Three years ago, one Sequoia partner did some quick math on a napkin and started a debate that Silicon Valley has been trying to answer ever since. The number was big then. It is much bigger now. And a growing chorus of economists is warning that if the answer comes back wrong, the pain will not stay inside the tech sector.
Here is the setup, the counter-argument almost nobody is pricing in, and what it means if you build, invest, or operate anywhere near this market.
The $3 trillion question: The total revenue the AI industry must eventually generate to justify the roughly $1.5 trillion in data center and chip spending going into 2026 alone, once you add operating costs and investor return expectations on top of the hardware. It is the AI era’s version of a simple test every business faces: does the money coming in eventually cover the money going out?
Where the $3 trillion number comes from
Back in 2023, Sequoia’s David Cahn looked at Nvidia’s reported annual GPU revenue of around $50 billion and worked backward. Add in the cost of running the data centers those chips live in, layer on the margins the operators need, and he calculated the industry would have to produce roughly $200 billion in revenue just to pay back the up-front investment. He framed it as a challenge to founders: go build products worth all that compute.
Fast forward through three years of hyperscaling. Cahn’s updated figure for AI infrastructure spending in 2026 is $1.5 trillion. Stack up the full bill, and he now estimates the industry needs to earn about $3 trillion to justify the chips and data centers. He also thinks that is probably a floor, not a ceiling, because rising memory costs and the shift toward exotic, inference-specific silicon keep pushing the required return per gigawatt of capacity higher.
If you want the wider frame on this buildout, we mapped the full spending picture in our look at the best AI companies to watch in 2026, where Goldman Sachs pegs hyperscaler capex from 2025 through 2027 at $1.15 trillion, more than double the prior three years.
The revenue side does not add up yet
Now flip to the other side of the ledger. The frontier labs are growing fast, but the gap is still enormous.
| Company | Reported figure | Notes |
|---|---|---|
| Anthropic | ~$60 billion ARR | Estimated run rate |
| OpenAI | ~$13 billion (2025 revenue) | Said $20 billion ARR in November 2025, likely higher now |
Even generously, the frontier labs’ combined run rate sits in the low hundreds of billions against a $3 trillion bill. And OpenAI’s revenue comes attached to staggering losses. Audited 2025 statements showed a $38.5 billion net loss on that $13 billion in revenue, a picture we broke down in our OpenAI leaked financials analysis. Revenue is not the same thing as profit, and in this industry the distance between them is measured in compute.
The 2028 deadline nobody is talking about
Here is the part that turns a tech story into a macro story. Torsten Slok, chief economist at asset manager Apollo, points out that the four hyperscalers footing most of this bill, Google, Meta, Microsoft, and Amazon, are all forecasting a massive acceleration in free cash flow starting in 2028. In plain terms, they are telling investors the payback arrives in about two years.
That is a promise, not a fact. And it sets a clock. If the cash does not show up on schedule, the story that has propped up some of the most valuable companies on earth starts to wobble.
The token deflation problem
Slok’s specific worry is one this site has tracked for months: the price of AI is falling, and it is falling for reasons the labs cannot fully control.
More organizations are switching to cheaper open-weight models, often Chinese ones, instead of paying frontier prices. We covered exactly this when a free, MIT-licensed Chinese model matched a restricted US frontier system at roughly a sixth of the cost. On top of that, the frontier models themselves keep getting more efficient. OpenAI’s latest release is, per Sam Altman, 54% more token-efficient on coding tasks.
For a company running AI agents, that efficiency is a gift. For a company that built a token factory and needs those tokens sold, it is a threat, unless total usage climbs fast enough to more than offset the falling price per token. That tension is the heart of the enterprise token-bill scramble we documented, where finance teams discovered that cheaper tokens did not mean cheaper bills.
The counter-argument almost nobody is pricing in
Here is where the bear case has a hole. The assumption that falling token prices shrink revenue quietly ignores Jevons paradox: when something gets cheaper, people often use so much more of it that total spending rises anyway.
We have seen this movie already. In January 2025, when DeepSeek proved a Chinese lab could match Western reasoning models cheaply, the market concluded that cheaper intelligence meant collapsing chip demand and erased roughly $600 billion from Nvidia in a single session. The thesis was exactly backward. Cheaper intelligence expanded compute demand, and Nvidia went on to cross $5 trillion in value. We laid out the full mechanism in our margin migration breakdown, which argues the real story is not whether AI pays off but where the money pools as it flows down the stack.
So the honest read on Slok’s warning is that a 54% efficiency gain is only bearish if usage stays flat. Every prior efficiency jump in this cycle triggered more consumption, not less. The bulls are betting demand keeps outrunning price. The bears are betting it finally does not. Nobody actually knows which, and that is the whole game.
Why “so few names” is the real risk
The scariest line in Slok’s note is not about AI at all. It is about concentration. With so much of the market’s value riding on so few companies, he argues a slower-than-expected payoff would not stay contained as a sector problem. It could tip the broader economy toward recession and the S&P 500 into a correction.
That is the difference between 2023 and now. Three years ago this was a venture-capital math problem. Today the same handful of stocks carrying the AI bet are also carrying the index, which means the ROI question has quietly become everyone’s question, whether or not you own a single AI product.
What this means if you build or operate
Three practical takeaways, regardless of where you sit:
First, if you sell AI-powered products, price for a world where the underlying model cost keeps dropping. Margin captured on top of a commoditizing input is durable. Margin that depends on the input staying expensive is not.
Second, if you buy AI, the efficiency gains are real and worth chasing, but watch your total spend, not your per-token price. The companies that got burned this year were the ones who confused a cheaper unit with a cheaper bill.
Third, if you are exposed to the market at all, understand that the AI trade and the index trade are now the same trade. The 2028 free-cash-flow promise is the load-bearing assumption underneath both.
FAQ
How was the $3 trillion figure calculated? Sequoia’s David Cahn started with roughly $1.5 trillion in projected 2026 AI infrastructure spending, then added operating costs and the return operators and investors expect on that capital. He considers $3 trillion a conservative estimate because memory and specialized-chip costs are still rising.
Are the AI labs anywhere close to that revenue? No. Anthropic is estimated near $60 billion ARR and OpenAI reported about $13 billion in 2025 revenue. Even combined, the frontier labs are far below the required figure, and OpenAI in particular is losing far more than it earns.
Why would this affect the whole economy? Because a small number of companies now drive both the AI buildout and a large share of the S&P 500. If their promised 2028 payback slips, Apollo’s Slok warns the fallout could push the economy toward recession and the index into a correction.
Is cheaper AI actually bad for the industry? Only if usage stays flat. Historically, every drop in AI cost has driven a larger jump in usage, which is why the bears and bulls are really arguing about demand, not price.
The Business Model Analyst Take
The $3 trillion headline is designed to scare you, and the concentration risk it points at is real. But the logic underneath Slok’s specific warning is shakier than it looks. Betting that falling token prices will starve the AI economy is the same bet that wiped $600 billion off Nvidia in early 2025, right before compute demand went vertical and the stock tripled toward $5 trillion. Efficiency has not shrunk this market once. It has expanded it every single time.
The sharper question is not whether AI earns back $3 trillion. It is where that revenue lands as it moves through the stack. The frontier labs are burning record cash on products that are converging toward interchangeable, which means the money is leaking away from them and pooling somewhere else: in compute, in cloud, and in the disciplined operators who point AI at the few jobs where it actually pays. If you want to be on the right side of the $3 trillion question, stop asking whether the industry pays off and start asking who captures the margin when it does. That is the answer worth building around.
