/, and more money stories at /articles/topic/economics/. Not financial advice: this is our reading of public filings and prices, not a recommendation to buy or sell any security.

Sources

First published here on tensorust.com on Oct 25, 2026. Related: Opinion, All opinion pieces, GPU price tracker, GPU depreciation.

Up next

More on OpinionIs college worth it? Grads 5.7% jobless vs 4.1%. The degree is a tollRecent grads: 5.7% jobless vs 4.1% for all workers, 42% in jobs needing no degree, tuition +208% since 2000. The degree is a toll, and AI can check skill.Read it next →

Cite this

Free to cite and reuse under CC BY 4.0 with a link back. How it was collected: methods. Spotted an error? Corrections.

Get the next one by email.

What my AI agents tried this week, with the real numbers: what worked, what flopped, what made money. One email when there is something worth reading.

No spam, never sold. Every email has a one-click unsubscribe.

What are people saying about this? Read and post on The Floor →

More: the author, related reading, share this, a free game

Try it yourself · Following AI?

Check what the models really cost

71 API prices, each copied from the provider page with the date, plus every change.

In The Manual: on this topic

Explainers answer one question, short answer first. Guides walk you through it step by step.

Open The Manual →

Read next

All articles →

1M+ X impressions in 4 days, run by AI agents · 8,000+ crates.io downloads across our 70 Rust crates, all time · 529 GitHub stars · 8 iPhone apps live on the App Store · O'BLOCK out now, free on Steam

Built by Matthew Busel · code on github.com/Mattbusel · numbers as of Oct 25, 2026, click any to check

Popular in the shop

All 124 tools →

Browse: Bundles and kitsHome, family and schoolMoney and businessAI toolsSports and fitnessMaking and hobbiesFood and drinkIndie dev and launchTrading and quant

Free: Landing page teardown iPhone app cost calculator LLM API cost calculator Paid online: $49 Launch Check App builds Articles FAQ