Published November 18, 2026 · Financial and regional evidence through October 25, 2026. Scenarios are illustrations. Sources and assumptions

Strategy · Infrastructure · Economic risk

Who Pays for the AI Flywheel

Follow the AI trade from silicon and power to customers, cash and consequences.

A brass flywheel connects silicon, power and data centers to a small customer workshop.
Editorial illustration · AI-generatedThe mechanism needs an outside customer. Financing alone does not close the economic loop.
Explore this essay · 15 sections

An investor funds an AI company. The AI company buys computing from the investor. The investor reports growing cloud demand, orders more infrastructure, and gains another reason to invest. Somewhere outside this circle, a customer is supposed to receive something valuable enough to keep paying for it.

That last transaction is the one that matters most.

The circular AI economy is neither imaginary nor, by itself, evidence of fraud. It is an unusually concentrated system in which investors, suppliers, distributors, and competitors frequently occupy the same corporate balance sheet. Its productive possibility is extraordinary: finance builds capacity, capacity improves products, and useful products generate cash that supports further investment. Its vulnerability appears when the evidence for the next round increasingly consists of commitments financed by the previous one.

My conclusion is conditional. AI demand is real; durable returns on every layer of the current investment are not established. A period of consolidation, repricing, and uneven losses is easier to defend than a confident prediction of either uninterrupted acceleration or an economy-wide collapse. The distinction turns on cash generation, contractual rigidity, and who ultimately bears the losses.

One transaction can pass through several revenue lines
Independent customer budgetsUseful work → paid renewal
  1. AI productCustomer revenue
  2. Cloud infrastructureThe product’s compute bill
  3. Chips and powerThe cloud’s suppliers
Some suppliers and cloud providers also invest in AI developers

Equity financing supports the chain. Customer cash must eventually support its economics. Arrows show relationships, not measured transaction sizes.

Follow the money beyond the circle

In January 2025, FTC staff documented cloud agreements requiring AI developers to spend a substantial portion of their partners’ investments on those partners’ services. The report examined Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic. It also described switching costs and overlapping competitive interests. Those findings concern arrangements examined through 2024; they establish the mechanism, not the current terms of every partnership. FTC staff study

Later announcements make the financial interdependence tangible. OpenAI’s February 2026 Amazon agreement paired an announced $50 billion investment with an expansion of AWS commitments by $100 billion over eight years. Anthropic’s April agreement committed more than $100 billion over ten years to AWS technologies, alongside a new Amazon investment and possible future investment. These are different kinds of commitments over different periods, not interchangeable piles of cash. OpenAI and Amazon, Anthropic and Amazon

NVIDIA’s September 2025 announcement was a letter of intent: up to $100 billion of investment as infrastructure is deployed. Treating that ceiling as money already received would manufacture liquidity on paper. Similarly, a funding-round total can include previously announced commitments: Anthropic explicitly said its May 2026 round included $15 billion of previously committed hyperscaler investments. Adding those announcements together again would count the same financing twice. NVIDIA partnership announcement, Anthropic Series H

There is a second counting problem. A customer’s payment can become a lab’s revenue, then a cloud provider’s revenue, then a chip supplier’s revenue. Each transaction can be legitimate. Their sum is not an equivalent amount of new final demand or economic value added. Nor is investment income another customer purchase. Equity financing, cloud revenue, backlog, and unrealized investment gains belong in different columns.

The practical test is whether customers with their own operating budgets renew because the product improves their economics. Investor-funded startups can be genuine customers, but their spending is also more exposed to the financing cycle. Independent customer demand is not a perfectly observable binary; it is a question of concentration, funding dependence, and renewal behavior.

One economy · different claims on its gains
↓ Capability is assembled↑ Cash must earn its way back
  1. Physical foundationPower · sites · fabricationConnections, construction, and manufacturing must arrive on time.
  2. The scarce componentsGPUs · memory · networksNVIDIA, Micron, and peers sell capacity before its owner proves a return.
  3. The operating assetNeoclouds · hyperscalersUtilization and collections have to cover finance and replacement.
  4. The intelligence layerModels · agents · softwareUseful outcomes must command more than their full delivery cost.
  5. The independent testCustomers who keep payingInvestor financing can buy time. Renewals must earn it.

The supplier’s ambition
Keep its contribution scarce.

The buyer’s ambition
Make useful work cheaper.

Conceptual dependency map; no arrow width or spacing represents measured financial flows. The same company may occupy several layers. Representative names are exposures, not recommendations.

What the AI trade actually contains

There is no single AI trade. There are linked claims on different cash flows, assets, and bottlenecks. A chip designer sells systems; a memory producer sells bits and bandwidth; a cloud operator rents access; a lab sells intelligence; an application sells a completed workflow. An investor buying one layer is not buying an equal share of the others’ eventual success.

LayerRepresentative exposureWhat must remain trueWhere the thesis can fail
Electricity and sitesGeneration, transmission, powered landDeliverable power and viable customer contracts justify constructionGrid delays, cost allocation, stranded sites, or a weak counterparty
Electrical equipment and coolingVertiv and other infrastructure suppliersDense computing installations convert into delivered equipment ordersProject deferrals, competing designs, or backlog that converts too slowly
Fabrication and packagingTSMC; semiconductor-equipment suppliers such as ASMLCustomers keep funding advanced manufacturing and packagingExport restrictions, geographic concentration, capital cycles, and execution
Accelerators and networksNVIDIA, AMD; Broadcom and Arista in networking or custom-chip ecosystemsSystems earn their price through useful throughput and adoptionCustom silicon, software portability, price pressure, or slower customer capex
Memory and storageMicron, SK hynix, SamsungBandwidth and capacity demand outgrow economically available supplyNew capacity, weaker pricing, qualification losses, or inventory reversal
GPU clouds and neocloudsCoreWeave and other specialized operatorsContracted utilization covers equipment, power, financing, and replacementConcentration, refinancing, service delays, and declining residual values
HyperscalersMicrosoft, Google, Amazon; Oracle in cloud infrastructureScale, distribution, and utilization recover incremental investmentOverbuilding, internal cannibalization, and inadequate returns on capital
Models and applicationsOpenAI, Anthropic, SaaS vendors and agentsAccepted outcomes command more than their total delivery costModel substitution, lower prices, weak retention, and costly verification
Workers and local economiesDigital occupations; housing supported by their compensationProductivity creates enough paid demand and income in the relevant placeReduced hiring, weaker bargaining power, and gains accruing elsewhere

These are analytical exposures, not recommendations or interchangeable business models. Vertiv describes power and cooling infrastructure; Arista describes Ethernet fabrics; TSMC reports advanced-node manufacturing. Those products are complements, but their order cycles and margins need not move together. Vertiv results and business description, Arista AI networking, TSMC third-quarter release

The dependency runs both ways. Physical capacity makes digital work possible; economically valuable digital work must eventually support the physical bill. Adding every company’s revenue along that route does not measure the final market several times over.

A chip order is also not an energized data center. Connections, transformers, cooling, permits, and transmission have to arrive together. The IEA’s 2025 base case projected global data-center electricity demand of around 945 TWh in 2030, more than double its 2024 level; that is a conditional projection for all data centers, not measured future AI-only consumption. Local grid constraints can matter even when national supply appears adequate. My investment inference is that a physical bottleneck can support an equipment supplier’s pricing while delaying a cloud customer’s revenue. For utilities and developers, the question is who pays if the promised load arrives late or never does—not simply how many gigawatts were announced. IEA energy-demand analysis

The GPU and memory trade is not the cloud trade

NVIDIA’s strongest argument is already visible in its income statement. For the quarter ending July 26, 2026, it reported $96.2 billion in revenue, including $89.0 billion from Data Center, and a 75% company gross margin. This is substantial realized business, not merely a promise that someone will eventually buy a chatbot. But a supplier’s sale proves the buyer purchased capacity; it does not prove the buyer will earn its required return on that capacity. NVIDIA second-quarter fiscal 2027 results

The NVIDIA thesis combines accelerators with software, interconnects, systems integration, and an installed developer ecosystem. Its risk is not limited to another chip winning a benchmark. Buyers may slow the next installation, shift selected workloads to custom chips, or demand a lower cost per useful result. Concentrated manufacturing and export restrictions add vulnerabilities that a strong product cannot eliminate. NVIDIA’s filing discusses inventory and purchase-obligation charges as well as China-related restrictions. NVIDIA quarterly filing

Micron expresses a related but different thesis: expensive arithmetic is wasted when the processor cannot obtain data fast enough. High-bandwidth memory helps feed accelerators; other DRAM and storage serve different parts of the system. They should not be combined into one undifferentiated “AI memory” price. Google’s Ironwood design, for example, emphasizes both memory capacity and bandwidth, demonstrating why a custom accelerator still needs a memory subsystem. Google’s Ironwood architecture

Micron’s fiscal 2026 release reported $133.19 billion of annual revenue and an 80.7% GAAP gross margin, against $37.38 billion and 39.8% in the prior year. Those company-wide results show a dramatic change in realized economics; they are not HBM-only figures or evidence that the margin is permanent. Its $27.37 billion of annual net capital investment also shows that the supplier has a capital budget of its own. Micron fiscal 2026 results

My analytical concern is the cycle: high prices make expansion attractive, while the capacity response arrives later. Contract duration, customer qualification, yields, and product mix determine how much protection a producer has when conditions change. A low multiple of unusually high earnings can be less cheap than it looks. In a purely illustrative example, earnings per share falling from 10 to 5 while the multiple falls from 20 to 15 takes a share price from 200 to 75—a 62.5% decline despite continued profitability.

Custom chips complicate the trade rather than simply ending it. Google’s TPUs and Amazon’s Trainium demonstrate that the largest customers are also pursuing alternative computing architectures. Broadcom’s announced Meta collaboration extends that custom-silicon effort. A shift away from merchant GPUs can redistribute revenue among accelerator vendors while preserving demand for fabrication, memory, networking, and power. It does not guarantee that every supplier benefits equally. Amazon’s Trainium discussion, Broadcom and Meta

The buyer wants computation to become cheaper. The supplier wants its own contribution to remain scarce. The boom can serve both for a while; their interests are not identical.

A stone aqueduct carries light through data centers toward productive workshops.
Editorial illustration · AI-generatedUse, revenue and cash recovery are different tests. Each connection has to hold.

Neoclouds turn growth into a credit question

A neocloud is a specialized cloud operator built around accelerated computing. Specialization can shorten deployment, improve workload support, and serve customers a general-purpose cloud does not satisfy. The business can also concentrate equipment, customer, and financing risk in the same company. It should be assessed as an operating and credit structure, not merely assigned a software revenue multiple.

CoreWeave makes the distinction concrete. Its June-quarter release reported $2.575 billion in revenue and $1.510 billion of adjusted EBITDA, but a GAAP operating loss of $49 million and a net loss of $626 million. Depreciation and financing cannot be wished away because an infrastructure company has a high EBITDA margin. Its approximately $104 billion revenue backlog also includes estimated future amounts subject to delivery and service-availability requirements; backlog is neither cash in the bank nor a profit measure. CoreWeave second-quarter results

The corresponding filing states $35.6 billion of total indebtedness and $10.0 billion of undrawn facility availability at June 30. It identifies customer concentration and warns that some general-purpose cloud competitors are also customers. The debt figure is a point-in-time disclosed measure, not a sum of every lease, equipment purchase, and future commitment. CoreWeave quarterly filing

The underwriting questions follow: Who owes the payments? What conditions precede acceptance? Does cash arrive before debt amortizes? Who absorbs a construction delay? What utilization and rental rate are required after the first contract? A hyperscaler can buy specialized capacity now and become a tougher competitor later. That possibility does not invalidate a contract; it changes the reinvestment assumptions.

An older GPU need not become useless when a newer one arrives. It may serve less demanding work at a lower price. But physical usefulness is different from sufficient rental income or collateral recovery. Illustratively, equipment costing 100 and financed with 70 of debt has 30 of initial equity. If recoverable asset value falls to 50 before debt is repaid, equity is exhausted and creditors face a shortfall before other recoveries. Actual loans have different collateral, guarantees, amortization, and recourse. The example is a loss waterfall, not an estimate for CoreWeave or any chip generation.

What the hyperscalers are racing to own

The race is for more than the largest training run. It is for recurring inference demand, customer distribution, scarce infrastructure, and control of the interface through which work is commissioned. Training creates a capability; inference repeatedly spends compute to use it. Agents add potential tool calls, retries, and longer tasks. That can enlarge the compute market, but only if useful output and customer willingness to pay cover the extra work.

Microsoft can sell intelligence through enterprise software and Azure. Google can support Cloud while defending Search and its broader distribution. Amazon can sell computing, host multiple models, and use custom chips to improve its own economics. Meta can monetize better recommendations and advertising without charging a separate subscription for every AI feature. These are distinct strategic routes; an AI revenue line alone will miss some returns, while total corporate growth will overstate what AI has proven.

Amazon is therefore essential even though the four detailed SWOT cases below remain OpenAI, Anthropic, Google, and Microsoft. Its June-quarter release reported $42.2 billion of AWS sales and $16.6 billion of AWS operating income. It also reported negative $7.6 billion of trailing-twelve-month company free cash flow and said the increase in infrastructure purchases primarily reflected AI investment. Those periods and scopes differ; they cannot be divided to produce an AI return. The same release attributed most of $53.4 billion in non-operating pre-tax other income to investments in Anthropic—another reason to distinguish investment marks from operating customer cash. Amazon second-quarter results

Strategy adds a game-theory problem. Being late may endanger a valuable franchise, so spending can be rational for one firm even when simultaneous spending by rivals produces excessive industry capacity. A real option to compete is valuable, but its value is not unlimited. Commitments that cannot be delayed turn an option into an obligation. Management ambitions about general intelligence do not themselves establish the discounted cash flows required to fund them.

The critical unit is consequently shifting from a chip, token, or benchmark score to an accepted outcome: a resolved case, validated analysis, or functioning workflow. The winning system need not use the most expensive model on every step. Routing, caching, smaller models, human review, and better software can lower cost. Whether those savings expand total spending or reduce it depends on demand response. Efficiency is neither automatically bearish for compute nor a guarantee of unlimited consumption.

Four firms competing for different positions

Microsoft is the fourth company in this comparison because its cloud distribution, enterprise software, and OpenAI relationship make the overlap between supplier, investor, and competitor especially clear. Amazon remains indispensable to the economic map, even though it is not a fifth full SWOT case.

OpenAI’s strategic asset is direct access to users and workflows. Its opportunity is to turn broad adoption into persistent, valuable work. Its constraint is that a successful software interface can carry a substantial physical infrastructure bill. Distribution helps acquire customers; it does not establish that the marginal customer produces enough contribution to fund inference, support, research, and future capacity.

Anthropic’s strategic argument emphasizes enterprise use and coding. Reuters reported in October that the company told prospective investors third-quarter revenue reached $16.8 billion and that it was profitable on an adjusted basis. That is consequential evidence against the claim that frontier AI has no commercial demand. It remains a reported private-company presentation, not a complete public cash-flow statement; adjusted profitability does not settle cash generation or comparability with a rival’s revenue accounting. Reuters reporting

OpenAI’s reported September annualized revenue was approaching $50 billion, according to Reuters’ account of Financial Times reporting. Earlier higher figures involved attempts to make its accounting comparable with Anthropic’s. An annualized pace is not a full year’s recognized sales, and gross-versus-net presentation can change a headline without changing the underlying customer transaction. A league table that ignores those differences rewards presentation rather than analysis. Reuters on revenue comparability

Google has a different strategic problem: monetizing and defending a broad existing system. It owns consumer distribution, cloud infrastructure, models, and custom computing capabilities. That integration can lower coordination costs and offer several routes to recovery. It also exposes the company to the cost of defending businesses that a new interface could reshape. Google’s second-quarter filing reported $24.77 billion of Cloud revenue and $8.81 billion of Cloud operating income. Neither is a standalone return on Gemini, and some shared AI research costs sit outside the segment. Alphabet second-quarter results

Microsoft’s advantage is the ability to distribute AI through enterprise purchasing relationships. Its June-quarter results included $90 billion of revenue and $40.6 billion of operating income across the company. These existing earnings provide financing capacity that a standalone lab cannot simply assume. They also create a demanding capital-allocation question: how much additional investment is needed to preserve those economics, and how much earns an attractive incremental return? Microsoft June-quarter results

The alliances are changing. Microsoft’s April agreement made its OpenAI IP license non-exclusive and allowed OpenAI to serve products across cloud providers. OpenAI’s October announcement added Google Cloud distribution. The strategic contest increasingly concerns who owns procurement and the customer relationship, even when several providers sell access to similar models. Microsoft partnership update, OpenAI and Google Cloud

A SWOT that has consequences

The following is an analyst’s strategic assessment, not a company forecast. Strengths and weaknesses describe internal positions; opportunities and threats describe the external environment.

Compare the four companies · full SWOT
CompanyStrengthWeaknessOpportunityThreat
OpenAIDirect consumer reach and developer relationshipsLarge compute needs; limited public visibility into full cash economicsOwn recurring workflows and valuable completed outcomesModel substitution, procurement intermediaries, infrastructure commitments outrunning collections
AnthropicEnterprise and coding position; several cloud distribution channelsDependence on partners and incomplete public financial comparabilityConvert experimentation into embedded production usePrice compression, platform bargaining power, concentrated large customers
GoogleDistribution, cloud, research, and custom silicon in one organizationCoordination complexity and exposure of established search economicsMonetize several layers while defending existing demandAI interfaces changing search behavior; regulatory limits; excess capacity
MicrosoftEnterprise procurement, software distribution, and diversified cash generationRapidly rising infrastructure intensity and complex partner relationshipsBundle trusted AI into existing workflows and earn cloud demandMulti-cloud distribution reducing exclusivity; weak realized returns on new capacity

SWOT becomes useful when it changes a decision. For OpenAI and Anthropic, an attractive response is to earn workflow dependence through reliability, integration, and demonstrated outcomes while negotiating capacity flexibility. For Google and Microsoft, the response is to demand incremental economics from both internal models and partner distribution, rather than treating infrastructure growth as its own justification.

Porter’s Five Forces sharpens this distinction. Chip capacity, power, and specialized infrastructure can give suppliers bargaining power. Large enterprise buyers can bargain, multi-home, and route easier tasks to cheaper systems. Substitute products include smaller specialized models, conventional software, and human work. Entry barriers are immense at the frontier but much lower for a thin application layer. Rivalry can therefore coexist with concentration: a few powerful firms may spend aggressively while still competing away portions of their margins. Harvard’s Five Forces framework

A resource-based analysis reaches a related conclusion. A benchmark lead is valuable, but its duration is uncertain. Distribution, security approvals, useful proprietary workflows, and customer relationships may be harder to displace. None is automatically a permanent moat. The question is whether the organization can capture value from a scarce resource before rivals reproduce it or customers route around it.

The bill arrives in cash before the return is proven

Capital budgeting asks whether the present value of incremental future cash flows exceeds the investment and its financing cost. Revenue growth alone cannot answer that question. Neither can a positive accounting margin, because capital spending and economic depreciation follow different clocks.

The same calendar quarter provides a useful comparison. Microsoft’s fiscal fourth quarter ended June 30, 2026, so it aligns with Alphabet’s and Meta’s second calendar quarters.

June 2026 quarter, $ billionsOperating cash flowCash purchases of property and equipmentDifference, approximately
Microsoft55.435.819.6
Alphabet39.06944.924−5.855
Meta31.86230.1161.746

These are company-wide figures, covering AI and other operations. The last column is operating cash flow less cash property-and-equipment purchases, not a fully standardized measure of residual cash. Meta also subtracts $0.962 billion of finance-lease principal in its own free-cash-flow definition, producing $0.784 billion. Microsoft’s broader $41 billion capex figure includes finance leases and must not be mixed with its $35.8 billion cash purchase figure. Microsoft cash discussion, Alphabet cash-flow reconciliation, Meta cash-flow reconciliation

Meta’s newer third-quarter release makes the tension sharper. It reported $64.60 billion in revenue and $32.79 billion in operating cash flow, but $34.91 billion in capex including lease principal, leaving negative $2.12 billion of company-defined free cash flow. Its annual capex outlook was $135–145 billion. The operating-income decline also included a $9.89 billion litigation charge, which prevents treating the entire decline as an AI expense. Meta’s remaining cash and marketable securities were substantial. The evidence describes investment pressure, not imminent insolvency. Meta third-quarter results

It is therefore too crude to say Meta has built infrastructure without income. Its existing advertising business generates considerable cash, and AI can improve recommendations and monetization without appearing as a separate AI subscription. What remains unresolved is the incremental return on the next dollar of infrastructure. Management’s attribution of better business results to AI is relevant; it is not a disclosed causal estimate of project-level return on invested capital.

The right unit of analysis also changes. A cheaper token is valuable only insofar as it lowers the cost of a useful completed outcome. Reliability failures, repeated attempts, human review, tool execution, and support can absorb the apparent saving. At enterprise scale, the denominator should approach an accepted result, not merely a model response.

The same quarter · June 30, 2026 · US$ billions
Operating cash flowCash property and equipment

Microsoft

55.40
35.80

CFO less cash PP&E
≈19.6B

Alphabet

39.07
44.92

CFO less cash PP&E
−5.86B

Meta

31.86
30.12

CFO less cash PP&E
1.75B

Company-wide cash flows, not isolated AI returns. Meta’s reported free cash flow also subtracts lease principal. Definitions and sources in the text.

What an extraordinary valuation requires

A strong company and an attractive security are separate propositions. NVIDIA can remain a leading supplier while its shares disappoint if the purchase price requires even more. Micron can post record earnings while investors overpay for the portion that will prove cyclical. A lab can transform work while issuing so much new equity to finance that transformation that early ownership is diluted.

Anthropic’s May financing announcement placed its post-money equity value at $965 billion. That is a financing-round equity valuation, not enterprise value, an executable public quote, or a market price for all existing share classes. Its associated growth claims and the later reported quarterly revenues should not be converted into a precise like-for-like multiple against a public company without reconciling accounting and capitalization. Anthropic financing announcement

Reverse valuation is more informative than calling a large number frightening. Consider a hypothetical enterprise worth $1 trillion. If sustainable free cash flow begins with a payment at the end of year one, grows forever at 3%, and is discounted at 10%, the first annual cash flow required is $70 billion. If there is no distributable free cash flow in years one through five, the required year-six cash flow rises to approximately $112.7 billion. At a 20% free-cash-flow margin, that implies about $563.7 billion of year-six revenue. These are deliberately simplified assumptions, not a valuation of Anthropic or a forecast for any company.

The arithmetic exposes the bargain: a high valuation asks for some combination of enormous scale, durable margins, a short wait, or a low required return. Assume more competition, greater reinvestment, or a longer delay and the supported value falls. Free cash flow here is after the investment needed to sustain growth, not EBITDA with replacement equipment ignored. Enterprise value must also be bridged to equity through debt, cash, other claims, and dilution.

What a hypothetical $1 trillion enterprise value asks of the future

A lower margin means a much larger business must arrive.

Enterprise value today$1T
Required year-six FCF$112.7B
Implied year-six revenue$563.7B

Illustration, not a company valuation. No distributable FCF in years 1–5; first annual FCF at the end of year 6, growing 3% forever. EV = FCF₆ ÷ (r − g) ÷ (1 + r)⁵. FCF is after reinvestment. Debt, cash, other claims, and dilution would be needed to bridge enterprise value to per-share equity value.

The dot-com comparison that actually helps

The useful historical analogy has two parts: internet-company valuations and the telecommunications capacity built to serve them. An NBER study of the dot-com episode examined prices relative to fundamentals and the role of optimistic investors and selling constraints. A Richmond Fed study documented telecom overinvestment while arguing that mistaken forecasts of fundamentals could explain the episode without its preferred definition of a speculative bubble. The historical diagnosis itself is contested. Ofek and Richardson on dot-com prices, Couper, Hejkal, and Wolman on telecom

The lesson is narrower and stronger than “AI is 1999 again.” Real adoption does not guarantee that every capacity owner earns an adequate return at the price it paid. Demand forecasts can be directionally correct and still too early, too concentrated, or too generous about margins. Cheap capacity after restructuring can benefit downstream users while original equity holders suffer.

There are also material differences. Today’s largest buyers include diversified, profitable incumbents; the record already contains substantial realized revenue. AI hardware faces a different replacement cycle from fiber or buildings. Buildings, grid connections, GPUs, and model research do not share one salvage value or useful life. Treating the whole complex as either worthless speculation or permanent infrastructure conceals where losses would actually fall.

The technology can win, the customer can benefit, and the investor can still lose.

How a flywheel becomes fragile

Consider an illustrative business with revenue indexed to 100. After variable serving costs, it retains a 40% contribution margin: 40 to cover fixed cash obligations of 50. It needs 10 of additional financing. If revenue falls to 80 and contribution margin falls to 30%, contribution becomes 24 and the financing gap becomes 26. The gap has more than doubled although revenue fell only 20%.

This is a sensitivity example, not an estimate of any company’s undisclosed finances. Its point is operating leverage. Infrastructure contracts with minimum commitments, research payroll, and debt service can adjust more slowly than prices or demand. Other contracts may have flexibility. A serious review must establish which obligations are actually fixed rather than assuming every announced dollar is irrevocable.

The circular system can amplify the same shock. Lower expected demand reduces a lab’s financing valuation. Financing becomes harder just as supplier bills arrive. Weaker purchases affect cloud utilization and supplier orders. Lower expectations then affect equity prices and collateral values. If outside customers continue generating adequate contribution, the loop can stabilize. If financing has substituted for that contribution, each participant’s apparent strength depends more heavily on the others’ confidence.

There is no general law requiring acceleration until a sudden break. Capacity can be delayed, workloads shifted, contracts renegotiated, or projects abandoned. Those adjustments may be painful and expensive, but gradual repricing and consolidation are possible outcomes too.

Try the financing gap · illustrative index units

What happens when revenue and contribution margin move together?

Contribution40
Fixed cash obligations50
Financing gap10

Revenue × contribution margin − fixed obligations. Teaching model only; these are not estimates for any company. No growth, renewal, or borrowing assumptions are hidden in the calculation.

The strongest case on each side

The bullish argument is that scarce capacity is restricting an already valuable general-purpose technology. Cheaper inference can open new uses; enterprise adoption can compound as software is redesigned around it; and existing advertising, search, and productivity franchises can monetize improvements indirectly. Building infrastructure ahead of fully visible demand can be rational when construction takes years and being late is costly.

The bearish argument is that competition and technical improvement may send much of the benefit to customers while expensive infrastructure retains its original purchase price. Lower unit costs do not guarantee higher supplier profit. More usage may coexist with falling revenue per task, lower utilization of older hardware, or greater spending needed merely to stay competitive. The technology can succeed while a particular financing structure fails to earn its cost of capital.

Both arguments can be true in different layers. An application can create useful work, a cloud platform can earn healthy margins, a lab can struggle to capture enough of the value, and a leveraged infrastructure project can lose money. “AI is a bubble” is too coarse to identify who is exposed.

My assessment is that parts of the financing system display bubble-like vulnerabilities: expensive expectations, overlapping financial dependencies, and investment that can outrun the cash available to support it. That does not establish that every AI business is mispriced. A valuation verdict ultimately depends on future cash flows and the price paid for them; capex intensity alone cannot settle it. The strongest concern is the financing gap, particularly where it meets leverage and inflexible obligations.

Possible pathWhat would support itWhat would weaken it
Productive expansionRenewals from operating budgets; improving cost per accepted task; capacity used at adequate marginsGrowth sustained mainly by subsidies, credits, or financially dependent customers
Consolidation and repricingStrong usage alongside pricing pressure, delayed capacity, and weaker firms seeking partnersBroad, durable margin improvement across model and infrastructure providers
Financial contagionDefaults, guarantees being called, leveraged lenders impaired, or funding withdrawals forcing asset salesLosses absorbed by long-term equity without disrupting credit or essential spending

These are conditional scenarios, not calibrated probabilities. My working view favors continued adoption accompanied by a harder contest over margins, capital discipline, and consolidation. That view should change if independently financed demand and cash recovery improve—or if financing stress propagates into credit.

When digital labor gets cheaper, SaaS changes too

The downstream thesis is more disruptive than another chatbot subscription. Some digital tasks become easier to reproduce: a first draft, a routine query, a translation, a code scaffold. Lower production cost can push prices down and expand supply. It does not make every result interchangeable. Verification, liability, privileged data, trusted distribution, and the ability to complete a messy real workflow can remain scarce.

This creates two plausible SaaS paths. In one, AI weakens the interface layer: an agent operates across applications, fewer employees need paid seats, and small teams build substitutes for narrow tools. In the other, the incumbent becomes more valuable because it holds the records, permissions, integrations, and audit trail an agent needs. An attractive demonstration is not evidence that a regulated system of record can be cheaply replaced.

Seat pricing is especially exposed when software promises to reduce the headcount that purchases those seats. Vendors can respond by charging for actions, consumption, outcomes, or platform access. Salesforce’s published Agentforce pricing already combines several such approaches; that is evidence of a changing billing architecture, not proof that the transition will preserve every vendor’s margins. Salesforce Agentforce pricing

An illustrative support department shows why the effects must be separated. A firm might resolve more cases with fewer staff, spend less on seat licenses, and spend more on AI usage and verification. Its own profit can rise while the displaced worker’s income falls and the software supplier’s revenue shifts. Alternatively, lower cost can make previously uneconomic customers serviceable and increase hiring elsewhere. These are competing mechanisms, not a measured prediction for the whole economy.

Evidence about jobs is already more nuanced than either denial or apocalypse. Stanford’s revised August 2026 payroll study found employment among workers aged 22–25 in exposed occupations 19% below a trajectory keeping pace with less-exposed peers. That is a relative employment gap, not a claim that 19% were laid off; adjustment was concentrated in reduced hiring. The authors describe early descriptive indicators, not causal estimates, and note sample and specification limitations. Yale’s October tracking analysis, using a different approach, did not yet find a clear aggregate labor-market footprint aligned with AI exposure. These results can coexist: a concentrated entry-level problem need not immediately appear as economy-wide displacement. Stanford employment study, Yale labor-market tracking

The distribution matters for the investment thesis. If savings accrue mainly to owners and buyers, productivity growth need not reproduce the same payroll, career ladder, or local spending. If cheaper capabilities enable new products and firms, employment can expand in different roles. The relevant measures are new hiring, hours, contractor income, wages, occupational transitions, and verified adoption—not every layoff an executive mentions alongside AI.

What should everyone else fear

A market correction, a technology-sector recession, and a systemic financial crisis are different events. A crisis requires transmission: leverage, vulnerable lenders, liquidity demands, or losses large enough to impair broader spending and credit. The Federal Reserve’s May 2026 stability report recorded market contacts’ concerns about AI valuations, debt-funded investment, and labor disruption. It was a survey of concerns, not a probability estimate or a declaration that collapse was inevitable. Federal Reserve risk survey

The practical exposure is often cumulative. Someone can depend on a technology salary, receive compensation in technology shares, own concentrated investments, and carry an expensive mortgage in the same regional economy. A reversal can reach all four positions without the national financial system collapsing. A small business can face a different exposure: depending on one subsidized AI supplier whose prices or terms later change.

The appropriate response is to identify those dependencies, preserve liquidity appropriate to existing obligations, test alternative suppliers, and avoid confusing a rising valuation with durable operating strength. For corporate boards, the equivalent discipline is to stage capital, disclose commitments clearly, and evaluate whether each increment of investment earns its return.

The most revealing measures will be paid renewals, customer concentration, contribution after actual serving costs, utilization, cash recovery, and the flexibility of future commitments. A large backlog matters; the creditworthiness and independence of the customers behind it matter too.

The transmission to Kirkland has missing links
Corporate AI gainsRevenue and productivity can rise.

Local work expands

More well-paid roles or stronger compensation can increase purchasing power.

Work becomes less labor-intensive

Profit can rise while headcount, entry roles, or seat demand weaken.

Neither route settles the housing questionWho receives the income or equity—and do they buy here?

Buyer budgets

Available homes

Actual clearing prices

Scenario mechanisms, not estimated effects or probabilities. Existing owners’ ability to wait can delay adjustment. A location premium does not establish a floor under a purchase price.

Follow the trade all the way to Kirkland

This is where a global investment thesis becomes a local housing question. Corporate profit, shareholder wealth, local wages, and homebuying power are connected, but they are not the same variable. A hyperscaler can report stronger earnings while employing fewer people to produce them. A chip supplier can prosper without adding its workforce in Kirkland. A local worker can retain a job while losing stock-based purchasing power.

The supportive chain is plausible: durable cloud and software earnings support well-paid regional work; compensation and accumulated equity support bids for scarce desirable homes. But the chain needs every link. “AI spending is growing” does not establish that the next Kirkland buyer’s salary or down payment is growing. Nor does an office location establish where remote workers live, spend, or pay tax.

The downside has two distinct routes. AI disappointment can reduce equity wealth, infrastructure orders, and hiring. AI success with labor substitution can increase corporate margins without preserving the same number of high-paid local buyers. Neither mechanically produces a housing crash. Existing owners’ fixed mortgages, equity, and willingness to wait can slow the adjustment. Both routes challenge the claim that corporate AI success automatically validates an expensive entry price.

Our retained public Kirkland data already show why level and direction must be separated: the July–September 2026 all-residential median remained around $1.22 million while seller conditions weakened across several measures. That is a mix-sensitive city snapshot, not a valuation of a specific house. The companion essay examines the migration, jobs, and financing evidence in detail. Kirkland public market data

A defensible location premium is therefore different from a defensible purchase. The latter needs a carrying-cost budget, an alternative-rent comparison, adequate holding capacity, and assumptions about income and resale that can survive disappointment. Without property-specific rents and costs, it would be false precision to announce a fair price or a date when prices must break.

The complete AI thesis ends with distribution: who gets paid, who keeps the margin, who carries the debt, and where that purchasing power settles. That is the bridge between a semiconductor earnings release and an Eastside mortgage.

AI does not need universal investor success to transform the economy. It needs useful work that people and organizations can keep paying for. The central question is how far the financing system can run ahead of that work—and who remains responsible when the distance stops growing.

Sources and further reading · 37
  1. FTC staff study
  2. OpenAI and Amazon
  3. Anthropic and Amazon
  4. NVIDIA partnership announcement
  5. Anthropic Series H
  6. Vertiv results and business description
  7. Arista AI networking
  8. TSMC third-quarter release
  9. IEA energy-demand analysis
  10. NVIDIA second-quarter fiscal 2027 results
  11. NVIDIA quarterly filing
  12. Google’s Ironwood architecture
  13. Micron fiscal 2026 results
  14. Amazon’s Trainium discussion
  15. Broadcom and Meta
  16. CoreWeave second-quarter results
  17. CoreWeave quarterly filing
  18. Amazon second-quarter results
  19. Reuters reporting
  20. Reuters on revenue comparability
  21. Alphabet second-quarter results
  22. Microsoft June-quarter results
  23. Microsoft partnership update
  24. OpenAI and Google Cloud
  25. Harvard’s Five Forces framework
  26. Microsoft cash discussion
  27. Alphabet cash-flow reconciliation
  28. Meta cash-flow reconciliation
  29. Meta third-quarter results
  30. Anthropic financing announcement
  31. Ofek and Richardson on dot-com prices
  32. Couper, Hejkal, and Wolman on telecom
  33. Salesforce Agentforce pricing
  34. Stanford employment study
  35. Yale labor-market tracking
  36. Federal Reserve risk survey
  37. Kirkland public market data
Back to the beginning ↑

Cite and share

Vale, Thaddeus. “Who Pays for the AI Flywheel.” ThaddeusVale.com, November 18, 2026. https://thaddeusvale.com/ideas/who-pays-for-the-ai-flywheel

Structured article (JSON)