Patel sees $700 billion at two AI labs by 2027 while Levie gives enterprises a decade

Bain says the AI industry needs $6 trillion of revenue a year by 2031 to pay for its data centers. On a podcast published October 9, Dylan Patel of SemiAnalysis called that number low and said OpenAI and Anthropic alone could book $600 to $700 billion by December 2027. The same week Aaron Levie, who sells software to the banks and drugmakers that would have to write those checks, told two audiences that enterprise adoption is a job of ten to twenty years. Both men are bullish. Their calendars do not match, and the buildout is financed on Patel's.

Patel thinks the bears are undercounting the spending

Host Alex Kantrowitz read Patel the Bain figure and waited for a defense. Bain's Global Technology Report says today's AI products top out at $1.2 to $1.8 trillion a year by 2031, which leaves about $4.2 trillion to come from markets that do not exist yet. It puts 2026 capital spending at five hyperscalers near $780 billion. Patel's first move was to say the spending is larger than that. Count the suppliers building fabs, memory lines and power gear, he said, and US capex next year is on the order of $2 trillion, or 5% to 6% of GDP. Then he said revenue clears the bar anyway.

But you know, I'm of the belief that we can get to like 600, 700 billion across OpenAI and Anthropic alone by December of next year.

Dylan Patel, October 9, 2026, source

His payback rule is simple. A trillion dollars of chips depreciated over six years needs about $250 billion of revenue a year. Apply that to his own $2 trillion and a single year of spending has to earn roughly $500 billion a year for six years. That last step is my arithmetic, and his $2 trillion includes supplier capex that Bain's hyperscaler count leaves out. One caution on the forecast itself. In his next sentence Patel said the two labs could reach $1.5 trillion "maybe even by end of 2027", which is the same date as the $700 billion. I am using the lower figure and assuming the higher one belongs to 2028.

The Anthropic number runs ahead of anything reported

Patel's case leans on one company. He said Anthropic grew tenfold this year, from under $10 billion of revenue to more than $100 billion, and that it now earns more than it spends on training and inference compute combined. The last public figure I could find is lower. In August, Bloomberg's sources put Anthropic's annualized run rate above $65 billion at the end of July, up from about $9 billion at the end of 2025. A run rate is one month multiplied by twelve. Anthropic may well have passed $100 billion since July at that slope, but the S-1 is still confidential and nobody outside can check.

And Anthropic on their inference is doing 75% gross margins, a little bit higher even, you know, especially on the frontier models.

Dylan Patel, October 9, 2026, source

Asked where customers find $700 billion, Patel pointed at research budgets, cost of goods and payroll. He expects labor's share of the economy to keep falling while wages hold flat in dollars because everything else grows. He was candid about how much of that is hope.

And so, you know, I would hope, and for society, that that's my belief, you know, and maybe it's like an irrational belief, but I believe that we will just grow the economy so fast that labor in terms of dollars will not shrink.

Dylan Patel, October 9, 2026, source

Levie says the buyers are not ready

Levie spends his weeks in the meetings where that payroll money would be reallocated. At an Axios event on October 10 he described coding agents in "vertical takeoff" and almost everything else stuck behind chores. Companies have not sorted out which data an agent may read. They have not decided where a person signs off. Their compliance teams do not know who is liable when an agent approves a loan.

But we are still truly in, you know, the moments before the first inning of this rollout and it's because there's just a tremendous amount of change management that has to happen in the enterprise.

Aaron Levie, October 10, 2026, source

He does not think this is bad news. Slow diffusion, he told Axios, gives the economy time to work out jobs and safety, and it leaves room for companies like his that connect models to old workflows.

Ten years on one stage, twenty on the other

Two days earlier, at AGI House during SF Tech Week, Levie had been blunter about the timeline. He compared agents to cloud computing, which every startup in the Valley adopted within a year or two of Amazon launching S3.

But it still took two decades for the rest of the economy to even get to the point where they even agreed that the cloud was the future.

Aaron Levie, October 8, 2026, source

Later in the same session he said re-engineering sales, recruiting and research around agents is "a decade of diffusion". Most companies, he said, have handed staff a chatbot that they use the way they used Google, and that saves at best 30% of the time on one task. He also said he can predict a customer's progress only by guessing what its executives read on X.

Box's own numbers fit the slow clock. In August Box reported quarterly revenue of $321.1 million, up 9%, and guided to about $1.29 billion for the fiscal year, up 10%. Box is the applied layer Levie keeps describing. If the layer that carries models into banks grows 10% in the year coding went vertical, I read that as evidence for his timeline over Patel's.

Nobody selling tokens knows the margin yet

A founder at AGI House asked Levie about the labs subsidizing consumer plans. He said he had seen reports of plans carrying as much as $10,000 of subsidized compute, paid for in the end by the endowments that back venture funds. The labs, he said, are hoping efficiency gains shrink that bill while capital is still cheap.

Like the big, I think the big open question for anybody at the lab side is like, are we at, is the terminal business model of AI a 10% margin business? Is it an 80% margin business?

Aaron Levie, October 8, 2026, source

Patel's 75% and Levie's question can both be true, since one describes API inference and the other describes flat-rate subscriptions. But Patel also said the labs have almost no customer acquisition cost. A subsidized plan is a customer acquisition cost. It just shows up on the compute line.

Meta is aiming at the people who pay nothing

Alexandr Wang, Meta's chief AI officer, gave David Senra the consumer version of the bet in an interview recorded October 5 and published October 11. Meta launched its Muse agent on September 8 with a free tier and plans at $20 and $100 a month. Wang said the rest of the industry crowded into coding agents after Anthropic pulled ahead, and Meta chose not to follow.

Meta has distribution to three and a half billion people using the apps every day. And the natural area where Meta was going to be successful is consumer AI.

Alexandr Wang, October 5, 2026, source

Pressed for a user count, he would only say "We have said millions." That is a real launch, and it is still a cost before it is revenue. Meta can carry it out of advertising cash flow, which is why Patel waved off talk of Meta's solvency. It does nothing for the $700 billion.

Musk paid $60 billion partly for a sales team

Elon Musk made Levie's point from the other side in an interview at Giga Texas posted October 10. Asked why SpaceX bought Cursor, he listed coding data first and then something less glamorous.

So basically for success of AI, you need data, talented people, and compute, and you also need a good enterprise sales team. And Cursor also had a good enterprise sales team with relationships with thousands of companies.

Elon Musk, October 10, 2026, source

SpaceX agreed in June to pay $60 billion in stock for a company with about $2.6 billion of annualized business revenue. Levie told the AGI House crowd there is no end of jobs right now for account executives who understand the technology. When the man building rockets and chip fabs says the fourth ingredient is salespeople, I believe the bottleneck has moved to the customer.

Can AI revenue arrive faster than enterprises adopt it?

For a year or two, yes. Coding and customer support have already gone vertical, those buyers need little change management, and Patel's 2027 figure mostly requires them to keep compounding. Bain's 2031 figure is different. It needs law, banking, accounting and drug research to pay at the same intensity, and those sit on Levie's timeline. The first hard evidence will be Anthropic's public S-1. If most of its revenue still comes from software developers, the labs have sold to the easy customers and the next trillion depends on the slow ones.