Abiola BabsalaamEnterprise Architect
GitHub

Writing / Article

The AI Bubble Is Four Different Questions

AI products can create value while providers lose money, infrastructure earns poor returns, and assets remain overpriced. Each claim needs its own evidence.

Abiola Babsalaam

11 min readAI

Independent writing, human reviewed

“Is AI a bubble?” sounds like one question. In practice, it usually refers to four:

  • Do AI products help people complete valuable work?
  • Can providers turn usage into durable profit?
  • Will demand justify the infrastructure being built?
  • Do current asset prices assume plausible future cash flows?

These questions are connected, but evidence for one does not answer the others. A coding assistant can save time while its provider loses money. A model provider can become profitable while data-center projects earn poor returns. AI can improve real work while related stocks or private companies remain overpriced.

My view is not that speculative excess is absent. It is that “the AI bubble” is too imprecise to diagnose where that excess sits.

Use four ledgers

LedgerQuestionUseful evidence
Product utilityDoes AI improve a valuable task?Retained usage, successful completions, error rates, measured time or quality changes
Operating economicsCan a provider earn durable profit from demand?Revenue, gross profit, compute expense, sales costs, operating income, cash flow
Infrastructure paybackWill capacity earn an adequate return over its useful life?Utilization, realized revenue per unit, energy and operating costs, replacement spending, residual value
Asset pricesDoes the price imply plausible future cash flows?Valuation, dilution, competitive position, discount rates, scenario analysis

The framework matters because the ledgers can move in opposite directions.

Suppose better models attract more users while competition lowers inference prices. Product utility and adoption rise. Customers benefit. Yet infrastructure owners may receive less revenue per unit of capacity while remaining committed to power, leases, debt, and hardware replacement. In that scenario, AI succeeds as a product while infrastructure returns deteriorate.

Builders and technology buyers can skip to the project test below. Readers evaluating the investment cycle should first examine what the public evidence does and does not establish.

Product utility: substantial use, uneven results

US Census Bureau data show that businesses and workers are using AI in real workflows. The evidence does not determine how much of that use is paid, retained after pilots, subsidized, or economically valuable.

From December 2025 through May 2026, between 17% and 20% of US businesses reported using AI in at least one business function during the previous two weeks. Use reached 37% among firms with at least 250 employees. Information and finance reported higher adoption than the national rate, while use among the smallest firms remained below 20%.

A March 2026 Census household survey found that 56% of US workers had used AI for at least one of 11 job tasks. Among respondents who used it during the previous week, 31% estimated that it saved one to two hours, while 4% said it required additional time.

That 56% figure covers a broad definition of AI, including machine learning, voice recognition, and virtual agents. It is not an estimate of large language model or generative AI adoption. The time estimates are also self-reported rather than audited productivity measurements.

A six-month NBER field experiment involving 7,137 knowledge workers at 66 firms provides narrower evidence. During the latter half of the experiment, active users of an integrated generative AI tool spent about two fewer hours on email each week. The researchers did not detect broader changes in the quantity or composition of work from individual access alone. Some authors worked for Microsoft, which made the tool, during the study.

The defensible conclusion is limited: AI has substantial real-world adoption, and some workflows show measured benefits. These findings do not establish universal productivity gains, willingness to pay, retained demand, or adequate returns on industry investment.

Operating economics: revenue is real, durable margins are unresolved

The current cycle is not built entirely on companies waiting for their first customers.

NVIDIA reported $89 billion in data-center revenue for the quarter ended July 26, 2026, up 117% from the prior year. Total quarterly revenue was $96.2 billion, with a 75% GAAP gross margin. Infrastructure vendors are collecting substantial revenue now. That does not show whether purchasers will earn sufficient returns from the hardware.

Anthropic’s amended S-1 provides a view inside one model provider. It reported $33.2 billion in revenue and $18.8 billion in compute and infrastructure expense for the first nine months of 2026. Compute expense declined from 172% of revenue in the comparable 2025 period to 57%.

This was an improvement in compute economics relative to revenue, not necessarily proof of greater technical efficiency. Rapid revenue growth against capacity contracted earlier can lower the ratio even without a comparable improvement in the physical efficiency of computation.

Anthropic reported a $1.26 billion GAAP operating loss for the nine months. It also reported $4.08 billion in adjusted operating income after excluding stock-based compensation and non-cash charitable contributions. Both figures matter because they answer different questions about accounting cost and management’s adjusted view of operations.

The filing says compute is generally contracted 12 to 18 months before it becomes available. That creates a forecasting risk: capacity commitments are made before the corresponding customer demand is known.

These disclosures establish that customers pay for model access and that one provider’s compute-expense ratio improved as revenue grew. They do not establish durable industry margins. Prices can fall, research and sales expenses can remain high, and capacity purchased in advance can become costly when forecasts are wrong.

Infrastructure payback: large spending must survive replacement cycles

Microsoft reported $45.2 billion of capital expenditure in its fiscal 2027 first quarter. Roughly three quarters went to short-lived assets, primarily CPUs and GPUs, including memory and storage. Microsoft generated $58 billion in operating cash flow and $18.6 billion in free cash flow, while attributing pressure on free cash flow to higher capital spending.

Alphabet reported $51.4 billion of third-quarter capital expenditure, most of it supporting AI infrastructure. About 65% of technical-infrastructure investment went to servers and networking equipment with useful lives of six years. The rest went to data centers with useful lives ranging from seven to 40 years.

Alphabet also disclosed $924 billion of future purchase commitments and $152 billion of pre-commencement leases and other obligations. The company described the purchase-commitment figure as a maximum contractual amount. Some agreements include variable pricing and volume flexibility, so actual spending could be lower. These figures should not be read as fixed minimum cash outflows.

Amazon said it expected approximately $220 billion in cash capital expenditure for 2026. Its trailing 12-month free cash flow was an outflow of $42.4 billion, driven primarily by a $94.8 billion year-over-year increase in property and equipment purchases. Amazon said the increase mainly reflected AI investment. Its broader AWS business was growing at the same time, but Amazon did not isolate AI-specific cloud revenue in this disclosure.

The Bank for International Settlements describes hyperscalers financing AI infrastructure through bonds, private credit, joint ventures, long-term leases, and capacity-purchase agreements. Some structures place project debt outside the hyperscaler’s balance sheet while leaving it with multi-year payment or purchase obligations.

The company disclosures show that several major buyers have profitable businesses and existing cloud customers. It is reasonable to infer that these buyers are not relying solely on speculative startup funding. The evidence does not establish the funding mix or customer quality of the entire data-center market.

A compact infrastructure payback test

A data-center project should be evaluated on two related measures.

First, estimate its unlevered return:

project value = discounted operating cash flow + residual value - initial capital expenditure - replacement capital expenditure

Operating cash flow should reflect:

  • utilization rather than installed capacity
  • realized revenue per billable unit after discounts
  • electricity, cooling, networking, maintenance, and staffing
  • downtime and capacity reserved but not sold
  • the useful life of servers and networking equipment
  • the cost and timing of replacements

Second, test financing resilience separately:

debt-service coverage = cash available for debt service / required debt service

This avoids counting financing costs twice while still asking whether a viable asset can survive its funding structure.

Microsoft’s disclosure that roughly three quarters of quarterly capital expenditure went to short-lived assets, and Alphabet’s six-year useful life for servers and networking equipment, make replacement spending central to any payback estimate. A project can remain busy and technologically useful yet disappoint investors if realized prices fall faster than operating and replacement costs.

Asset prices: utility cannot settle valuation

Asset valuation requires assumptions that adoption surveys cannot supply: future revenue, margins, competition, replacement costs, financing, dilution, discount rates, and market share.

A productive AI system is evidence for product utility. It is not evidence that every supplier deserves its current valuation. Conversely, a collapsing stock price would not prove that the underlying technology had stopped working.

Falling inference prices illustrate the distinction. They can expand usage and customer surplus while compressing provider margins. They can also make previously purchased capacity less valuable than projected, particularly when buyers face fixed commitments or accelerated hardware replacement.

There is therefore no market-wide valuation conclusion available from the operating evidence alone. Each company and infrastructure project needs its own cash-flow scenarios, including cases with lower prices, slower demand, higher replacement costs, and weaker market share.

A safer, measurable AI project test

A successful demonstration is not a business case. Evaluate one recurring workflow with a concurrent control, randomized rollout, or crossover design where feasible. A 90-day period is often long enough to observe repeated behavior, but the required duration and sample depend on task frequency and outcome variability.

1. Establish deployment prerequisites

Before exposing real data or consequential decisions to a model:

  • Classify the data involved, including personal, confidential, regulated, and credential-bearing information.
  • Use approved vendors and model terms, with explicit retention and training controls.
  • Apply least-privilege access to models, tools, retrieval systems, and logs.
  • Test prompt injection, unauthorized tool use, and data exfiltration paths.
  • Define logging, monitoring, incident response, and rollback procedures.
  • Require human approval for outputs that affect rights, safety, employment, credit, health, legal obligations, or other high-impact decisions.

A project that cannot satisfy these conditions is not ready for a production-value test.

2. Predefine the experiment

Specify before launch:

  • the unit of analysis, such as a task, case, customer, or employee-week
  • eligible and excluded tasks
  • primary success metrics and acceptable error thresholds
  • the minimum observation count
  • treatment and control assignment
  • how failed, abandoned, escalated, and retried tasks will be counted
  • the uncertainty measure, such as confidence intervals around the estimated effect

A simple before-and-after comparison can be distorted by seasonality, staffing changes, task mix, learning effects, or unrelated process improvements.

3. Record the baseline and full cost

Measure completion time, successful output, errors, rework, escalation, abandonment, loaded labor cost, and any associated revenue or expected loss.

Include model charges, retrieval, human review, integration, maintenance, evaluation, monitoring, security, compliance, retries, and incident handling.

Convert released time into money before combining it with monetary benefits:

realizable labor value = usable hours released × loaded labor cost per hour × realizability factor

net monetized value = realizable labor value + incremental gross profit + expected loss avoided - total AI workflow cost

The realizability factor represents the share of released capacity that can actually be redirected to useful output, reduced hiring, or another measurable benefit. Keep quality, latency, customer experience, and risk outcomes as separate metrics when they cannot be credibly monetized.

4. Measure retained behavior and alternatives

Track repeat usage, override rates, successful completions, and continued use after any mandatory pilot ends. Compare the AI workflow with a credible alternative, such as better search, templates, conventional automation, training, or process redesign.

Set redesign or exit conditions in advance. Examples include review time erasing the measured benefit, no retained usage, an error rate above the workflow’s tolerance, or a cost per successful task above the control.

These measurements primarily answer the product-utility ledger and part of the buyer’s own operating economics. They do not reveal a model provider’s full cost structure, the utilization of an upstream data center, or whether an AI-related stock is fairly priced.

What evidence would distinguish durable investment from excess

Evidence of overinvestment would include capacity expanding faster than paid usage, declining utilization alongside rising depreciation or financing costs, model prices falling faster than serving costs, infrastructure obligations being repeatedly refinanced, and customers abandoning deployments after full costs are measured.

Evidence of more durable economics would include retained paid usage after discounts expire, improving GAAP operating margins at model providers, capital spending moderating while AI-related revenue continues growing, independent productivity results across larger field studies, and compute capacity being repurposed economically when particular models lose relevance.

The available evidence supports a narrower conclusion than either “AI is a bubble” or “AI is not a bubble.” Product utility exists in meaningful but uneven contexts. Provider revenue is real, while durable operating margins remain uncertain. Infrastructure payback depends on utilization, pricing, useful lives, and replacement costs that are not yet resolved. Asset prices require company-specific analysis.

AI is a real technology inside a large investment boom. That combination permits valuable products, profitable suppliers, overbuilt capacity, failed projects, and overpriced assets to exist at the same time.

Sources