Is AI a Bubble?

Executive summary

Is AI a bubble? The honest answer from the public data is that it depends on which number you look at, and serious people read the same figures in opposite ways. On the bull side, the demand is unmistakable: ChatGPT reached 900 million weekly active users by February 2026 (OpenAI), OpenAI's annualised revenue run-rate passed roughly $25 billion and Anthropic's reached about $47 billion by mid-2026 — growth rates no dot-com company ever produced. On the bear side, valuations have detached from revenue by historic margins. On Affärslivet's reading of Epoch AI revenue data and reported funding rounds, OpenAI is priced at about 34 times revenue, Anthropic about 21 times, and xAI — with only a few hundred million dollars of revenue — more than 500 times, against roughly 5-15 times for mature software. Meanwhile the money going in dwarfs the money coming out: the five largest US hyperscalers are on course to spend $700-900 billion of capex in 2026, while Bain & Company estimates the industry needs $2 trillion of new annual AI revenue by 2030 to justify the build-out and may fall about $800 billion short. A widely cited MIT Project NANDA study found roughly 95% of enterprise generative-AI pilots produced no measurable profit impact. Critics also point to 'circular financing' linking Nvidia, OpenAI, Oracle and CoreWeave. Skeptics from Michael Burry to, in their own words, Sam Altman and Jeff Bezos say a bubble is forming — even as several argue the underlying technology is real. This report lays out both cases from named sources so you can weigh them yourself. This is not financial advice.

Key findings

What would it actually mean to say AI is a bubble?

A bubble is when the price of an asset rises far above the cash flows it can plausibly generate, sustained by expectations of future gains rather than present fundamentals — so calling AI a 'bubble' is a claim about valuations and capital spending, not about whether the technology works. The two questions are separate: the internet was transformative and the dot-com bubble still burst. AI can be genuinely useful and simultaneously over-financed.

In practice, the AI-bubble debate is about three measurable things. First, valuations: are private AI labs and public AI-exposed stocks priced at multiples that only make sense if enormous future revenue arrives? Second, capital spending: is the money being poured into chips and data centres justified by the revenue those assets can earn? Third, returns: are the companies actually deploying AI making measurable money from it yet?

This report takes each in turn using named, dated sources, then lays out the bull case and the bear case side by side. It deliberately reaches no verdict. Whether these facts constitute a 'bubble' is a judgement each reader must make. This is not financial advice.

What do AI valuations actually look like?

The most striking data in the whole debate is the gap between what leading AI labs are worth on paper and how much revenue they earn. Using Epoch AI's revenue tracking alongside publicly reported funding rounds, Affärslivet calculated valuation-to-revenue multiples for the frontier labs as of 2026. Mature, profitable software companies typically trade at roughly 5-15 times revenue; several AI labs are many times higher.

The pattern is clearest at the extremes. OpenAI and Anthropic, with the largest revenues, sit at 'only' 21-34 times — rich, but in the range hyper-growth software has reached before. xAI, valued at around $230 billion on a few hundred million dollars of revenue, sits above 500 times, and China's Z.ai (Zhipu) above 250 times. These are early-stage companies where a single year of growth can reprice the multiple dramatically.

Two caveats matter. These are private-round marks negotiated by a handful of investors, not liquid public prices — they can be stale or optimistic. And revenue is compounding: Anthropic's run-rate reportedly went from about $1 billion at the end of 2024 to roughly $47 billion by mid-2026. A high multiple today can look ordinary in eighteen months if growth holds — or absurd if it stalls.

How do the AI labs compare on one number: valuation-to-revenue?

Stripped to a single comparable metric, the valuation-to-revenue multiple shows how far each lab's price sits above its current sales. The chart below plots that multiple for the frontier labs against a reference band for mature enterprise software (~10×). The higher the bar, the more of the valuation rests on future revenue that has not yet arrived.

The spread is enormous: from Anthropic and OpenAI at 21-34×, through Z.ai at 258×, to xAI above 500×. Two labs at 34× (OpenAI and Mistral) reach that figure from opposite ends — one with $25 billion of revenue, one with $0.4 billion — a reminder that the multiple alone does not tell you which businesses are durable. Figures are Affärslivet estimates derived from Epoch AI revenue data and reported valuations; treat them as order-of-magnitude, not precise prices.

How big is the gap between AI spending and AI revenue?

Beyond the labs, the bigger number is infrastructure: the five largest US hyperscalers — Amazon, Microsoft, Alphabet, Meta and Oracle — are on track to spend an estimated $700-900 billion of capital expenditure in 2026, the bulk of it on AI data centres and chips, a sharp increase over 2025. The question is whether the revenue exists to earn a return on that hardware.

Bain & Company's 2025 Technology Report put a figure on it: to profitably meet projected AI compute demand, the industry would need roughly $2 trillion in new annual AI revenue by 2030 — and Bain warned it could fall about $800 billion short on current trajectories. Independent estimates of today's actual AI revenue across the ecosystem sit in the low hundreds of billions, so the arithmetic requires steep, sustained growth to close.

Bulls note this is how every major infrastructure build works — railways, electricity, cloud — where spending precedes revenue by years and the assets are productive for a long time. Bears note that GPUs depreciate fast (a large share of that capex is chips with a useful life measured in a few years, not decades), which makes the timing of returns far less forgiving than fibre or rail.

Is AI actually making money for the companies using it?

The demand-side evidence is strong at the consumer level and weaker at the enterprise level, and that split is central to the debate. On one hand, ChatGPT reached 900 million weekly active users by February 2026 (OpenAI) and crossed a billion monthly app users by mid-2026 — genuine, paid, recurring usage that pets.com never had. On the other, the return for businesses deploying AI is much harder to find in the accounts.

The most-cited datapoint is MIT Project NANDA's 2025 study, which reported that roughly 95% of integrated enterprise generative-AI pilots showed no measurable profit-and-loss impact, with only about 5% capturing real value. McKinsey's 'State of AI in 2025' echoes the tension: a large majority of organisations now use AI somewhere, yet only a small share report material enterprise-wide earnings impact so far.

Both findings come with caveats — the NANDA sample is self-reported and not peer-reviewed, and early productivity gains often show up in speed and quality before they hit the P&L. Controlled studies do find real per-task productivity improvements for coding, writing and support. The open question is whether those gains scale into durable, measurable profit fast enough to justify the valuations and capex above.

What is 'circular financing' and why does it worry critics?

'Circular financing' describes deals where the same money appears to loop between suppliers, customers and investors — and in 2025-26 a web of such deals linked Nvidia, OpenAI, Oracle and CoreWeave. The concern is that revenue, funding and backlog can be different accounting views of the same dollars, inflating the apparent health of the whole chain.

The structure is roughly this: Nvidia invests in and supplies OpenAI; OpenAI commits hundreds of billions to cloud providers such as Oracle (a reported $300 billion, five-year deal); those providers buy Nvidia GPUs to build the capacity; and neoclouds such as CoreWeave sign multi-billion-dollar agreements financed against the same demand. A headline $100 billion Nvidia-OpenAI commitment from September 2025 was restructured in early 2026 into an equity stake — illustrating how fluid these arrangements are.

Critics compare this to dot-com-era vendor financing, where equipment makers lent customers the money to buy their gear, flattering revenue until the customers failed. Defenders argue these are strategic investments and genuine supply contracts between solvent, cash-rich companies, not hollow round-tripping. The data does not settle it; the interlock is real, its fragility is disputed.

How does this compare to the dot-com bubble?

The dot-com comparison is the most common historical frame, and the data shows real rhymes and real differences. What rhymes: extreme valuations, a capital-spending frenzy (telecom fibre then, GPUs and data centres now), circular/vendor financing, and a narrative that 'this changes everything.' What differs most is profitability and revenue.

At the 2000 peak the Nasdaq-100 forward P/E reached roughly 60× and only a minority of listed dot-coms were profitable; many had trivial revenue. In 2026 the Nasdaq-100 trades near 30× and the megacap AI leaders are among the most profitable companies in history, funding much of the build-out from operating cash flow rather than IPO proceeds. The revenue is also real and large — ChatGPT's 900 million weekly users versus pets.com's sock puppet.

The bear counter is that the epicentre has moved to private markets and infrastructure, where valuations are opaque and GPUs depreciate quickly — so a correction could hit private labs, neoclouds and capex-heavy balance sheets even if the profitable megacaps hold up. The table contrasts the two eras on the metrics that matter.

What is the bull case that this is not (just) a bubble?

The bull case rests on real, fast-growing revenue and genuine utility, not hype. ChatGPT's 900 million weekly active users (OpenAI, February 2026) and the labs' revenue run-rates — OpenAI ~$25 billion, Anthropic ~$47 billion by mid-2026 — represent adoption faster than any prior software category. Unlike 2000, the demand is paid and recurring, and the leading buyers of AI infrastructure are cash-generative megacaps funding it largely from profits.

Proponents also argue the productivity gains are early but real: controlled studies show meaningful per-task improvements in software development, customer support and writing, and Stanford HAI's AI Index documents rapid gains in model capability and steep falls in inference cost. Jeff Bezos framed it as 'a kind of industrial bubble' whose eventual benefits to society would be 'gigantic' — bubbles can misprice assets while the underlying technology still transforms the economy.

The strongest bull point is time: infrastructure booms routinely spend ahead of revenue, and if AI revenue compounds toward Bain's $2 trillion marker even partially, today's multiples normalise. In this reading, some capital is wasted and some valuations correct, but the category is durable — closer to the internet's long arc than to a pure mania.

What is the bear case that this is a bubble?

The bear case is that valuations and capex have detached from demonstrable returns. On the numbers above, several labs trade at hundreds of times revenue, the hyperscalers are spending $700-900 billion a year against a revenue base an order of magnitude smaller, and Bain warns of a possible $800 billion revenue shortfall by 2030. The MIT NANDA finding — ~95% of enterprise pilots with no measurable P&L impact — is the bears' proof that adoption is not yet profit.

Investor Michael Burry, of 'Big Short' fame, disclosed large put positions against Nvidia and Palantir in late 2025, citing dot-com-scale price-to-sales ratios and, crucially, GPU depreciation — the argument that hyperscalers may be understating the cost of chips that lose value in a few years. The 'circular financing' web adds fragility: if one large buyer falters, the interlocked revenue could unwind quickly.

Even Sam Altman, whose company sits at the centre of the boom, said in 2025 that investors are 'overexcited' and that 'when bubbles happen, smart people get overexcited about a kernel of truth,' calling some startup valuations 'insane.' The bear reading is not that AI is fake — it is that the price of AI exposure has run far ahead of the cash it currently produces.

What would signal a bubble bursting versus a boom sustaining?

Rather than predict, it is more useful to name the observable signals each side is watching — readers can track these themselves. Signals that would support the bear case: AI revenue growth decelerating while capex keeps rising (a widening gap); hyperscalers writing down GPU or data-centre assets; a marquee lab or neocloud struggling to raise at a flat or lower valuation ('down round'); enterprise ROI staying elusive in successive McKinsey/MIT-style surveys; or one node of the circular-financing web defaulting.

Signals that would support the bull case: the capex-revenue gap narrowing as enterprise AI revenue compounds; a rising share of companies reporting measurable earnings impact, not just usage; inference costs continuing to fall (improving unit economics); and durable, profitable AI products beyond chat — agents, coding and vertical tools — showing retention.

The genuinely uncertain variables are timing and concentration. Even most bears expect AI to matter enormously long-term; the disagreement is whether the near-term financing can bridge to that future without a sharp repricing. Because so much of the risk now sits in private valuations and capex-heavy balance sheets, any correction may look different from a public-market crash. All forward-looking statements here are scenarios, not forecasts.

What does this mean for you? (No advice)

This report presents public data and the arguments named sources draw from it; it does not tell you what to do with that data. The facts that are well-established: AI usage and lab revenue are growing exceptionally fast; valuations and capital spending are at historically extreme levels relative to current revenue; and enterprise return on investment is, so far, hard to measure at scale. Those three facts are simultaneously true, which is exactly why credible people reach opposite conclusions.

'Bubble' is ultimately a judgement about the future, and the future is uncertain by definition. Both the bull case (real demand, profitable leaders, infrastructure logic) and the bear case (stretched multiples, capex-revenue gap, unproven ROI, circular financing) are grounded in the same public record laid out above. Reasonable readers, weighing the same evidence, will land differently.

Affärslivet's role is to show the numbers, attribute every one to a named and dated source, and let you decide. We take no position on whether AI is a bubble, and nothing here is a recommendation to buy, sell or hold any asset. This is not financial advice.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

Every figure is attributed inline to a named, dated source. Valuation-to-revenue multiples are Affärslivet calculations that pair each lab's most recently reported private-round valuation with its annualised revenue run-rate as tracked by Epoch AI and reported by primary sources; they are order-of-magnitude estimates, not liquid market prices, and are marked Medium confidence. Multiples: OpenAI ~$852bn/~$25bn≈34×; Anthropic ~$965bn/~$47bn≈21×; xAI ~$230bn/~$0.4bn≈537×; Mistral ~$14bn/~$0.4bn≈34×; Z.ai/Zhipu ~$39bn/~$0.2bn≈258×. Hyperscaler capex ($700-900bn, 2026) is an analyst-estimate range across Amazon, Microsoft, Alphabet, Meta and Oracle. The $2 trillion revenue requirement and ~$800bn shortfall are from Bain & Company's 2025 Technology Report. The 95% no-ROI figure is from MIT Project NANDA's 2025 study (self-reported, not peer-reviewed; sample of 300+ initiatives, 52 interviews, 153 survey responses). ChatGPT's 900 million weekly active users is OpenAI's February 2026 figure. Nasdaq-100 P/E comparisons are market data. Quotes from Altman and Bezos are verbatim from 2025 interviews as reported. No figures were invented; any that could not be tied to a named source were omitted or labelled as estimates. This report is factual analysis, not investment advice.

Data dictionary

Frequently asked questions

Glossary

Valuation-to-revenue multiple

A company's valuation divided by its annual revenue; a quick gauge of how much of the price rests on future rather than current sales. Mature software typically trades at 5-15×. ↗

Capex (capital expenditure)

Money spent on long-lived assets such as data centres and chips. In AI, hyperscaler capex reached an estimated $700-900bn in 2026, largely for AI infrastructure. ↗

Circular financing

Interlocking deals in which the same capital appears to loop between a supplier, its customers and its investors, potentially flattering revenue across the chain. ↗

GPU depreciation

The decline in value of AI accelerator chips over their useful life (a few years), a key bear argument that hyperscalers may understate the true cost of their AI build-out. ↗

Sources

• Epoch AI — AI company revenue data
• Stanford HAI — 2026 AI Index Report
• Bain & Company — Technology Report 2025 ($2tn revenue / $800bn shortfall)
• MIT Project NANDA — State of AI in Business 2025 (95% no-ROI)
• McKinsey — The State of AI in 2025
• CNBC — Guide to $1 trillion of AI deals (circular financing)

Related reports & data

• The Largest AI Companies
• How Much It Costs to Train an AI Model
• National AI Power Index