AI’s Price Tag Reopens Debate Over a Possible Bubble

On August 1, 2024, the debate over a possible AI bubble could be framed with public data, but it required three separate questions: how much capital was being spent, what revenue it produced and what return each company expected. David Cahn’s analysis for Sequoia turned that gap into a revenue scenario; it was neither an accounting forecast nor proof that every investment was unproductive.

The investment was material. Microsoft reported $19 billion of capital expenditure including finance leases in the quarter ended in June; nearly all related to cloud and AI, split between data-centre infrastructure and servers. The figure measures quarterly assets and contracts, not the exclusive cost of training generative models.

Alphabet reported $13 billion of capital expenditure in the second quarter, dominated by servers and data centres, and expected a quarterly pace roughly at or above the first quarter. Meta placed its full-year forecast at $37 billion to $40 billion and tied future growth to AI research and products. The accounting definitions are not identical, so adding them shows scale rather than a perfect comparison.

The supplier collects while the buyer depreciates

NVIDIA showed the visible side of demand: it recorded $26 billion of quarterly revenue, up 262% year over year, for the period ended in April. That was NVIDIA’s revenue; for customers, chips and systems became investment that had to be depreciated and produce capacity over several years.

The timing difference is central. A supplier books a sale; an operator deploys equipment, sells compute or products and recovers spending gradually. High investment and little attributable revenue in one quarter can be normal for infrastructure. It becomes concerning when utilisation, pricing or demand miss the plan across the asset’s useful life.

What Sequoia’s question calculated

Cahn compared possible infrastructure revenue with the spending needed to sustain data centres and proposed a $600 billion annual gap. The value depended on assumptions about GPU supply, centre costs, margins and market allocation. Change utilisation, prices or asset life and the answer changes.

The number is therefore a sensitivity test, not a bill the sector must earn by an exact date. A rigorous reading copies the variables and builds scenarios: occupied capacity, revenue per unit of compute, energy, maintenance and margin. It then shows which assumption creates the shortfall.

Unit economics live in completed work

For a customer, the return is not “having AI”. Measure it by task: time saved that does not return as review, errors prevented, incremental sales or a service that was previously impossible. Subtract licence, integration, security, data, supervision and failures. A frequently used pilot may still destroy value if every output must be rebuilt.

For a cloud provider, track occupancy, revenue, variable cost and contract duration. For an application, retention and margin after inference cost. For an advertising platform, incremental revenue without loss of trust. Mixing those units can celebrate adoption while hiding that the product subsidises each query.

A denominator prevents the most common trick. Saying AI revenue is growing quickly does not show whether it starts from a small base or how much capital supports it. Compare revenue with employed assets, operating expense and time. Separate new income from cloud sales that would have existed without the generative feature.

Utilisation requires measurable capacity. A centre may be contracted, built, energised or occupied by paid workloads; these are different states. Advance reservations help forecast demand but are not usage or margin. The record keeps every stage so a commercial promise does not look like cash flow.

Marginal cost changes with model, length, hardware, optimisation and customer mix. A price covering inference today may fail when users add context or choose a more intensive feature. Track margin by cohort and use case rather than only as a company average.

When “bubble” is a useful description

A technology can work and attract overinvestment. The signal is not a failed demo but prices or valuations that require improbable growth, capacity built without customers or projects unable to cover their cost of capital. There may also be excess in one layer and shortage in another: many similar models but insufficient energy, networking or integration.

The internet comparison shows that durable infrastructure and investor losses can coexist. It does not prove that every present investment will find a use. Look for contracts, utilisation and cash flow; historical narratives provide analogies only.

Scenarios rather than one verdict

A conservative scenario lowers growth, price and utilisation while raising energy or review; another tests what happens if capacity fills and query cost falls. Both use the same horizon and definitions. Their difference reveals the controlling variables and signals to seek in later results.

Companies may justify defensive investment: failing to build could surrender a future market. That option value exists but has a price. State it as a strategic bet rather than mixing it with observed return. A reasonable decision under uncertainty can still end with underused assets.

Early indicators include paid adoption, renewal, intensity of use and demonstrated savings. Later ones include depreciation, impairment and margin. Watching both avoids waiting years to identify a problem and avoids declaring failure before infrastructure enters service.

Energy and time are not footnotes

A facility may contain servers while remaining constrained by power connection, networking or cooling. Commissioning time affects return because idle capital produces no paid work. The audit records announced capacity and actual operating date alongside energy cost per workload, without assigning all centre electricity to one product.

Model efficiency changes the calculation too. Quantisation, better chips, batching and software can lower response cost while larger models or longer contexts consume the saving. Measure the cost of a comparable task over time. A lower token cost does not ensure a lower result cost if the application sends far more tokens.

Competition pressures prices and can raise utilisation, but it also reduces margin. A sound scenario tests both forces. Assuming every query becomes cheaper while price holds describes a best case; assuming only a price war ignores differentiation and contracts. Sensitivity shows which combination sustains the investment.

Finally, inspect who bears the loss if the scenario fails: chip supplier, cloud operator, developer or customer. Minimum contracts, purchase commitments and depreciation distribute risk unevenly. Following that allocation explains how a boom can enrich one layer and leave excess capacity in another without a single moment when the bubble bursts.

A record for following the debate

Each quarter, record spending with its accounting definition, commissioned capacity, attributable revenue, margin, paying customers and expected useful life. Separate a published fact, company guidance and an analyst estimate. If a number cannot be assigned to one column, do not use it to prove return.

The transferable skill is to audit AI economics as a bridge from capital to utilisation, revenue and margin while keeping accounting measures and scenarios separate. It reveals overinvestment without denying technical utility and recognises a sound business without mistaking supplier growth for buyer returns.