Keywords: AI capex, AI revenue, David Cahn Sequoia, AI bubble, hyper scaler capex, AI ROI, data center capex, AI economy 2026.The largest funding round in the history of corporate expenditure in AI. Every hyper scaler ( Microsoft, Google, Amazon, Meta) and chip-maker ( Nvidia) is pouring billions of dollars into AI infrastructures, data centers and GPUs. However, under the fanfare, there remains one question: will AI be able to make any money This is now known as the "$3 trillion question", and looks to be shaping up as the financial test of the AI age.
Where the $3 Trillion Number Came From
2023 was the year when this all started, when Sequoia Capital partner David Cahn, who was among the minority of people in Silicon Valley at the time to do the numbers, and by accounting also for the costs of datacenters at the margins that the datacenter creators were otherwise planning – took Nvidia's stated annual revenue from GPUs of c. $50 billion, and accounted for datacenter costs felt that the AI industry would ultimately need to capture around one-sixth of global GDP of $200 billion per year to be sustainable.Prior to that, it wasn't really framing it as a warning - Cahn was challenging entrepreneurs; to produce AI products and services that were able to leverage all the new computing capacity.
Moving on three years and the buildout is running so far ahead of expectations that Cahn's revised projection for the buildout for 2026 alone for AI infrastructure is now nearly$1.5tn. If you add in finance costs, sum costs, the return on shareholder investment AND allow for the likely demand-then which is sellingk3tn-worth of products cycling towards chips,GPUs and datacenters-in the chain-we are in the right ballpark.
And that's probably on the low side, Cahn suggests-though, with higher costs for memory and ever more inference-specialized chips, the dollar per gigawatt revenue line for data center capacity is, he says, "starting to push well past $1 trillion." It may very well get even higher, to more than $3 trillion.
Morgan Stanley Sees a Similar Gap
He's not the only one sounding the alarm. A report commissioned by Morgan Stanley itself, discussed on the bank's Thoughts on the Market podcast by its analysts Stephen Byrd, Josh Baer and Lindsey Tyler, produces a picture very similar to Cahn's, but starting out from quite a different premise. By estimating the global capital investment needed for data centers turned out to be about $3 trillion in four years.From that perspective, there is a clearly defined playing field that high-end hyperscalers could raise roughly half of that via their own internal operating cashflows, while the rest of the shortfall-possibly north of a trillion dollars-battle has to be borrowed on the credit and fixed income markets, turning Wall Street's bond desks just as crucial to the AI story as Silicon Valley engineers.
The Revenue Side: Impressive, But Still Far Behind
So how close is the entire industry, as a whole, to closing the gap? To be honest, we're not close at all.Although this figure is a guess, estimates say that the second is raking in roughly $60 billion annually in recurring revenue. Openai claimed that was worth $13 billion in profitable revenue In 2025 and that by the end of the year, was on pace to hit $20 billion in sales and gaining going into 2026.That is very good growth for a new company-but as part of a $3 trillion bill, that is only a small fraction of the revenue necessary to justify this expenditure on infrastructure. In fact, to break the trillion-dollar mark, you would need to add all revenue generated by all frontier AI firms to the mix.
The Falling Price of Intelligence
However there's one more thing add to the chaos, there's a quiet trend that's been simmering for some time already- "AI is becoming cheaper to use even as it's being used more."Torsten Slok, Chief Investment Economist, Apollo, has pinpointed a risk of this. More and more firms-big coa's et al.-are experimenting with cheaper, open-sourced, open-weight AI models-many of them located in China or short, as a net, of voluntarily freezing their reliance on laboratory frontier models from, say, OpenAI, Anthropic, and so on. And at the same moment, overall token prices-the basic unit of price for AI-are crashing across the board.
Even the frontier labs are contributing to the downwards pressure. OpenAI's newest model is claimed to be 54% more token efficient for coding tasks, claimed CEO Sam Altman (A better Carbon footprint for AI developed). Handy for companies and developers who are concerned with the cost of operating AI agents - but possibly catastrophic for companies who rely on people ingesting tokens at large volumes (per token prices falling so quickly that revenues level off prior to the growth in access being able to compensate).
Why This Isn't Just a Tech Industry Problem
One might be inclined to write it off as a problem of Silicon Valley, yet Slok has warned that the implications have a far-reaching impact. A small number of AI related companies now represent an uncommonly high proportion of the total value of all stock market value. As a consequence, the prosperity of the AI industry and the overall stock market S&P 500 have become deeply intertwined.If the hyperscalers fall short of their anticipated cash-flow targets, Slok forewarns that investor reaction could be rapid and punishing enough to rock the entire world economy into a recession-not merely a correction in the tech universe. With so much riding on so few companies, if the achievement isn't worth the hype, it will go beyond the AI share.
So, Can AI Really Answer the $3 Trillion Question?
Perhaps a more useful question than will AI ever make back the 3 trillion dollars in total is where does that 3 trillion dollars go?As products at the application level start to look similar, competitive separation will get harder. That suggests that margins may not stay confined to frontier labs creating the base models, but may move south: toward the layers below, like cloud infrastructure companies and compute providers, and beyond that, more niche firms taking AI to solve specific problems, where its benefits are clear and undeniable, rather than chasing the single, all-dominating, mythical AI messiah.
There are three key lessons that you can learn about whether to buy or sell in the current climate that applies to both businesses, investors or individual users:
If you sell anything based on Aidanote:consider pricing for a world in which the world average cost of models ever falls. A mark up on top of an input that is getting ever cheaper is its own everlasting defender; a mark up that depends on that input still being expensive is not.
If using or buying AI tools: Watch your overall cost and not only the minimum token. At this year many companies were surprised that,companies were providing at lower unit price so the usage increased which eventually balance out your bill.
If you're participating in finance in some way: just understand that the AI trade, and the whole stock index trade, have become one and the same trade. This is the 'load-bearing' central post supporting both, of hopeful free cash flow by 2028.
The Bottom Line
If you’re looking at AI infrastructure spend in the last 3 years and compare the needed revenue that should justify it - it’s increased by 15x - from $200B to $3T. There’s still, however, some maths which just doesn’t add up even though companies like Anthropic and (now) OpenAI are reporting genuine revenue growth. As we see the high-end, paid, and costly providers getting replaced by cheaper alternatives, and with token prices tumbling, the pressure is on the hyperscalers to deliver on the promised multi-year cash-flow growth and revenue that they committed to over several years.
If you’re looking at AI infrastructure spend in the last 3 years and compare the needed revenue that should justify it - it’s increased by 15x - from $200B to $3T. There’s still, however, some maths which just doesn’t add up even though companies like Anthropic and (now) OpenAI are reporting genuine revenue growth. As we see the high-end, paid, and costly providers getting replaced by cheaper alternatives, and with token prices tumbling, the pressure is on the hyperscalers to deliver on the promised multi-year cash-flow growth and revenue that they committed to over several years.
Over the last 3 years, the amount invested in AI infrastructure has jumped by a factor of 15 – from $200bn to $3tn. While there’s real revenue coming from OpenAI, and most recently Anthropic, some of the numbers don’t add up yet. Lower-cost models eating into usage at high-end, with tokens continuing to lose value, mean that the pressure on the hyperscalers is increasing for them to deliver on their run-rate cash-flow projections over the multi-year period.
Whether AI can answer its own $3trn question – however, the whole world appears to be betting yes.
FAQs Q1.
What is the '3 trillion dollar question' in AI?
It's how much money AI has to make to justify the chips, GPUs and the datacenters spent in 2026 (estimated $1.5tn in spending to make up for $200bn for that year as projected by David Cahn).
Q2.Who is David Cahn and why does his number matter?
David Cahn is a partner at Sequoia Capital.In 2023, his group was one of the earliest to produce figures predicting how much revenue AI must generate in 2026 to make the underlying investments, a figure he began at $200bn. This recent upward revision to $3tr will be keenly watched, as it highlights just how fast AI infra spending – and risk – has accelerated.
Q3.What is the explanation for the magnitude increasing from $200 billion to $3 trillion?
Over the past three years of so-called “hyperscaling” or increasing investment in GPUs and related infra and compute, combined with increased price on memory and deployment of inference dedicated hardware, it’s projected to be even higher on an increase in revenue per Gigawatt of capacity.
Q4.How much monetary value are AI companies generating right now?
So far Anthropic is doing about $60Bn on an annua recurring revenue basis, whereas OpenAI generated $13Bn for the 2025 fiscal year and has a current run rate of about $20Bn, both of which are a tiny percentage of the $3Tr the industry requires.
Q5.Why are falling token prices a risk?
The cost per token is expected to fall due to competition from less costly alternatives (many originating in China).
Without sufficient growth in the use of the models to offset the decrease in token prices per usage unit, many businesses using token count to determine profitability will have difficulty scaling up profits.
Q6.However, can this turn into a wider economic issue beyond tech stocks?
According to Torsten Slok, chief economist at Apollo, as the top fewAI-related companies constitute a disproportionate segment of the total stock market capitalization, a below-average result from theAIbet could have ripple effects across other technology stocks, resulting in a market correction or even the risk of recession.
Q7.If AI does ultimately work, who wins the most?
The primaryAIcompanies themselves might not benefit as much, as manyAIapplication layers have already been made generic, or become commoditized, with providers who apply AI to specific high value areas rather than a broad-brushed AI story.
Q8.What conclusions should businesses and investors draw from this?
There are three conclusions with less clear implications – 1. Assume future price of AI products reflect the ongoing decline of costs to train models 2. Instead of looking at the cost of each token, monitor the overall investment inAI3. The correlation between the AI investment trade and the stock market as a whole – both are bet on Hyperscaler cash-flow for 2028 and higher.