Real customers,
uneven capture

Follow the customer’s benefit all the way to cash

A legally independent customer may still fund purchases with new venture capital. A supplier-backed buyer may purchase a genuinely valuable service and later become self-financing. “Connected” does not mean fake, and “unconnected” does not mean durable. The useful questions are what pays the invoice now, what outcome the service creates, and what pays the next invoice once promotional or financing support ends.

Public evidence usually joins only part of that chain. A retailer’s payment to an application, the application’s model bill and the model provider’s cloud bill can all serve the same outside purchase. Adding the sales would not reveal three independent sources of demand. A saved wage or contractor expense is a customer benefit, not automatically an AI seller’s revenue.

Why adoption percentages can disagree without contradicting each other

Census changed its AI-use question on 17 November 2025, broadening it from producing goods or services to any business functions; the new results began appearing on 4 December. A jump across that change is not an unchanged-question adoption trend. This repairs part of the baseline’s measurement gap, not its missing current national series.

A Federal Reserve note compared late-2025 measures of roughly 18% of firms, an employment-weighted 78% for firms using any AI, and about 41% of individuals using generative AI at work. The definitions, weighting and populations differ. Employment-weighting firms that use AI does not mean all their employees use it. These are late-2025 observations, not a September 2026 nowcast.

Sources: Census · AI question wording updates; Federal Reserve · monitoring AI adoption. Inspection notes.

Menlo’s $37bn enterprise estimate uses a November 2025 survey of 495 U.S. decision-makers at active adopters plus market modeling, with important infrastructure and bundled-product exclusions. Ramp observes payments in a commercial card/invoice panel. Company seat figures describe licensed deployment. Experiments measure specified workers and tasks. None independently reconciles all customer funding, credits, renewals and intermediary purchases into an outside-demand total.

Sources: Menlo Ventures · 2025 enterprise study; Ramp · how its data works. Inspection notes.

The missing quantity is not assumed to be zero. The cases below make the economic links inspectable without manufacturing an aggregate that the evidence cannot support.

01 · Customer support

A resolved issue is a useful unit. A billable outcome is not necessarily an avoided cost.

Support offers a relatively clear economic task: resolve a customer’s issue at an acceptable quality level using less human effort. The relevant buyer unit includes escalation, repeat inquiries, review and remedies after wrong answers—not just the inexpensive automated portion.

A real deployment gain, with the right denominator

Generative AI at Work studies 5,172 agents in one software-company support network, mostly outside the United States. Its main staggered-rollout analysis finds about 15% more issues resolved per hour, with stronger benefits for less-experienced workers. Average resolution and satisfaction measures did not deteriorate, though effects differed by skill. The main deployment was in 2020–21 and assisted human agents; it did not test autonomous replacement.

The principal evidence is a comparison across a staggered rollout, not random assignment of the entire workforce. A small pilot does not change that: the paper lacked the pilot’s identified control-group records. Training and a finite license budget were part of deployment. These limits matter when moving from a useful study to a procurement decision.

Sources: Generative AI at Work · served arXiv v2. Inspection notes.

Klarna: a forecast, a workload equivalent and an expense line are different evidence

Klarna’s February 2024 launch announcement described 2.3 million assistant conversations in the first month and work equivalent to 700 full-time agents. It claimed fewer repeat inquiries and comparable satisfaction, while retaining access to human help. The widely repeated $40m profit improvement was a forecast for 2024, not a realized result in that announcement.

Sources: Klarna · first-month assistant announcement. Inspection notes.

On a narrow screen, scroll the table sideways.

Expense rose more slowly than revenue, but did not fall in dollars. That is consistent with better operating leverage; it neither measures the part caused by AI nor disproves savings against a higher no-AI counterfactual. The expense line includes more than the assistant, and revenue includes financial-business components such as receivables-sale gains. Revenue is not a count of comparable support requests.

The table uses reported expenses of 51 and 58, not the adjusted 48 and 57. No verified $40m annual AI saving, causal assistant share of the ratio improvement or model-contract renewal is inferred. The underlying supplier invoice, full implementation budget and no-AI cost comparison remain unavailable.

Sources: Klarna · Q2 2026 interim financial statements. Inspection notes.

The billing definition decides whether incentives line up

Intercom’s Fin tariff inspected on 19 September is $0.99 per billable outcome. The definition can include confirmed or assumed resolution and qualifying procedure completion, including handoffs. A quiet customer or completed routing step need not represent a human task permanently eliminated. This is a dated list tariff, not a measured average invoice or a provider cost.

Outcome pricing can bring the seller’s incentive closer to the buyer’s problem. It works only as well as the outcome definition and the buyer’s audit of errors, repeats and escalations. The provider must also cover failed attempts, supporting software and customer service.

Sources: Intercom Fin · pricing and outcome definitions. Inspection notes.

When staffing cannot change and extra output has no value, the cash-avoidable $4 may be much smaller or zero. Better customer retention or valuable additional service can make the benefit larger. Either direction needs measurement.

What support adds: a credible route to durable outside demand in repeatable, quality-controlled work. What a particular seller can retain depends on avoided cost, contract definitions, integration and alternatives. This supports useful utilization more directly than it establishes a large model-provider margin or the recovery of a particular data center.

02 · Software development

More completed code is promising. It is not the same as cheaper, better software.

The first investigation located positive developer field evidence without incorporating its full methods. This chapter does that work. The result is a stronger positive case—not a reason to erase the negative experiment or average incompatible percentages.

Sources: Cui and coauthors · three developer field experiments; METR · early-2025 developer experiment. Inspection notes.

The positive study’s early-window, two-company appendix gives an offer-of-access estimate of 4.66%, with a 3.56-point standard error. It is statistically inconclusive and uses a different sample window; it is not a universal rival estimate to the pooled induced-adopter effect. A buyer paying for every assigned license has to count non-use too.

Pull requests completed over a work period and minutes to finish a particular accepted task are different outcomes. Less-contextualized or standard implementation work can plausibly benefit from suggestions that an expert maintainer must inspect or rewrite. Those differences rule out a mechanical average; they do not identify every cause of the studies’ disagreement. Quality proxies in the positive experiments were mixed, not a comprehensive measure of defects, security or customer value.

Sources: Cui and coauthors · three developer field experiments. Inspection notes.

METR’s February 2026 follow-up encountered selection and timing problems, including willingness to work without AI and how to measure parallel activity. It did not supply a reliable updated causal effect. Its May survey of 349 technical workers records perceived value, not measured employer output or procurement bids. Neither freezing the old slowdown nor treating current enthusiasm as a randomized proof is justified.

Sources: METR · February 2026 follow-up limitations; METR · technical-worker usage survey. Inspection notes.

Follow the bottleneck, not the volume of generated code

The buyer needs accepted, maintained functionality. AI may shift the bottleneck to specification, review, integration, testing or deployment. More output helps when worthwhile projects are waiting; it can add maintenance obligations when the extra work is unnecessary or poor.

Buyer return = incremental contribution from shipped functionality + genuinely avoided labor or contractor costs − AI fees − integration, review, rework and risk costs.

This is an evaluation framework, not a monetary industry estimate. The public studies do not provide enough common evidence on project value, implementation and long-run defects to fill it with a defensible average. Multiplying all software wages by a headline effect would hide the missing terms.

Capture is difficult for the provider too. A subscription collects a fixed fee while inference, tool calls and retries can expand. Metering limits that exposure but confronts the customer with a marginal price. Microsoft’s July call describes usage-based monetization and says GitHub Copilot consumption affected segment margins, with improvement associated with the business-model change. It does not disclose a standalone Copilot cash margin.

Sources: Microsoft · FY2026 Q4 earnings call. Inspection notes.

A buyer may replace outsourcing, launch a previously uneconomic project, use a cheaper model for routine steps or switch assistant. Continued demand for coding help is not a guarantee of receipts for one provider or its compute supplier. No same-buyer annual coding-tool renewal cohort was obtained. Experimental participation is not evidence that participants subsequently renewed paid subscriptions.

What coding adds: a material positive demand case, more fully established than in the first investigation, alongside meaningful negative evidence. Continued paid work and higher utilization are plausible; stable per-task spending, vendor loyalty and a particular long-run cash margin remain unestablished.

03 · General office assistance

Time saved can justify a license without becoming payroll savings.

Dillon and coauthors study randomized access for 7,137 workers across 66 firms in 2023–24. In later experiment months, access reduced Outlook email time by about 1.4 hours per week per assigned worker. The estimate among workers induced to use the tool was about two hours. Weekly use stabilized below 40%, though many more tried it at least once.

The researchers did not directly observe work content, employer productivity or performance evaluations. Nor did they find comparable average reorganization in measured task quantities. That is bounded evidence of time savings, not proof of zero benefit: quality and work outside observed applications may matter. Participating firms were interested early adopters using an earlier product.

Sources: Shifting Work Patterns with Generative AI · served arXiv v4. Inspection notes.

A shorter email session can support concentration, better decisions or less after-hours work. Those can be valuable to a worker or employer without producing a separately priced output. Requiring immediate headcount cuts would miss real value; valuing every measured hour at wages and calling it available cash would invent a saving. Paid renewal depends on the outcomes the buyer actually values.

Sources: Microsoft 365 Copilot · enterprise pricing; Shifting Work Patterns with Generative AI · served arXiv v4. Inspection notes.

A well-targeted deployment can plausibly clear this hurdle. A poorly targeted one may not. A low fee can sustain a sensible purchase without generating a huge revenue pool; a higher fee or intensive paid agent use changes that comparison.

Paid deployment and repeat use are distinct evidence

Microsoft’s July 2026 call reported more than 30 million paid Microsoft 365 Copilot seats, faster additions and expansion at named large organizations. That is stronger evidence of commercial deployment than a free-user count. It is not an independent invoice audit, matched renewal cohort or a basis for multiplying seats by today’s list price and twelve months.

The earlier randomized trial estimates work patterns; the current vendor disclosure reports commercial scale. One does not update the other’s population or outcome. Better usefulness, bundling and large-scale experimentation may each help explain buying. Continued active paid use and subsequent renewal on known terms would distinguish them.

Sources: Microsoft · FY2026 Q4 earnings call. Inspection notes.

OpenAI’s September 2026 workplace study adds recurring-use evidence with a crucial boundary: it analyzes individual-account messages linked to Business onboarding information, not Business-account messages. Among selected active users, a cross-occupation task recurred the following month in 23.6% of prior-use cases versus 8.4% of matched no-prior-use cases. Sampling, persistent-user selection and nonrandom matching mean this is neither causal productivity nor paid-business retention.

Sources: OpenAI · Work at the Frontier. Inspection notes.

What office assistance adds: a wider basis for recurring payment at moderate prices without assuming dramatic payroll cuts. How much time value employers capture—and whether high-intensity agent consumption fits that budget—remains uncertain.

Repeat buying is evidence. Pricing power is another question.

Ramp’s February 2026 update, using January observations, reported roughly 4% monthly business-count churn for both OpenAI and Anthropic, and that 79% of Anthropic-paying businesses also paid OpenAI. These observations support repeat purchasing and show that customers need not choose one exclusive provider.

They are not dollar net retention, annual contract renewal or national proportions. A business can keep a small subscription while cutting spending or expanding elsewhere. Compounding one monthly retention rate across twelve months assumes a stable cohort and constant hazard not established by the report.

Sources: Ramp · February 2026 payment update. Inspection notes.

Card and invoice records establish buying more directly than intentions do. Ramp’s panel is nevertheless commercially selected and relatively technology-forward; it does not trace every customer’s funding, credit or discount. The June 2026 expansion changed coverage and vendor identification, so earlier and later levels should not be spliced into an unqualified series.

Sources: Ramp · how its data works; Ramp · June 2026 coverage expansion. Inspection notes.

A recent capture warning—not a sector-wide demand collapse

On a narrow screen, scroll the table sideways.

The arithmetic is about a 40.9% effective-price decline and a 9.7% decline in the high-spending measure. July was initially estimated near $7,400. The revised value must be used for that comparison. Neither series is a fixed-quality price index or a stable matched cohort of the same heavy users. Mix, channels, ranked-tail composition, volatility and possible seasonality matter.

The observations justify checking receipts separately from tokens. They do not identify how much is within-product repricing, sufficient cheaper work, a holiday effect or retrenchment. They do not establish a universal infrastructure-contract price decline.

Sources: Ramp · September 2026 pricing and spending update. Inspection notes.

The favorable mechanism matters too: lower prices can make more tasks worthwhile, and providers can preserve cash if their own costs fall sufficiently. Multiple models may complement one another. The evidence weakens both “nobody pays” and “more usage guarantees durable pricing power.”

Net customer revenue − inference and tools − support and distribution − development and training − taxes and replacement/reinvestment = cash available for financial claims.

These costs occur in different accounting lines and at different times. They still have to be counted somewhere. An application gross margin, a cloud-segment margin and complete model-provider cash margin are not interchangeable. This investigation does not turn the baseline’s attributed private-lab adjusted-profit reporting into audited distributable cash.

Enterprise licenses are not the entire market

An OpenAI September 2025 analysis classified about 70% of sampled consumer ChatGPT use as non-work by mid-2025. That establishes a substantial non-work boundary, not the fraction paying, subscription retention or 2026 profitability. Free assistance can have value without funding an equivalent amount of infrastructure.

Sources: OpenAI · how people use ChatGPT. Inspection notes.

Meta reported $59.363bn of advertising revenue in Q2 2026. Advertising is a route to monetization without a separately billed AI subscription; the revenue figure is not an AI-attributed cash receipt or a measure of incremental generative-AI profit. This chapter assigns no fraction of it to AI.

Sources: Meta · Q2 2026 results. Inspection notes.

Internal or defensive investment can improve an existing product or protect it from deterioration. The relevant comparison is incremental cash with and without the investment, including cannibalization and replacement. A plausible strategic rationale cannot validate any spending amount. Some new AI revenue can also displace consultants, outsourcing or ordinary software rather than add to total end-market spending.

More volume can restore revenue without restoring the cash needed for recovery.

The first investigation’s model used a $14.117bn hypothetical investment, calibrated only to a disclosed CoreWeave half-year cash-investment scale. At four operating years, a 10% required return, zero residual value and a 50% pre-financing cash margin, it requires $8.907bn of level annual revenue. This is not a measured fleet return or company target.

Sources: Whole-system foundation · recovery model. Inspection notes.

The new demand evidence strengthens the existence of useful paid work, but does not supply an empirical 50% margin or four-year revenue life. Instead, it separates price, volume, unit cost and buyer realization—variables compressed into the baseline’s constant-margin assumption.

On a narrow screen, scroll the table sideways.

A 20% price cut needs 25% more volume just to restore revenue. With unchanged cost per unit, the extra work raises variable costs from 35 to 43.75. Cash is only 41.25 rather than 50, leaving a $2.470bn net-present-value shortfall at the inherited investment scale.

With the assumed cost structure, restoring the original cash instead requires 44.4% more volume. If unit variable costs also fall by 20%, 25% more volume is enough. These are model identities, not measured customer responses to prices.

The 20% stresses are not calibrated to Ramp’s effective-token-price series. That mix-sensitive series cannot identify a cloud contract, loan or accelerator cohort’s repricing. “Usage rises,” “revenue holds” and “compute is cheaper” describe different parts of the economics; none alone establishes recovery.

The support and office models add another constraint: sensible buyer economics do not allocate the buyer’s entire surplus to a supplier. A moderate price can make a tool worth purchasing without supporting unlimited premium-model spending. Buyers may retain useful AI and still pay less per task.

What changes in the whole-pool answer: genuine paid demand is more strongly supported; price-to-cash conversion needs a tougher test; complete margins and asset lives remain assumptions. Sustainable recovery requires independent paid activity, adequate realized prices and controlled costs to endure over the asset and liability horizons—together.

The next link: paid demand must also fit an investment’s capacity and renewal bill

The 22 September capacity investigation retains the logic of the price/volume/cost example above but applies a separate disclosed-budget case. IREN’s initial Microsoft commitment does not automatically reprice when retail model prices fall. The future replacement cycle is a different question: what price and paid utilization will be available after another equipment purchase?

Under that case’s explicit replacement and operating-cost assumptions, 85.95% paid utilization at the selected price marker clears the whole ten-year hurdle. With a 20% lower price and unchanged unit costs, the requirement is 117.21%—infeasible within the modeled 200-MW IT cohort. This is a capacity-bound scenario, not a measured utilization level, market forecast or estimate derived from Ramp’s token-price series.

More efficient throughput, cheaper replacement or more valuable work could change the boundary. So could a customer’s decision to pay for capacity that is not fully used. Commercial reservation, physical load and useful processing must remain separate. The original 35/15 cost split and buyer studies above are not recast as measured IREN economics. See the replacement cycle and its assumptions →

What would turn the commercial case into an investment case?

The financing question distinguishes useful service, the customer’s ability to pay and the creditor’s right to collect. The documented next link is Meta’s take-or-pay relationship supporting DDTL 4.0, identified by DBRS. A well-funded customer can pay despite poor utilization, and the facility is expected to amortize fully from contracted cash. That expectation is not observed borrower compliance. Rating analysis; full repayment proposition.

Do not transplant the price stress into that contract. The 20% price changes and 35/15 cost split in this chapter are conditional illustrations, not observed DDTL costs or a forecast of its invoices. Lower retail model prices can first squeeze a buyer carrying fixed compute commitments; service availability deductions and costs can instead weaken the project’s collections. Contract terms determine the immediate transmission. Cash thresholds and loss allocation test that separately.

On a narrow screen, scroll the table sideways.

Better results across these measures would strengthen sustainable expansion. Repeated weak renewal, worsening cash conversion and greater financing dependence would strengthen concern. This is not a numerical sector-wide trigger or assigned crisis probability.

Continued useful demand therefore answers neither the current cash certificate nor the next renewal. DDTL 4.0’s expected full amortization differs from DDTL 5.5’s approximately five-year financing against initial contracts averaging about three years. Replacement customers matter differently in the two structures. Dated financing contrast; favorable and adverse paths.

A good customer purchase and adequate provider recovery remain separate tests. The capacity chapter now shows why a high first-term project measure can omit the facility bill or rely on unobserved post-contract value. It preserves a serious profitable-continuation case alongside an adverse case with useful output but inadequate investment return. Favorable and adverse paths.

Claim-level correction · 19 September 2026

The NVIDIA forward establishes an obligation, not verified receipt.

The underlying first whole-system investigation said that an executed prepaid-forward agreement acknowledged “$1.5 billion paid by NVIDIA”; its source note repeated that characterization. The second investigation withdraws those evidentiary assertions. The reader account had retained the narrower obligation/receipt distinction while the issue was checked.

The Prepaid Forward Contract between Energy Global, LP and NVIDIA Corporation, dated 17 August 2026, defines the purchase amount. Section 1(f) requires cleared-funds transfer within three business days. Annex A’s wire instructions are provided separately. The opening consideration language is not confirmation of settlement, and generic acknowledgement wording in an attached form of lock-up agreement is not a receipt for this investment.

Sources: Energy Global, LP / NVIDIA · Prepaid Forward Contract. Inspection notes.

The URL filename contains exhibit1027, but the September 4 amendment’s index lists this forward as 10.26 and the separate share-purchase agreement as 10.27. Identify the contract by title, parties, date, URL and section, not the filename number alone.

Sources: SB Energy · 4 September 2026 exhibit-only amendment. Inspection notes.

The inspected NVIDIA August current report and July-quarter filing did not supply confirmation; the July balance-sheet date also precedes the contract. The SB Energy registration-statement body could not be retrieved and was not treated as inspected evidence in either direction.

Sources: NVIDIA · 17 August 2026 current report; NVIDIA · quarter ended 26 July 2026 filing. Inspection notes.

Receipt unverified does not mean nonpayment established. No amount belongs in verified received cash on the strength of this contract. This is not customer revenue and is not combined with the separate offering-linked share purchase or contingent guarantee.

The correction weakens the claimed certainty of that liquidity example, but changes none of the first investigation’s numerical model outputs: its ledger and code contain no prepaid-forward receipt input. A later receipt-confirming disclosure could settle cash status; it would not make the original description of what this contract proves correct.

This is separate from the Amazon payment correction already made in Cash, commitments and capacity. That earlier correction remains credited there.

Methods, limitations and reproducible work

This chapter draws on the second investigation and the existing whole-system foundation. The underlying demand investigation inspected original study methods, buyer accounts, vendor tariffs, payment-panel explanations and the named contract material on 19 September 2026. Public prices and vendor claims remain dated to that inspection, not asserted as live tariffs.

Case selection was purposeful: support has bounded output, coding has disputed productivity and intensive consumption, and office assistance has diffuse benefits. These are not three representative slices from which to estimate a national total. Private-lab profitability, security valuations and wider credit exposure were not re-underwritten; those parts of the foundation retain their 18 September evidence boundary.

The largest gaps are economically consequential

No dataset reconciled final-customer receipts, credits, venture-funded use and inside-network transfers. Complete buyer invoices and implementation costs, same-buyer annual renewal cohorts and standalone all-cost provider cash margins were not obtained. Monthly buying and short experiments cannot establish a four- or six-year cash stream.

The stronger causal studies concern older tools and selected early adopters. Current payment panels and company disclosures improve timeliness, not representativeness or causal identification. Vendor connections in the positive developer and office studies, and cooperating firms in the support study, warrant scrutiny without automatically invalidating results. Underlying confidential microdata were not independently re-estimated.

The support and office texts have unresolved header/internal-date discrepancies: the served arXiv headers say November 2024 and November 2025, respectively, while both carry an internal August 24, 2026 date. The underlying fieldwork periods are retained. The pages are not presented as new September experiments or evidence of specific changes between manuscript releases.

Full current Census downloads, Ramp raw panel/API data and the SB Energy registration-statement body were not obtained. Some scholarly landing pages failed; the accessible author/arXiv versions named below were used. Missing retrieval is not evidence of absence.

Inspect the numbers

30 selected observations, qualifiers and source locations (CSV) · Buyer and recovery calculations (Python).

The CSV is a reader extract of the supplied observation ledger, not a market-size dataset. Numerical inputs and units are retained. Save both files together and run python demand-calculations.py, or pass --observations with the CSV path. The Python standard-library script needs no network access and makes no file changes. It reproduces the support, office, Klarna, payment-panel and recovery arithmetic; checking those identities does not verify sources or validate assumptions.

Research model: GPT-6 Pro. These are source-linked findings and conditional calculations, not an independent audit of private accounts or confidential study microdata. The scope and source limitations determine what each claim supports.

The remaining question is specific: which independently funded customers keep paying enough, after the cost of useful service, for the particular obligations the build-out has created?

Return to the financing mechanisms · The account of changes