Part 3 of 3 in a series reading Coatue’s October 2026 Public Markets Update. The other parts: Part 1: Measuring the Ages and Part 2: AI Beyond Software. All charts and figures discussed come from Coatue’s own deck, which you can read at coatue.com. Coatue labels many of its slides illustrative or forward-looking, and so do I.
The question underneath every AI bubble argument is simple. Is the revenue arriving fast enough to pay for the spending? Coatue’s October update tries to answer it with a comparison most commentary skips: what the last giant build-out looked like at the same stage.
The cloud benchmark
One slide measures incremental cloud revenue as a percentage of annual cloud capex, using the combined hyperscalers, from 2010 to 2020. In 2010 it was 83%. It fell to 33% in 2013 and 30% in 2014, recovered to the mid 60s by 2016 and 2018, and ended 2020 at 59%. Coatue puts the average at 60%.
Put differently, for each dollar the hyperscalers spent on cloud capacity in a year, they added roughly 60 cents of new annual revenue that same year. There was a scary stretch in 2013 and 2014 where it looked like a poor trade. Then it recovered and stayed near 60.
The AI version
The next slide does the same math for AI and calls it “Is AGI worth the wait?” The ratio is 9% in 2024, 11% in 2025 and 21% in 2026 on Coatue’s estimates. The bars for 2027 through 2030 are grey and labelled E: 23%, 24%, 26% and 30%.
Those later bars come with an important footnote. They are an illustrative scenario in which each year’s revenue equals the prior year’s revenue plus 30% of selected-company capex, anchored to an assumed 2026 revenue of about $0.277 trillion. Coatue says it is not a forecast.
Even on that scenario, the ratio stays well under the cloud average of 60%. The dotted arrow between 30% and 60% is labelled “AGI gap?” The deck’s own framing is that the missing revenue has to come from somewhere, and the something is AGI-level capability that opens markets today’s products cannot reach.
The 2013 and 2014 dip in cloud is the steelman. A ratio of 30% did not mean cloud was a bad business. It meant the capacity was ahead of the demand, and demand caught up. A skeptic should concede that the early years of a build-out can look like this.
How big the spending gets
The comparison matters more because of the size of the bill. Coatue estimates 2028 AI capex from hyperscalers, labs and neoclouds at $2.8 trillion. Set that against annual flows into other large US capital pools: about $2.1 trillion of single-family mortgage originations in 2026, $1.0 trillion for the national defense budget, and $0.8 trillion for all non-AI S&P 500 capex in 2025.
On those numbers AI capex would be roughly 1.5 times mortgage originations, 3 times the defense budget and 3.5 times everything the rest of the S&P 500 spends on capex. Coatue’s own line on the slide is that AI capex “will soon need its own Fannie and Freddie to standardize, finance and scale.”
The slide carries a caveat worth repeating: the comparisons mix periods and categories, and the AI figures are estimates.
The financing is changing shape
The next slide shows how the money is being raised. At the start, a hyperscaler pays capex for GPUs and a data center. Then it moved off the balance sheet: a neocloud owns the GPUs under a take-or-pay contract, financed 75% debt and 25% equity with a chipmaker guarantee, while a separate data center carries a lease guarantee.
Now the picture is layered. A data center joint venture is funded with 80% from a credit fund and 20% from hyperscaler equity plus debt, backed by a residual value guarantee from the hyperscaler. A separate GPU layer sits underneath with chipmaker guarantees, debt and equity, and a compute contract from the hyperscaler.
Coatue’s footnote says the structures are simplified illustrations based on public reporting, including Meta’s Hyperion joint venture and Nvidia’s capacity agreement with CoreWeave. Whatever the details, the point is clear. Guarantees and take-or-pay contracts mean that risk is being moved around the system, not removed from it. If revenue per dollar of capex stays low, the people who feel it are the ones holding the guarantees.
What the deck itself says about the risk
Coatue is not neutral here, and it does not pretend to be. A different slide lists the bear case: financing and credit concerns, open-weight models taking share, job displacement and security fears. The frontier lab P&L slide flags the same open-weight question with a red cross, the only red mark on the page.
The firm’s position is to value labs on earnings before training costs, and to treat training as a discretionary bet on AGI. That is coherent. It also means the whole case rests on training costs being optional, and optional is a hard thing for a lab to be when competitors keep spending.
What to watch
The deck gives you one number to track: incremental AI revenue divided by annual AI capex. It was 9% in 2024 and is an estimated 21% in 2026. The cloud analogy says a trough like this is survivable and the ratio can recover to about 60% once demand catches up. The deck’s own scenario says it might not get there on current products alone.
Watch which way it moves. If the ratio keeps climbing toward the cloud range, the spending starts to look like cloud in 2015. If it flattens around 25 to 30 cents while capex keeps rising past $2 trillion, the ones holding the guarantees are the ones who need to be right about AGI.
Source and attribution. This post is commentary on Coatue’s Public Markets Update (October 7, 2026), published by Coatue Management at www.coatue.com/blog/perspective/public-markets-update-2026-10-07. The figures, slide titles and scenarios discussed belong to Coatue and are quoted for analysis; they are Coatue’s estimates, not mine, and not investment advice. Read the original deck and its disclosures before relying on any number here.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
