What Counts as 'AI Revenue'? Mostly Numbers That Predate the Boom
Hyperscalers have quietly folded a decade of ordinary machine learning into the 'AI revenue' they report, just as those figures are being used to justify historic capital commitments. The denominator is contaminated, and boards are building on it.
The most important number in technology right now is one nobody audits. 'AI revenue' is the figure vendors report, boards benchmark against and analysts stretch into trillion-dollar market sizes. Look closely at what sits inside it and a large share is old revenue wearing a new badge: ad targeting, recommendation engines, search ranking, fraud detection. Machine learning, certainly. Generative AI, mostly not.
That matters because a denominator problem is never just a reporting problem. Every ROI model and market-sizing exercise inherits the inflation, and so does every 'our competitors are monetising AI faster than us' board slide. Feed a contaminated base rate into a forecast and it fails quietly, at scale.
What counts as AI revenue?
Almost anything, at the moment, which is the problem. The category has no accounting definition and no auditor. So when the boom arrived, a decade of pre-transformer machine learning was grandfathered in: systems that were profitably ranking adverts and recommending products long before anyone typed a prompt now sit in the AI column. The logic has been stretched to the point where the industry can assert, more or less openly, that essentially all platform revenue is AI revenue. If everything counts, the category measures nothing, and that may be the point.
The incentive is straightforward. These are the figures being used to validate infrastructure commitments measured in the hundreds of billions. A big AI-revenue number tells investors the demand is already here, and the reclassification costs nothing while moving the market. My base rate for any unaudited, undefined, incentive-laden metric is that it flatters whoever reports it. Nothing in the current disclosures updates me away from that.
The demand signal is partly manufactured
Set reclassification aside and the demand picture still needs discounting. Consider the arrangement between Nvidia and OpenAI. In the companies' joint announcement, Nvidia says it intends to invest 'up to $100 billion' in OpenAI, released progressively as OpenAI deploys ten gigawatts of Nvidia systems, and reporting on the deal in the Wall Street Journal put the initial tranche at $10 billion. A chipmaker investing tens of billions in its most important customer, so that money flows out and comes back as chip orders. None of this is hidden and none of it is illegal. But it is circular, and circular financing manufactures a demand signal that looks organic until you trace the loop. One party locks in demand, the other locks in valuation, and the headline does the rest.
The headline deserves its own scepticism, because the load-bearing words in that announcement are the qualifiers: 'up to', not 'will', and investment gated on deployment rather than cash at signing. The announced figure exists to move a valuation. The committed figure, the one with signatures and near-term cash attached, is the initial tranche, a tenth of the headline number. Treating letters of intent as commitments is how a probable overbuild gets rounded up to a certain one.
Then there is the game theory. A public-company chief executive who announces restrained, evidence-led AI spending gets punished by the market the same afternoon. The rational move is to profess maximal capex regardless of what the internal ROI dashboards say, and every player knows every other player faces the same payoff matrix.
That is a prisoner's dilemma, and its output is spending decoupled from returns: not because executives are fools, but because the incentive structure pays them to overbuild in public.
When does AI capex become everyone's problem?
Here is the distinction that matters more than any bubble debate: how the build-out is funded. While it is paid for out of corporate cash flows, a correction is a shareholder problem, painful but contained. The moment the marginal data centre is debt-financed, the risk changes character, because debt is the transmission mechanism that turns a sector's write-down into an economy's problem. The scale is now on the record. Axios's tally of the 2026 build-out puts combined AI capital spending by the major technology companies, Oracle included, near $800 billion, while noting that none of them discloses the sales or profits attributable to it. Meta alone tells investors in its first-quarter 2026 Form 10-Q to expect approximately $125 billion to $145 billion of capital expenditure this year. Set spending at that pace against the operating cash flow Microsoft, Alphabet, Amazon and Meta report each quarter and, on my arithmetic, the build-out consumes the bulk of it, plausibly all of it once buybacks and dividends leave the door. Treat that ratio as my estimate rather than a disclosed figure: no company publishes it, which tells you something in itself. When internal cash stops covering the bill, the balance arrives as borrowing.
The timing problem compounds it. The filings depreciate this hardware gently: Microsoft's fiscal 2022 annual report records its decision to 'increase the estimated useful lives of both server and network equipment from four years to six years', and its peers book similar schedules. My own estimate of a frontier accelerator's economic life, the period before newer silicon outcompetes it for the work it was bought to do, is harsher: perhaps three to four years, roughly one product generation. Work backwards from the scale of committed spend on that harsher assumption and the revenue required to justify it implies a new trillion-dollar software market materialising inside a single chip generation. That is possible. Whether it is probable enough to lever a balance sheet against is a different question, and one not being asked out loud.
There is a quieter technical asymmetry underneath. The capex thesis is premised on ever-larger frontier training runs, yet the visible action has moved to post-training, inference-time compute and radical efficiency, with open-weight models reproducing expensive capabilities for a small fraction of the original cost. If capability per pound keeps falling, compute demand may keep growing while the pricing power needed to repay this particular vintage of hardware does not. That is the scenario no headline number prices in.
What would change my mind
Position stated, here is the evidence that would flip it. Independently verifiable segmentation showing genuinely incremental, generative-attributable revenue growing faster than the reclassified base. Capex ratios stabilising against operating cash flow rather than climbing past it. Announced commitments converting into cash at high rates instead of quietly lapsing. Show me two of those and the boring story, an unusually loud but fundamentally ordinary infrastructure cycle, becomes the better bet. So far the disclosures run the other way.
For a board, all of this converts into a single diligence move. When a supplier, a consultancy or your own strategy deck quotes an AI revenue figure, demand the incremental cut: revenue that exists only because of generative AI, net of everything that was already being sold under an older name. Treat the grandfathered total as marketing. It is the same discipline behind establishing AI readiness before you build: work out what the technology actually changes for your economics before committing capital, ideally with an independent technical strategy view that is not paid to inflate the denominator.
The asymmetry to hold onto is this: if the sceptics are wrong, careful buyers lose a little time. If the headline numbers are wrong, levered builders lose the decade. Price capital for the second mistake, not the first.
Questions people ask
Is AI revenue the same as generative AI revenue?
No. Most reported 'AI revenue' includes machine-learning products that predate generative AI entirely, such as ad targeting, ranking and recommendation systems. Genuinely incremental generative revenue is a far smaller subset, and vendors rarely break it out unless pushed.
How should a board test a vendor's AI revenue claims?
Ask for the incremental, generative-attributable cut: revenue that would not exist without generative AI, net of reclassified legacy lines. Ask how the category is defined, whether prior periods were restated, and treat any refusal to disaggregate as an answer in itself.
Does a plateau in AI scaling mean the spending stops?
Not necessarily. Investment is shifting towards post-training, inference-time compute and efficiency rather than ever-larger training runs. Compute demand can keep growing while the returns on this specific generation of hardware still disappoint, which is why funding source matters more than volume.
Related
- Retention Bonuses Don't Reduce Key Person Risk. Doctrine Does.
- Bot Traffic Is Splitting Your Ad Budget, and a Human Signed It Off
- Your Cloud Provider Is Financing Its Own AI Revenue. Do the Financial Due Diligence
- Data & Strategy
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.