Part 2 of 3 in a series reading Coatue’s October 2026 Public Markets Update. The other parts: Part 1: Measuring the Ages and Part 3: Does the Money Show Up. 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.
For thirty years, “software” meant something specific: write it once, copy it for free, and charge for every seat. Gross margins are high and the marginal cost of the next customer is close to zero.
Coatue’s October update argues that AI does not work that way. Their slide on the token puts three physical inputs on the left (land with grid, water and permits; 24/7 electricity measured in gigawatts; chips, memory and networking) and a factory in the middle. Out the other side comes “intelligence,” a list of products that includes answers, code, research, forecasts, agents, designs, analysis, contracts and support.
The slide that follows says it plainly: intelligence will be measured in gigawatts. That is a sentence about power plants, not about apps.
A frontier lab, in one table
The most useful slide for understanding the shift is a deliberately simple profit and loss statement for a frontier lab. Coatue labels it illustrative and says the figures are a hypothetical $100 revenue framework, not any company’s results.
Revenue is $100 per gigawatt. Inference costs take $20, leaving an 80% gross margin that looks like classic software. Operating expenses take another $20, so adjusted EBITDA before training is $60, a 60% margin. Then training costs take $60 and EBITDA is zero. Capex takes a further $50 per gigawatt, and free cash flow is negative $50.
Look at where the money goes. A software company’s biggest cost is people. Here the biggest costs are training and capex, which are electricity and silicon. Coatue’s argument is that you should value the business on the line before training, because training spend is “easy to turn on and off” and is elevated in the chase for AGI. Their steady-state view is that earnings power “could be very high.”
That is a legitimate way to value the business. It is also a bet that training costs are discretionary and not a permanent cost of staying competitive, and the slide itself puts a question mark next to the one risk that could break it: the role of open-weight models.
Who buys the intelligence: agents
If the product is intelligence per gigawatt, somebody has to buy it. Coatue says the buyers will be agents.
One slide sets today against the future. Right now your phone is a small computer and your apps run on it. Coatue pictures the phone as a remote control with a big screen, while agents do the actual work on your computer and in the cloud. The drawing shows a person talking to a phone, and one agent passing work to five others.
Then the deck turns the picture into infrastructure math. It assumes 8 billion AI agents, 64 agents per server CPU and an average price of $10,000. That works out to about 125 million server CPUs and roughly $1.25 trillion over a five-year life, or around $250 billion a year. On Coatue’s estimate, server CPU revenue climbs from $40 billion in 2026 to $300 billion in 2031. The slide is really about Intel, and Coatue says plainly that the inputs are assumptions.
The consumer version
The same logic is applied to commerce on a separate slide about Meta. Start with $21 trillion of 2031 global retail sales outside China. Assume 40% moves online, which is $8 trillion. Assume 30% of that is agentic, which is $2.5 trillion. Assume the agent takes 10%, and you get $250 billion of agentic commerce revenue, which the slide notes is similar to total 2026E revenue.
The shift Coatue is describing is from monetizing attention to monetizing intent. An ad shows you something. An agent buys it for you and keeps a cut of the transaction.
Every number in that chain is an assumption, and the slide says so. A change of a few points at any step moves the answer by tens of billions. It is a framework for thinking about the size of the prize, not a forecast.
Why “beyond software” is more than a slogan
Take the slides together and the claim holds together as a chain. The input is physical (power, land, chips). The product is priced by throughput, so revenue scales with gigawatts. The demand side is agents running continuously, not people opening an app. And the financial profile looks less like a software license and more like an industrial plant: huge up-front capex, high fixed costs relative to running costs once it is built, and a cost of capital that matters a lot.
That last point is where the argument has teeth for a skeptical reader. If AI is a utility-like business, then returns depend on utilization, financing costs and how long the assets stay useful, not on how clever the interface is. Those are the questions you would ask about a power company.
The counterargument deserves a hearing. Software margins came from near-zero marginal cost, and a lab with 80% gross margins still looks like software on the first lines of the table. If training costs fall, or if open-weight models cap what labs can charge per token, the utility framing could turn out to be a transitional phase and not a permanent business model. Coatue flags exactly that risk.
What the deck does not do is prove its assumptions. It gives you a better set of questions than “is AI the next software?”: how many gigawatts, at what revenue per gigawatt, financed how, and consumed by whom.
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
