Sequoia's $600B AI Revenue Gap, Rerun: In 2026 It Is a $1.4 Trillion Question

Nvidia booked $89.0 billion of data centre revenue in the quarter to 26 July 2026. Run that single number through the arithmetic David Cahn used for Sequoia's "$600B question" in 2024, and the AI revenue gap becomes a $1.42 trillion requirement. The disclosed revenue of the two largest AI labs covers about 8% of it. The hole is far wider in dollars than it was in 2024. The revenue meant to fill it is growing much faster than the hole.

What Sequoia's $600B question actually measured

Sequoia partner David Cahn published "AI's $600B Question" on 20 June 2024. It updated his September 2023 essay, which had framed the same problem as a $200 billion question with a $125 billion hole. The 2024 version became the most quoted single number in the AI bubble debate.

The essay is short, and the method fits in two lines. That brevity is why it spread. It is also why it gets misquoted, because people repeat the $600 billion headline without the two assumptions underneath it.

The two doublings

Cahn started with Nvidia's run-rate revenue forecast. He multiplied it by 2 because, in his words, "GPUs are half of the total cost of ownership." The other half covers energy, buildings and backup generators.

He then multiplied by 2 again, "to reflect a 50% gross margin for the end-user of the GPU." The logic is that whoever rents the compute has to sell something worth twice what they paid. The essay does not print the Nvidia figure it started from. Working backwards, $600 billion divided by 4 implies a run rate near $150 billion.

The revenue he subtracted

Cahn then subtracted the AI revenue he could plausibly see. He "generously assumed" $10 billion each of new AI revenue for Google, Microsoft, Apple and Meta. He added $5 billion each for Oracle, ByteDance, Alibaba, Tencent, X and Tesla. He cited OpenAI at $3.4 billion.

Those named figures sum to about $73 billion. The essay states the result as "the $125B hole is now a $500B hole," which implies roughly $100 billion was netted off. The text does not reconcile the two. This matters for the rerun, because the 2024 revenue side was an assumption, not a disclosure.

The 2026 input: Nvidia's $89.0B data centre quarter

The rerun uses one primary document for the cost side. Nvidia's results for the second quarter of fiscal 2027 reported total revenue of $96.2 billion, up 106% year on year. Data centre revenue was $89.0 billion, up 117% year on year and 18% on the prior quarter.

Using a reported quarter rather than a forecast is a deliberate choice. Cahn's essays start from a projected run rate. A reported quarter is smaller and less exciting, and it cannot be argued with. If the gap is large on reported numbers, it is not an artefact of someone's forecast.

From one quarter to a run rate

Multiplying the quarter by 4 gives a data centre run rate of $356 billion. That is about 2.4 times the roughly $150 billion implied by the 2024 essay. Nvidia's own gross margin of 75.0% in the quarter is not part of the method, which only cares what customers paid Nvidia.

The run rate is a floor, not an estimate. Nvidia guided third-quarter revenue to $108.0 billion, plus or minus 2%, and assumed no data centre compute revenue from China. If the data centre share holds at its second-quarter level of about 92.5%, the next run rate lands near $400 billion. That figure is my extrapolation, not company guidance.

The AI revenue gap rerun: $1.42 trillion required

Apply both doublings to the $356 billion run rate. Total data centre cost becomes $712 billion. The end-customer revenue needed to give every renter a 50% gross margin becomes $1,424 billion, or $1.42 trillion. My view is that this is the cleanest single statement of the AI revenue gap, because every input is either a filing or a stated assumption.

The table puts the two versions side by side. The 2024 column uses Cahn's own published figures. The 2026 column uses Nvidia's filing and the two largest disclosed lab revenue lines, and nothing else.

The 2024 revenue line was a generous assumption about big tech, not a measurement. The 2026 line is narrower and real. Comparing the coverage rows therefore flatters 2024. On disclosed revenue alone, 2024 coverage was a small fraction of 17%.

Cahn's own update lands in the same place

Cahn reran the numbers himself in "AI's $1.5T Question" on 8 July 2026. He applied the same two doublings to Nvidia's projected fourth-quarter run rate. He cross-checked it by doubling roughly $750 billion of expected 2026 hyperscaler data centre capex.

Both routes gave him $1.5 trillion. The reported-quarter rerun here lands at $1.42 trillion, close enough that the method is clearly stable. What his update does not do is total the revenue side. That is the gap this post tries to fill, and it is also where the arithmetic gets contested.

What disclosed AI revenue can be counted against the gap

Cahn's test asks about end-customer revenue. That is money paid by people and businesses who use an AI product, not money one AI company pays another. Getting this right matters more than any single figure, because the AI economy is full of revenue that is someone else's cost.

The lab run rates

Anthropic's annualised run rate passed $65 billion at the end of July 2026, according to Bloomberg reporting carried by TechCrunch. That was up from $47 billion in May and $9 billion at the end of 2025. Anthropic did not comment on the figure.

OpenAI told investors its run rate was approaching $50 billion at the end of September, the Financial Times reported, via Reuters. That was about $20 billion below a widely circulated figure near $70 billion. The difference is accounting. Anthropic counts gross sales made through cloud partners such as AWS and Google Cloud, and OpenAI counts only its share.

For Cahn's test, the gross basis is arguably the right one. A business buying Claude through a cloud marketplace is still an end customer paying for an AI product. On that reading, OpenAI's comparable figure sits nearer the $70 billion investors had normalised to, and lab coverage rises from $115 billion to about $135 billion.

Microsoft and the double-count problem

Microsoft disclosed that its AI business passed a $37 billion annual run rate, up 123% year on year, in results published on 29 April 2026. It is the largest AI revenue figure any public filer has put in a press release.

It cannot simply be added to the lab totals. A material share of Microsoft's AI revenue is Azure capacity sold to OpenAI, which is OpenAI's cost rather than an end customer's spend. Adding both counts the same dollar twice. Microsoft does not split the figure, so the honest treatment is to exclude it from the floor and show it as an upper bound.

Taking every row at face value, including Microsoft and the grossed-up OpenAI figure, gives about $172 billion, or 12% of the requirement. That total double counts and should be read as a ceiling.

The last row is the real limitation of any revenue gap calculation. The hyperscalers with the most AI revenue disclose the least about it. Our ranking of how companies disclose AI revenue shows why the numbers above are the only clean ones available.

The AI revenue gap is wider in dollars and narrower in growth

The headline version of this story is that the gap grew 2.6 times in two years. That is true, and I think it is the less important number. The requirement is driven by Nvidia's revenue, which grew about 2.4 times between Cahn's two essays. The revenue meant to fill it grew several times faster.

OpenAI's run rate went from the $3.4 billion Cahn cited in June 2024 to roughly $50 billion at the end of September 2026. That is about 14.7 times. Anthropic went from $9 billion at the end of 2025 to above $65 billion at the end of July 2026. That is about 7.2 times in 7 months.

Why the coverage ratio still fell

If revenue is growing faster, why did coverage fall from about 17% to about 8%? Because the 2024 figure was not real revenue. Cahn's $100 billion was a generous guess about what big tech might earn. The 2026 figure is narrower, counts only two companies, and is press-reported but real.

On disclosed revenue, 2024 coverage was close to nothing. OpenAI's $3.4 billion against $600 billion is under 1%. On that like-for-like basis, the two labs alone moved coverage from under 1% to about 8% in roughly two years. That is the finding I would put in front of a board.

How sensitive the AI revenue gap is to the markup

Both doublings are assumptions, and Cahn says so. The first assumes GPUs are half of total data centre cost. The second assumes the renter of the compute needs a 50% gross margin. Small changes to either move the answer by hundreds of billions, so the table below shows what happens when you loosen them.

All rows use Nvidia's $356B data centre run rate except the last, which assumes the Q3 FY2027 guide of $108.0B keeps the Q2 data centre share of about 92.5%. Coverage uses the $115B lab floor. Zan Digital calculations.

Even the most generous row leaves a gap larger than the 2024 hole. My position is that no plausible markup closes this in 2026. The useful debate is not whether a gap exists, but how many years the revenue has to close it. That depends on asset life, which is the subject of the next section.

Where this rerun is weakest

Every rerun of the $600B question published so far skips at least one of these. I would rather list them than have a reader find them.

It compares a lifetime bill with one year of revenue

Cahn describes his output as "the lifetime end-customer revenue requirement for a single year of CapEx." The $1.42 trillion does not have to arrive in a single year. It has to arrive across the working life of the hardware, which nobody outside these companies can measure.

As a purely illustrative assumption, spread the requirement over 5 years. The annual bill becomes about $285 billion, and the two labs alone cover about 40% of it. That assumption is mine, not a disclosure, and the right life is contested. Our analysis of how GPU depreciation schedules move AI earnings covers why. The point is that the headline gap overstates the annual shortfall, possibly by a lot.

The opposite pressure also applies. Each new year of capex adds another lifetime requirement on top. Cahn's own cumulative figure since ChatGPT launched is about $3 trillion. A stacked requirement grows faster than any single year suggests.

Run rates are not revenue, and the labs are not audited

Both lab figures annualise a recent month. Both come from press reporting of investor documents or people familiar with the numbers, not from a filing. OpenAI's figure was effectively revised down by about $20 billion in October, which shows how much the basis matters. Treat the $115 billion as directional until a prospectus exists, a point developed in our breakdown of what OpenAI's run-rate multiple assumes.

Some of the revenue is circular

Part of what the labs earn is ultimately funded by the same suppliers selling them compute. Nvidia has invested in both OpenAI and Anthropic. A dollar of revenue that started as a supplier's investment is weaker evidence of demand than a dollar from an unrelated customer. The structure is set out in how the AI circular deals actually work, and this rerun makes no attempt to net it out.

Frequently asked questions

How to use this before your next budget review

Rebuild the calculation yourself after each Nvidia report, which takes ten minutes. Take the data centre line from the release, run the two doublings, and subtract only revenue you can trace to a named disclosure. Track the coverage percentage, not the gap. The gap will keep growing in dollars for as long as Nvidia grows, and that alone tells you little.

Then ask the question that matters for your own spend. If your AI vendors are funded by investors betting that coverage keeps rising, their prices reflect that bet. Write down what you would do if a key vendor raised prices by a third at renewal, because a slower close of this gap is how that happens. The cost side of that risk is laid out in the AI capex payback math.

Last verified 25 October 2026. Nvidia's next results, for the third quarter of fiscal 2027, will reset the cost side of this calculation. Lab run rates move monthly.

Related reading