Who Captures the Returns from AI?

Fact cutoff: September 5, 2026

---

The thesis

AI can create a large amount of economic value without distributing profit in proportion to each participant’s technical contribution. Profit tends to remain at the layer that controls a scarce input, a billable interface, and the right to reprice. Risk tends to remain at the layer that locks in fixed cost, hardware generations, and long-duration capital before end demand is proven.

The most observable fact today is not the eventual contribution of AI to GDP. It is the route taken by the first dollars: enterprises pay for Copilot seats and cloud consumption; advertisers pay for impressions and actions; cloud platforms procure compute; compute operators buy accelerators, power, and data-center capacity; systems vendors reserve memory and manufacturing; and capital providers fund the time gap.

Every layer can report growth. Growth becomes a return to a specific class of capital only after it passes five gates:

Verifiable productivity or revenue uplift → end-customer payment → corporate margin and operating cash → interest, hardware refresh, and principal → residual cash to common equity

Technical success opens only the first gate. Contracts are not cash; revenue is not free cash flow; EBITDA is not distributable equity cash; scheduled interest does not prove a positive return for project equity.

1. Find the terminal payment before allocating industry profit

Productivity value is a net outcome: qualified time saved, incremental output, and avoided errors, less integration, supervision, retry, and change-management costs. Without workflow acceptance it remains a capability; without continued payment it is not yet industry revenue.

Microsoft offers evidence close to a value-to-payment loop, with an important boundary. On its FY2026 fourth-quarter call, management said Microsoft 365 Copilot had passed 30 million paid seats. It also relayed that an NHS England trial saved employees an average of 43 minutes per day before a planned rollout to 505,000 staff. The paid-seat count is company-reported commercial adoption, not audited standalone Copilot cash collections. The time saving is a customer result repeated by Microsoft, not an independently replicated causal estimate in this report.

On the same call, Microsoft said annual Azure revenue exceeded $100 billion for the first time, up 41%. Fourth-quarter Intelligent Cloud revenue was $39.3 billion, up 32%, with a 41% segment operating margin. This shows that newly delivered capacity can be monetized at scale. It does not establish a positive return on investment for every customer or isolate how much revenue came from AI rather than the wider cloud portfolio. Microsoft FY2026 Q4 earnings call

Meta shows a different terminal-payment mechanism. Users do not pay directly for the recommendation model; marketers pay for impressions and actions. In the second quarter of 2026, advertising revenue rose 27% year over year, ad impressions rose 14%, and average price per ad rose 12%. Meta said AI recommendations had improved engagement and monetization, while also stating that it could not precisely attribute advertiser spending decisions to individual trends.

The separation between value, revenue, and profit is visible in the same filing. Total revenue rose 28%, but operating income fell 8%, while capital expenditures, including principal payments on finance leases, reached $31.08 billion. A better product and higher revenue can coexist with infrastructure, third-party cloud, token, personnel, and other costs moving current-period profit in the opposite direction. Meta 2026 Q2 Form 10-Q

Token volume is therefore a weak endpoint. Track workflow acceptance; the payment unit; price versus cost per successful task; and whether renewal, expansion, and cash collection cover inference, sales, support, and integration.

End-customer payment is the external revenue of the entire capital stack. Many contracts between the other layers merely redistribute that payment—or redistribute an expectation that it will arrive later.

2. Where profit sits now: a cash-flow cross-section, not a league table

The six primary disclosures used in this report show a clear but non-comparable set of layers.

Sources: NVIDIA FY2027 Q2 Form 10-Q · TSMC 2026 Q2 results · CoreWeave 2026 Q2 Form 10-Q

These margins are not directly comparable: NVIDIA and TSMC report consolidated gross margin, Microsoft reports segment operating margin, and CoreWeave has a different ownership and capital-intensity profile. The point is to locate timing and seniority, not rank companies.

A vendor can recognize profit at delivery and a platform as service is consumed, while project equity waits until operating costs, leases, debt service, maintenance, and refresh. One unit of compute can therefore produce supplier profit, platform revenue, lender cash, and an equity loss in the same period.

Five rights largely determine who captures the economics:

1. Control of a binding bottleneck.
2. The ability to define the billing unit.
3. Ownership of the customer relationship and outcome data.
4. Repricing, routing, and migration options.
5. The ability to place fixed-duration risk with capital suited to hold it.

A replaceable layer with no billing right may retain little profit; a smaller technical contributor may earn substantial rent by controlling customer access or non-substitutable capacity. As capacity normalizes and workflows become sticky, the allocation can reverse.

This is why “AI revenue” is not “AI return.” Operating cash reflects business costs and working-capital timing; debt return requires interest and principal recovery; equity receives the residual after fixed claims and competitive reinvestment. A company can clear one gate and fail the next.

3. When financing support reduces total loss—and when it only moves loss

Credit support is neither inherently safe nor inherently dangerous. The test is whether it changes a real state variable.

Customer prepayments, completion support, enforceable step-in rights, and residual-value arrangements can preserve project value by reducing idle time, bringing forward revenue, or improving migration, takeover, and re-leasing. A lower cost of capital also benefits the borrower, but its components matter: some reflect lower risk or real execution costs; others transfer income from lenders or support providers. A lower financing price cannot be counted in full as a system-wide resource saving.

Support that changes only the payer for an existing asset pool primarily reallocates loss. If it also induces new capacity without sufficient end demand, it can increase aggregate loss.

NVIDIA’s latest 10-Q makes the distinction unusually visible. As of July 26, the company reported $279 billion of supply and capacity commitments, up from $119 billion in the prior quarter. It separately reported $56 billion of AI cloud agreements and third-party data-center leases not yet commenced, a maximum gross exposure of $3.5 billion on land, power, and shell guarantees, and an August agreement with a total $105 billion cap for guarantees supporting the SB Energy project.

Those numbers have different legal and economic meanings and cannot be added into “financing.” The SB Energy guarantees generally activate only as ready-for-service conditions are met, leases commence, and nine construction phases enter service. Exposure then declines with tenant performance. The activation curve, coverage definition, and recovery path matter more than the headline cap. NVIDIA FY2027 Q2 Form 10-Q

The Meta–Blue Owl Hyperion joint venture shows why moving assets off one balance sheet does not necessarily move all economic risk. Blue Owl-managed funds own 80% and Meta owns 20%; the parties committed to fund their shares of approximately $27 billion of development cost for buildings and long-lived power, cooling, and connectivity infrastructure. Meta signed initial four-year operating leases for all facilities and retained extension options, while also providing a conditional, capped residual-value guarantee covering the first 16 years of operation.

External infrastructure capital gives Meta operating flexibility. The guarantee supplies part of the contract tail sought by capital providers. Risk allocation therefore cannot be inferred from the 80/20 ownership split alone. It also depends on lease renewal, residual value, campus specialization, replacement tenants, and the guarantee trigger. Meta–Blue Owl Hyperion announcement

At the system level, the identity is simple:

Total loss falls only if support raises terminal cash, lowers a real cost, shortens idle time, increases completion probability, or improves recovery.

A guarantee can lower loss given default for a lender while leaving total project loss unchanged. It can even increase system loss if the lower cost of funding induces capacity that will never be paid for. The guarantee is then an allocation rule, not a productivity technology.

Ten separately underwritten projects may still share one model buyer, GPU generation, power market, and refinancing window. Treating guarantee calls as independent understates exposure when the support provider is also under pressure. A strong balance sheet moves the default boundary; it cannot manufacture independent terminal demand.

4. A fully specified hypothetical

The following uses currency units and does not represent any real company, project, or security.

A customer pays 50 per year under a contract with a fixed four-year calendar term; a delay does not extend the end date. Power and operating cash cost are 20 per year. Equipment and installation cost 100, and that equipment has a terminal recovery value of 20 at the end of year four. Initial financing is 60 of debt and 40 of equity. Principal amortizes at 15 per year, beginning at the end of year one; interest is charged on beginning-of-year debt. We ignore tax, working capital, risk pricing, opportunity cost, and discounting to compare nominal cash. The residuals below are not IRRs or fair values.

Without effective support, the debt rate is 12%. Interest on beginning principal is 7.2, 5.4, 3.6, and 1.8, or 18 in total. A construction-coordination failure delays commissioning by one quarter while drawn debt continues to accrue interest. The project loses 12.5 of revenue but avoids 5 of operating cost, reducing operating cash by 7.5.

Cumulative cash returned to equity over four years is:

187.5 revenue − 75 operating cost + 20 terminal equipment recovery − 60 principal − 18 interest = 54.5

After the initial 40 of equity, the nominal cumulative residual is 14.5. Depreciation is an accounting allocation and is not deducted a second time from this cash calculation.

Now assume completion support and enforceable step-in rights reduce the debt rate to 8% and prevent the delay. Total interest becomes 4.8 + 3.6 + 2.4 + 1.2 = 12. Additional coordination and execution consume real resources of 2; assume no separate guarantee profit or other fee.

Cumulative cash returned to equity becomes:

200 revenue − 80 operating cost + 20 terminal equipment recovery − 60 principal − 12 interest − 2 execution cost = 66

The nominal residual after the initial equity is 26, an improvement of 11.5 for equity. Consolidating the project and its debt and equity providers gives a different result: without support, net cash is 187.5 − 75 + 20 − 100 = 32.5; with support, it is 200 − 80 + 20 − 100 − 2 = 38. The combined improvement is 5.5. Equity’s additional interest saving of 6 is offset by lower interest income for capital providers. It is not additional resource value. The 5.5 comes from avoiding 7.5 of lost operating cash, less 2 of real execution cost; this project-level measure is not economy-wide welfare.

Now change one fact. Suppose the guarantee finances a second equipment package costing 100, with ultimate recovery of 20, but terminal payment remains 200. The project and its capital providers incur an additional nominal net cash loss of 80 on a combined basis. If the guarantor later pays 30 to lenders, that payment transfers 30 of loss from lenders to the guarantor. It does not recover the missing 80.

The example shows why technical success and capital returns split. The equipment supplier records a sale early, lenders receive scheduled interest and principal, and project equity remains exposed to revenue timing and residual value. A gain for one capital provider is not automatically a gain for all project participants.

5. Three falsifiable propositions and the next observations

Proposition 1: paid-demand measures should explain operating cash better than total token volume

Where comparable data exist, use information available at the time—paid seats, pricing, retention, and service cost—to estimate subsequent operating cash, then compare the error against a simple token-growth baseline. If the paid-demand measures repeatedly fail to improve forecast error, the proposition is not supported. Rising usage alongside falling profit is not sufficient to perform this test. Observe the next two reporting quarters; where sample coverage is inadequate, leave the claim untested.

Next observation: the next Microsoft and Meta 10-Qs and earnings calls, with emphasis on paid expansion, usage-based pricing, infrastructure cost, and operating margin—not model benchmarks.

Proposition 2: as supply scarcity eases, profit will migrate from hardware bottlenecks toward owners of customer workflow

The proposition gains support if incremental margins at NVIDIA and TSMC normalize as capacity expands while cloud and application unit economics improve. It is falsified, at least for the observed period, if upstream margins keep rising while downstream capital intensity and service cost continue to increase.

Next observation: NVIDIA supply commitments and gross margin; TSMC utilization, leading-edge mix, and margin; Microsoft and Meta’s cost to serve paid AI usage.

Proposition 3: completion and step-in support should preserve operating value more effectively than a payment guarantee alone

Compare commissioning delays, cash recovery, and migration costs for projects with similar power conditions, customers, construction stages, and hardware generations. If the completion-and-step-in group repeatedly fails to improve these outcomes, the proposition is not supported. Growth in guarantee caps alone neither verifies nor falsifies it.

Next observation: activation and third-party resale of NVIDIA commitments; CoreWeave’s bridge from operating cash to capital expenditure and debt service; and Hyperion’s completion, lease commencement, and guarantee status.

Conclusion: profit belongs to the party that turns value into cash without locking its future

The AI value chain is not a conveyor belt that distributes profit evenly. The end customer first determines whether real value exists. Applications and platforms decide how it is billed. Cloud and compute operators translate demand into deliverable capacity. Chip and manufacturing vendors earn scarcity rents during constrained periods. Lenders take contractual cash before equity. Common shareholders absorb the final uncertainty around refresh, refinancing, and residual value.

The evidence supports an asymmetric conclusion. Scarce upstream nodes and scaled platforms show profit capture. The infrastructure middle shows strong revenue but must still prove equity cash after capital expenditure. Enterprise productivity has company-reported cases and observable payment, not one audited measure of economy-wide value.

The durable winner will not necessarily be the company producing the most tokens or announcing the most capacity. It will be the platform that repeatedly closes four arrows—customer outcome, terminal payment, unit economics, and capital return—while retaining the ability to migrate workloads and reprice through the next hardware cycle.

---

Evidence note: Facts come from the listed primary disclosures. Company-reported outcomes remain labeled as such. Mechanism claims and falsifiable propositions are AI Infra Credit structural inferences. The numerical example is an explicit hypothetical.