TECHNOLOGY & CAPITAL / A FINANCIAL PERSPECTIVE

Is AI a bubble?
Follow the cash,
not the hype.

The technology can be useful and the investment cycle can still overheat. My framework for separating capability, commercial returns and market expectations.

01 / THE QUESTION

A real technology can still attract unrealistic prices.

I would separate three questions that are often collapsed into one: does AI work, can a business earn attractive returns from it, and does its valuation already assume too much success?

The first question cannot settle the other two. A model can save time while its provider struggles to capture enough of that value. A profitable provider can still be a disappointing investment if the purchase price assumes extraordinary growth.

TechnologyUseful capability

Test it against real tasks.

Business modelReturns need proof

Follow costs and retained revenue.

Market priceExpectations matter

Ask what is already priced in.

My working view is that AI can be useful while parts of its investment cycle become overheated. That is a thesis to test, rather than a claim that every AI business is a bubble. Fraud is a separate allegation requiring evidence of deception; large spending or connected commercial relationships do not establish it.

02 / THE ECONOMICS

Count successful outcomes, not just tokens.

Traditional software can distribute another copy at very low cost, although hosting, support and customer acquisition still cost money. Generative AI adds inference: computing required each time a model runs. More use can therefore create both more revenue and more operating expense.

But the claim that AI margins must deteriorate with scale is too strong. Better hardware utilisation, smaller models and efficient serving can reduce cost. The question is whether these gains outweigh price cuts, longer workloads and the cost of checking unreliable outputs.

MICROSOFT / FY2025

Microsoft reported a 69% Microsoft Cloud gross margin, with AI infrastructure expansion putting pressure on it and Azure efficiencies partly offsetting that pressure. This is a broad cloud measure, not a standalone AI margin.

Microsoft FY2025 annual report ↗

A token is a unit of model usage, not a unit of business success. Usage billing is not inherently flawed; it simply leaves the buyer responsible for measuring whether the output was worth the full cost.

What changes when the product has to think on every request?
Economic questionTraditional softwareGenerative AI service
Another useDistribution can be cheap; hosting and support remain.Model execution adds compute and token costs.
Quality controlTest predefined behaviour and software defects.Also evaluate variable outputs, retries and human review.
Scaling returnsDepend on pricing, retention and operating costs.Also depend on inference efficiency, workload and acceptance rates.
Useful measureRevenue and contribution per customer.Also cost and contribution per accepted outcome.

Try the economics.

Illustrative example / INR per task

Assume ₹100 revenue per accepted task, ₹30 for review and ₹15 for integration and support. Adjust inference cost and the share of attempts accepted.

Inference per accepted task₹25.00
Contribution per accepted task₹30.00

Where the ₹100 goes

Revenue to contribution, per accepted task. Each cost steps down from the previous balance.

Revenue₹100.00
Less: inference−₹25.00
Less: review−₹30.00
Less: integration & support−₹15.00
Remaining contribution₹30.00

₹100 revenue less ₹70 in modelled costs leaves ₹30 contribution.

How much room is left as inference gets dearer?

Contribution per accepted task at different inference costs. Revenue, review and support stay fixed.

100% accepted80% accepted
Inference cost versus contributionTwo lines compare 100 percent acceptance with the selected acceptance rate. The highlighted point shows the selected cost.₹50₹0−₹70₹10₹35₹60Inference cost per attempt

At 80% acceptance, contribution reaches zero at ₹44.00 inference cost per attempt.

Inference cost per accepted task = cost per attempt ÷ acceptance rate. Contribution = ₹100 − adjusted inference − ₹30 − ₹15. This simplified model excludes fixed R&D, taxes and financing; contribution is not net profit. Review and support are held fixed for illustration.

03 / COMPETITION

Cheaper intelligence is good for users. Pricing power is another question.

Stanford's 2025 AI Index reported that the inference cost of reaching GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. That is a historical, capability-specific comparison, not a claim that every model or task became 280 times cheaper.

Stanford AI Index 2025 ↗
Nov 2022
100
Oct 2024
<0.36

Earlier cost indexed to 100. Equivalent GPT-3.5-level performance; derived from Stanford's reported decline. Not current API pricing.

Open-weight competition strengthens the buyer's alternatives. DeepSeek's R1 release includes smaller models distilled using R1-generated data. Distillation transfers behaviours from a larger model into another model; it does not justify treating all Chinese AI as copied US research.

DeepSeek R1 model documentation ↗

My inference: efficiency can expand demand while also compressing supplier pricing. The winner depends on which effect dominates. A single cheap coding run cannot establish a durable advantage: model version, retries, accuracy, latency and operating costs all matter. I would compare total cost per accepted task on the same workload.

04 / THE CAPITAL CYCLE

The risk sits in who pays, who borrows and who depends on whom.

AI infrastructure brings together chip suppliers, cloud operators, model developers, lenders and investors. When the same organisations appear in several roles, headline spending may overstate the diversity of independent end demand.

01Capital

Equity and borrowing

02Capacity

Chips, power and facilities

03Usage

Cloud and model contracts

04Cash return

Collections after costs

COREWEAVE / 2024 REVENUE

CoreWeave's 2025 IPO prospectus reported that Microsoft represented 62% of its 2024 revenue. That dated concentration illustrates why customer creditworthiness, renewal and contract terms matter alongside growth.

CoreWeave IPO prospectus, SEC ↗

One customer. Most of the revenue.

CoreWeave revenue mix, 2024 — as reported in its 2025 IPO prospectus.

62%Microsoft
38%All other customers

The remaining 38% is calculated as 100% less 62%. This is a historical mix, not a current concentration estimate. Source: CoreWeave prospectus ↗

Supplier financing or an investor becoming a customer is not automatically artificial revenue. The useful questions are whether contracts have commercial substance, payments are collected, exposure is concentrated and demand survives without another financing round.

I would distinguish announced commitments, contracted backlog, recognised revenue and cash receipts. They are different stages of the same economic story. I would also test whether equipment remains productive for the period assumed in depreciation and whether debt can be serviced under lower utilisation.

05 / TRUST & CONTROL

Enterprise adoption is a governance decision too.

A low model price does not settle where sensitive data should go. Procurement must consider retention, access permissions, training terms, auditability, service continuity and the cost of switching providers.

It would be inaccurate to assume every enterprise prompt trains a vendor's next model. Anthropic states that commercial inputs and outputs are not used for training by default, while describing exceptions such as explicit opt-in and submitted feedback. Product and contract terms matter.

Anthropic commercial data policy ↗

Local or sovereign infrastructure can change the allocation of control and spending. The European Commission's February 2025 InvestAI announcement sought to mobilise €200 billion, including a €20 billion fund for AI gigafactories. An announced funding initiative is not completed infrastructure.

European Commission, February 2025 ↗

My interpretation is that sovereignty can redirect infrastructure demand rather than eliminate it. Self-hosting also introduces operating costs and responsibilities; it is not automatically the cheaper or safer choice.

06 / WHAT I WOULD WATCH

A watchlist is more useful than a crash date.

What would strengthen or weaken the bubble argument
SignalWarning signsEvidence against a bubble
Customer economicsPilots grow but paid renewals stall.Customers renew after measuring savings and quality.
Capacity utilisationInstalled capacity outpaces paying demand.New capacity fills at sustainable pricing.
Cash conversionBacklog rises while collections lag.Contracted demand becomes operating cash.
FinancingRefinancing costs rise and funding options narrow.Debt service remains covered under stress.
Competitive pressurePrices fall faster than unit costs.Efficiency and usage gains preserve returns.

A capex reduction could signal weaker demand or sensible discipline. Wider credit spreads could reflect sector problems or a broader shock. Neither is a standalone countdown to a collapse; interpretation requires context and a consistent time series.

07 / WHAT WOULD CHANGE MY MIND

The optimistic case deserves a fair test.

The strongest counterargument is that cheaper AI unlocks enough new demand to absorb infrastructure, while better utilisation and workflow integration improve returns. Companies may also capture value through an existing distribution network rather than a standalone chatbot subscription.

I would become less concerned if repeat customers demonstrated measurable benefits, unit costs fell without sacrificing reliability, cash generation followed revenue and capacity additions remained supported by diversified demand.

Conversely, repeated funding rounds used to sustain uneconomic usage, weakening renewals and persistent gaps between commitments and cash would strengthen the overheating thesis. This makes the argument falsifiable: there is evidence that can change it.

08 / CONCLUSION

Be optimistic about capability. Be demanding about returns.

I do not need to call AI a fraud to question an investment narrative. The more useful distinction is between value created for users and value captured by the businesses funding it.

For a finance team, the task is to connect adoption to accepted outcomes, operating cash and return on capital. For an investor, useful technology alone cannot justify any price. Market-wide certainty is less valuable than knowing which assumptions a valuation requires.

THE TAKEAWAY

AI's capability can be real.
Its economic returns still have to be earned.

This is a business-analysis perspective, not a forecast or a recommendation to buy, sell or use a fixed portfolio allocation. The cited financial and cost figures are dated evidence, not a live market snapshot.

References

The figures above refer to the periods stated in each source. The calculator uses illustrative assumptions; it does not represent the results of a particular company.

  1. Microsoft FY2025 annual report ↗

    Cloud margin and AI infrastructure discussion.

  2. Stanford AI Index 2025 ↗

    Historical inference-cost comparison.

  3. DeepSeek R1 official repository ↗

    Distilled models and technical documentation; developer-reported results.

  4. CoreWeave 2025 IPO prospectus ↗

    Historical customer concentration and disclosed business risks.

  5. Anthropic commercial data policy ↗

    Training defaults and stated exceptions; accessed 25 October 2026.

  6. European Commission: InvestAI ↗

    Announcement dated 11 February 2025; planned funding, not realised spending.