AI can be a real technological revolution—and still contain a financial bubble

Updated 26 September 2026. All financial figures are in US dollars.

An AI service can be useful, popular and expensive to run. Its customers can get good value while its investors lose money. That combination is easy to miss when the debate becomes a choice between “AI is a bubble” and “AI will change everything.”

The useful questions are more specific. Does serving a customer make money? Does that surplus cover the cost of developing the next model? Will the data centres earn enough to repay their financing? And how much has an investor paid for a share of the eventual cash flow?

The answers already differ across companies. Anthropic reportedly achieved positive adjusted operating income in the second quarter of 2026, while OpenAI reported a substantial operating loss. Those measures also leave out much of what we need to know about the return to shareholders. ^[1]

Private labs publish far less financial detail than listed companies. Much of the evidence below comes from investor documents reported by journalists. Where the accounts do not provide a cost breakdown, I have left the gap visible. Forecasts and hypothetical examples are labelled separately from reported results.

What profit leaves out

Suppose an AI company receives $100 from customers and spends $20 running their requests. The remaining $80 still has to pay for research, training, product development, sales and administration, plus any interest and taxes. The word “profit” only helps once we know which of those costs have been deducted.

Term	What it actually tells you
Revenue	Sales earned from customers. Money raised from investors is not revenue.
Gross profit	Revenue minus the direct costs classified as delivering the product. It does not normally subtract all research and corporate expenses.
Operating profit	What remains after the business’s operating costs, including research and development.
Net profit	The bottom line after additional items such as interest and taxes.
Free cash flow, or FCF	Broadly, cash generated by operations minus capital spending. It addresses whether the business is generating cash after investing in physical assets.

Cash flow matters because equipment is usually paid for before its full cost appears in the income statement. Depreciation spreads that cost across the asset’s accounting life. A company can therefore report a profit while spending more cash than it receives. ^[2]

“Adjusted profit” excludes specified expenses. Read the exclusions: removing stock-based compensation can change the result considerably. Paying staff in shares saves cash now but dilutes existing shareholders. EBITDA, meanwhile, excludes interest, taxes, depreciation and amortisation. Ignoring depreciation is a large concession to a business whose expensive computing equipment will need replacing.

Annualised revenue is another common source of confusion. A company earning $1 billion in a month can describe a $12 billion annualised run rate. It has not earned $12 billion, and its customers have not necessarily committed to spending that much.

What the labs’ accounts tell us

The table separates AI labs from the larger businesses that own them. Google’s advertising income is not Gemini revenue. Meta’s overall profit does not tell us whether selling access to its models makes money. Here, “operating costs” means all costs deducted to reach operating income, rather than the narrower line sometimes called operating expenses. The public-company quarterly financial statements cited here are unaudited; private-lab figures are reported rather than fully disclosed accounts.

Business and period	Revenue	Operating costs or result	What we can conclude
OpenAI, Q2 2026	$6.7bn	Reported operating loss $12.3bn, including stock-based compensation. Implied total operating costs: approximately $19.0bn.	Loss-making on this reported operating measure. These are privately reported figures. ^[3]
Anthropic, Q2 2026	About $11.5bn	Reported positive adjusted operating income. A complete cost reconciliation is not public.	An important profitability milestone, but not proof of positive net income or free cash flow. ^[1]
Alphabet, Q2 2026	$119.80bn	Operating costs approximately $79.03bn; operating profit $40.77bn.	The parent company is highly profitable. Standalone Gemini profitability is not disclosed. ^[4]
Meta, Q2 2026	$60.80bn	Operating costs $42.03bn; operating profit $18.78bn.	The parent company is profitable. Its frontier-model business does not have a separately disclosed profit-and-loss account. ^[5]
SpaceX’s AI segment, including xAI, Q2 2026	$2.56bn	Reported operating loss $1.3bn, but positive adjusted EBITDA.	Operating-loss-making, despite a positive adjusted measure. This segment also earns revenue from renting computing capacity; it is not simply Grok subscriptions. ^[6]

OpenAI’s implied $19 billion of Q2 operating costs is revenue of $6.7 billion plus its reported $12.3 billion operating loss. It is not a cash-burn estimate. Separately, Reuters reported $3.7 billion of cash burn against $5.7 billion of revenue in Q1. Cash burn is the cash deficit after inflows, not total spending. ^[7]

Anthropic’s improvement needs an accounting footnote

September reporting said Anthropic expected another positive adjusted operating result in Q3, following its preliminary Q2 result. The quarter had not ended by 26 September. A complete public reconciliation of the adjusted measure was not available. ^[8]

Its reported gross margin also exceeded 80% before distribution revenue-sharing and model-training costs. That is a different measure from adjusted operating income. It should neither be called an 80% net margin nor used to assume that the operating-income calculation also excludes training. The published information does not provide a complete reconciliation from revenue to shareholder earnings and cash generation. ^[8]

Google’s profit and its cash spending

Google Cloud reported $24.77 billion of Q2 revenue and $8.81 billion of operating income, a margin of about 35.6%. Cloud sells plenty besides AI, and Alphabet reports some shared costs, including AI research, outside its individual operating segments. These figures cannot establish Gemini’s standalone profitability. ^[4]

Alphabet generated $39.069 billion of operating cash flow in Q2 and spent $44.924 billion on capital expenditure, leaving negative $5.855 billion of quarterly free cash flow, or about negative $5.86 billion. Trailing-twelve-month free cash flow was positive at $53.27 billion. A profitable business can have a cash-consuming quarter while building infrastructure. ^[9]

Break-even dates depend on the measure

A January report put Anthropic’s cash-flow-positive target in 2028. Its later adjusted operating profit does not demonstrate that it has reached that cash-flow target; nor should an earlier forecast be treated as unchanged after rapid growth. ^[10]

OpenAI’s September projections, discussed below, show substantial cumulative cash consumption through 2030. A cumulative deficit does not tell us which individual years are negative or give us a reliable first year of positive net income. ^[11]

Alphabet and Meta are profitable parent companies, but their reviewed disclosures give no standalone frontier-AI break-even dates. The reporting on xAI’s broader AI segment likewise does not establish a dependable standalone cash-flow break-even date. Positive adjusted EBITDA is not a substitute.

Training, serving customers and paying for the machines

Training costs extend beyond the successful run

Training changes a model’s internal parameters. A research programme also pays for data, evaluations, researchers and experiments that do not work. The successful training run is only part of that bill.

DeepSeek’s original V3 report estimated $5.576 million for its reported training process, using an assumed GPU rental price. It explicitly excluded earlier research and experimentation. The number is evidence of an efficient training process, not the cost of building the entire company. ^[12]

Inference costs depend on the task

Inference is running a trained model to answer a question, write code or choose an action. Each request can use substantial computing resources. The model, hardware, memory requirements and serving setup all affect the bill; processing a long input and generating a long answer impose different workloads. ^[12]

Batching lets hardware process several requests together, with possible trade-offs in response time. Caching avoids repeating work on material that has already been processed. These techniques help explain why two apparently similar requests can have different costs. ^[13] ^[34]

A token is a unit of text, such as a word, part of a word or punctuation. API providers usually price input and output separately. Anthropic lists Claude Opus 5.5, released on 22 September, at $4 per million input tokens and $20 per million output tokens. Without caching or extra tool charges, that produces these customer bills: ^[14]

Illustrative workload	Calculation	Customer’s API bill
10,000 input tokens and 2,000 output tokens	$0.04 input + $0.04 output	$0.08
500,000 input tokens and 100,000 output tokens	$2 input + $2 output	$4.00

Those prices do not reveal Anthropic’s underlying compute cost or profit margin. An agent can also make many calls, inspect the results and try again before completing one task. For a business buying the service, the relevant cost includes all those attempts and the human checking that follows.

The infrastructure bill has to land somewhere

A lab can buy machines, rent them or sign a capacity contract. Renting changes who finances the equipment; the supplier still needs to recover its costs. In a simplified example, a $1 billion cluster depreciated over five years with no residual value creates $200 million of annual depreciation, despite the upfront cash purchase.

Do not add the lab’s rental payments, its supplier’s equipment purchases and the full value of a multi-year capacity contract and call the sum this year’s AI losses. They cover different transactions and accounting periods.

What a five-times compute markup means

In an August interview, SemiAnalysis founder Dylan Patel estimated compute costs of roughly $10 million to $15 million per megawatt, alongside inference revenue reaching about $50 million per megawatt for Anthropic. These are his estimates, not audited industry margins. ^[15]

A megawatt measures power; a gigawatt is 1,000 megawatts. Data-centre operators use power capacity as shorthand for the equipment they can run, but it does not specify the hardware, utilisation or selling price. The time period matters too: an equipment purchase and a year’s revenue are not directly comparable. In the examples that follow, assume costs and revenue cover the same year.

Compute costs include more than electricity

One megawatt running continuously for a non-leap year consumes 8,760 megawatt-hours. At an illustrative electricity price of $70 per MWh, the bill is $613,200 before cooling and other supporting loads.

1 MW × 8,760 hours × $70/MWh = $613,200

An annual compute-capacity cost of $10 million to $15 million would therefore include much more than electricity: machines and supporting infrastructure, whether bought directly or paid for through a rental.

Margin and markup use different denominators

With revenue of $50 and serving costs of $10, the serving margin is 80%, while the markup on cost is 400%. Revenue is five times cost. If the cost rises to $15, the margin falls to 70%.

Margin: (50 − 10) ÷ 50 = 80%
Markup: (50 − 10) ÷ 10 = 400%

Here is a hypothetical annual business, with all amounts in millions:

Item	Amount
Customer revenue	$50
Fully loaded inference-serving costs	−$10
Distribution and partner payments	−$5
Training and research	−$30
Other operating costs	−$8
Operating result	−$3

It sells inference for five times the serving cost and still loses $3 million after the other expenses. Attractive inference economics can fund model development without covering the whole research budget.

Even that example is cleaner than the real business. A margin on paid API requests may not describe free users, subscriptions and partner channels. Competition can push selling prices down while scarce capacity gets more expensive, a risk Patel also discussed. ^[15]

A lab could cut training spending, but might then lose the product quality that attracts its customers. Some research opens new markets; some is the cost of staying competitive. The public accounts do not separate those neatly.

How much scale would cover the costs?

A larger customer base can spread development costs across more sales. The benefit depends on how much revenue remains after inference and distribution costs, and how fast the research bill grows.

Operating profit = Revenue × contribution margin − research and other fixed costs

In this simplified model, “fixed” means relatively fixed during the period being analysed. If research and other fixed operating costs total $30 billion a year, break-even revenue looks like this:

Contribution margin	Revenue needed to cover $30bn
40%	$75bn
60%	$50bn
80%	$37.5bn

The model explains why a lab can become profitable without making training cheap. It also explains why fast growth may not be enough: if research spending doubles along the way, the break-even point moves.

OpenAI’s spending plan still needs financing

Reuters’ September report, based on a company presentation reviewed by the FT, described the following forecasts. None is an actual future result. ^[11]

OpenAI forecast	Amount
2026 revenue	$36bn
2030 revenue	$350bn
Cumulative 2026–2030 revenue	$840bn
Computing and infrastructure spending over the period	$856bn
Cumulative negative free cash flow	$278bn

Moving from $36 billion of revenue in 2026 to $350 billion in 2030 requires about 76.6% annual growth over four intervals. The projected $856 billion of computing and infrastructure spending is about 102% of the $840 billion cumulative revenue. That spending category includes more than inference and should not be read as one year’s operating expense.

Even if those sales arrive, the plan consumes cash. Existing reserves can fund some of the projected $278 billion deficit, so that is not automatically the amount still to be raised. The same reporting suggested then-available cash would run out in 2028. New funding or a change in the spending plan can move that date; it is not a bankruptcy deadline. ^[11]

A lab may serve additional customers profitably while spending the resulting surplus, and much more, on research and future capacity. Customer economics and the affordability of the expansion plan need separate tests.

What businesses are buying, and what they get back

Microsoft reported more than 30 million paid Microsoft 365 Copilot seats in its July 2026 results. That is evidence of demand. Paid seats alone do not tell us how actively staff use the product or what their employers earn back. ^[16]

Eurostat reported that about 20% of EU enterprises with at least ten employees used some form of AI in 2025; among large enterprises, the figure was about 55%. This measure covers AI broadly, not just paid generative-AI tools. Adoption is substantial but uneven. ^[17]

Productivity gains vary with the job

A field study of roughly 5,000 customer-support agents found about a 15% average increase in issues resolved per hour, with different benefits for workers at different experience levels. ^[18]

METR’s early-2025 randomised study reached a different result in a different setting. Experienced developers working on familiar open-source repositories took 19% longer with the AI tools available to them. ^[19]

METR’s February 2026 update warned that participation and task-selection biases made newer measurements hard to interpret. The earlier slowdown cannot be treated as a settled result for current tools, just as a successful demonstration cannot establish a company-wide productivity gain. ^[20]

Turning time saved into something useful

Consider an illustrative business with 100 employees. If each saves one hour a week over 48 weeks, at a fully loaded labour cost of $60 an hour, the potential annual capacity value is $288,000.

100 employees × 48 hours saved × $60 = $288,000

Now suppose the first year costs:

Illustrative cost	Amount
Subscriptions at an assumed $30 per employee per month	$36,000
Integration and implementation	$100,000
Training, governance and review	$60,000
Total	$196,000

If all the freed time becomes useful output or avoided expenditure, the net benefit is $92,000. If only 30% does, the realised value is $86,400, less than the $196,000 cost. Faster service, better quality and additional sales can create value without anyone losing a job, but the business has to make that conversion happen.

An invoice-processing agent, for example, must do more than read an invoice. It needs to match records, respect approval limits, handle exceptions and leave an audit trail. Cheaper tokens do not solve those implementation problems. They are closely related to the work described in taking a predictive model from demo to production.

Who keeps the customer relationship?

The labs are moving beyond model access. OpenAI describes an enterprise-and-agent expansion strategy, while Anthropic markets products such as Claude Code and Cowork. They want a larger role in the work their models perform. ^[21] ^[35]

If an application vendor charges a customer $100 and pays a model provider $10, the provider has an incentive to build the application too. The vendor, in turn, can try to keep the customer relationship while switching among competing models. I expect a mixed market: neither side has an automatic claim to all the profit.

An enterprise system is more than its interface

A system of record holds a company’s authoritative customer, employee or financial data. An AI interface may make it easier to use without replacing the permissions, integrations and compliance processes underneath.

Reliability across a whole workflow matters. A hypothetical agent completing ten independent steps, each with a 95% chance of success, succeeds at all ten only about 60% of the time. Real failures need not be independent, but the example shows why a good score on individual actions can hide a poor completion rate.

0.95¹⁰ ≈ 60%

The companies with a defensible position may control a superior model, a trusted workflow, distribution, access to important data or low-cost infrastructure. Labs can expand into applications; established software companies can offer several models; specialists can make a narrow process work better than either.

Ramp said customers using its model router reduced inference costs by 40% on average by matching requests to cheaper models that met a quality bar. This is Ramp’s own reported customer result, not an independent industry estimate or a reduction in total AI spending. It illustrates how buyers can take some of the savings even as their usage grows. ^[27]

The total wage bill of knowledge workers is not automatically the labs’ future revenue. Customers can keep much of the value AI creates, especially when they can switch suppliers.

Where the investment case could succeed or fail

The case for growth

Lower prices make more tasks worth automating. Stanford’s 2025 AI Index reported a more than 280-fold fall in inference cost at roughly GPT-3.5-level capability between November 2022 and October 2024. That holds capability roughly constant; it does not describe every frontier workload. ^[22]

Epoch AI’s newer study estimates that the price of a fixed performance level fell about 13-fold per year since 2023 across five selected benchmarks in maths, science and games. It follows the cheapest available way to reach benchmark performance, assuming users switch to suitable models. It does not measure each lab’s production cost or productivity on real business tasks, and Epoch cautions that its data are incomplete and the estimate is not exact. ^[28]

A task that was not worth automating at $20 may be worthwhile at $0.20. Enough new customers could make continuing research affordable. Companies with existing profitable products also have ways to benefit from AI beyond selling chatbot subscriptions, including improving those products.

The commitments can outlast the demand

The financial risk is that customers can cut usage or switch providers faster than suppliers can escape their infrastructure obligations. A busy data centre can still earn less than its owners expected when they financed it.

Prices might fall faster than costs. A lab might need to fund every new model just to retain customers. New hardware might make an older cluster less competitive while it still works perfectly well. In all those cases, users can receive better services while suppliers earn disappointing returns.

More usage does not settle the revenue question. An 80% price cut combined with ten times as much usage doubles revenue. A 90% price cut combined with five times as much usage halves it. The size of the demand response and the margin on that demand decide the outcome.

The financing links between labs and suppliers

The concentration of that exposure is examined in Steve Eisman on who is funding the AI boom.

Anthropic announced a commitment to spend more than $100 billion on AWS over ten years, alongside an additional $5 billion Amazon investment immediately and potentially $20 billion later. A supplier investing in a customer can be a sensible commercial arrangement, but it also ties their financial fortunes together. ^[23]

OpenAI and SoftBank each announced a $500 million investment in SB Energy in January 2026, alongside a planned 1.2-gigawatt Texas data-centre lease for OpenAI. Nvidia then announced a separate $1.5 billion investment in SB Energy and an Ohio campus that SB Energy would build and own, with OpenAI as tenant for 20 years and Nvidia providing the compute infrastructure. Equity investments, capacity and leases are different measures; adding them does not produce an estimate of final-customer demand. ^[30] ^[31]

Nvidia’s filing caps its initial residual-value guarantees for Ohio leases at $105 billion, covering about 4.25 gigawatts of IT load. Payment depends on leases starting, readiness conditions being met and a shortfall following an OpenAI default. The cap is a conditional maximum exposure, not money already paid. ^[32]

SoftBank Group announced terms for roughly $11.1 billion of bonds on 24 September, partly to fund its final $10 billion OpenAI investment tranche. The 9.75% coupon applies to the $4.5 billion 2034 dollar tranche, not the whole issue. Issuance was expected on 29 September, after this article’s review date. ^[33]

Follow the money without counting it twice

Suppose a business pays an AI application company $100, the application company pays a lab $30, and the lab pays a cloud provider $20. Their revenues add to $150, but final-customer spending was $100. All three transactions can be legitimate; treating them as independent end-market demand is the error.

The same care is needed with a cloud provider’s contracted revenue and a lab’s capacity commitments. A contract can cover years of future activity. It is not cash already collected or spent.

How the risks can spread

Residual-value guarantees can transfer part of the risk of disappointing asset values to a guarantor. Nvidia’s Ohio arrangement above is a concrete example: a financing structure that links the lab, infrastructure owner and equipment supplier.

If customer revenue disappoints, labs may need more financing just as investors become less willing to provide it. Capacity commitments can then become a burden, and suppliers face renegotiations or defaults. That is a possible route to restructurings and investor losses even while AI use keeps growing. The guarantees themselves do not prove that it will happen.

Paying back the investment

Covering this year’s costs is a smaller achievement than repaying all the capital invested, with a return that compensates investors for the risk and the wait.

Suppose investors put $100 billion into a business that eventually generates $10 billion of annual free cash flow after necessary capital expenditure. Ignoring discounting, simple payback takes ten years after those cash flows begin. An individual shareholder can sell earlier for a gain, but that sale does not establish the ultimate return from operating the business.

What a trillion-dollar price requires

Anthropic’s May funding announcement put its post-money valuation at $965 billion. In August, Bloomberg reported an annualised revenue run rate of approximately $65 billion, as relayed by Fortune. That run rate extrapolates from recent sales; it is not an audited $65 billion of annual revenue. ^[25] ^[29]

A simplified valuation gives a sense of the scale required. For a mature business with steadily growing free cash flow:

Value = next year’s FCF ÷ (required return − growth)

At a $1 trillion value, a 10% required return and 3% perpetual growth, this model needs $70 billion of next-year free cash flow. A 20% free-cash-flow margin would imply $350 billion in annual revenue.

$1tn × (10% − 3%) = $70bn of next-year FCF

A fast-growing young company needs a fuller model with several years of forecasts. Debt and equity must also be handled consistently with the cash flow being valued. The example is not a price target; it shows how much more a trillion-dollar price demands than simply reaching profitability. Delays, further capital requirements and shareholder dilution make the return harder to achieve.

What a public listing would test

Patrick Boyle points out that private funding valuations have yet to face the same price discovery as public shares. Private rounds involve selected investors, sometimes suppliers with strategic interests in the lab. A listing brings a wider set of buyers and sellers and more standardised financial disclosure. That makes assumptions easier to challenge, without guaranteeing a fair price. ^[26]

Higher required returns reduce the present value of future cash flows. Higher borrowing costs can also make data-centre construction more expensive. These effects matter particularly when the earnings needed to justify today’s price lie years away.

A large total addressable market cannot fill that gap on its own. An investor still needs a plausible path from economic opportunity to customer spending, the company’s share of it, margins and cash left after investment. Forecasts of rapidly improving AI, discussed in the recursive self-improvement article, do not remove those steps.

What would change my view

AI has paying customers and evidence of productivity gains in specific jobs. Its usefulness leaves several investment questions unanswered: whether each lab can fund continuing research, whether each data centre earns its cost of capital, and whether a buyer of shares has paid too much.

I expect the outcomes to differ across companies. Some may flourish while others consolidate or restructure. Customers may continue getting better services through all of it. The optimistic case is that demand grows fast enough and serving economics improve enough to make research and infrastructure renewal affordable. That needs to show up in the cash flows.

I would be more confident if labs produced sustained cash after necessary infrastructure spending, improved margins after distribution costs, retained customers and let research spending grow more slowly than contribution profit. I would be more concerned if growth relied increasingly on discounts, capacity commitments ran ahead of credible demand, or cash-flow forecasts kept moving out.

When someone argues that a lab is profitable, I want to know which expenses the measure excludes. When they argue that its valuation is justified, I want to see the future cash flows that support the price. Useful technology can survive a bad investment cycle. Investors still have to pay a price the business can earn back.

Sources and further reading

Citation numbers are retained from earlier versions; a removed reference can leave a gap. Additional sources appear at the end. Some reporting requires a subscription.

Financial disclosures and forecasts are dated. Company-reported savings and private financial results are attributed as such; illustrative calculations are not reported company results.