Cash, not counted revenue
The test is whether unique users, enterprises and public buyers can fund the whole stack after internal payments are removed.
The Portfolio at a Glance
The following overview summarizes the strategic building blocks, position directions, and allocated assets for this research theme.
Long
Durable production and infrastructure bottlenecks
These positions have scarce capabilities that can remain valuable across several customers and hardware generations.
Long
Embedded workflow ownership
The position tests whether professional data, distribution and workflow integration capture value even if compute becomes less scarce.
Short
Financed compute capacity
These positions face the most direct test of contract quality, utilisation, refinancing and the economic life of accelerators.
Short
The memory-cycle expression
This position tests whether a genuine HBM shortage can sustain current profitability once capacity additions and customer inventories catch up.
The analysis
The investment case for AI infrastructure is often framed as a volume question: how many accelerators, megawatts and halls will be built. That framing is incomplete. The first wave is financed through several balance sheets at once. An application provider pays a model company; the model company pays a cloud provider; the cloud provider signs for equipment and capacity; a data-centre operator or a special-purpose vehicle raises debt against contracts that may ultimately depend on the same payer. Each invoice is valid. They are not independent sources of final demand.
This does not make the build-out fictitious. It makes cash-flow attribution decisive. A high-quality foundry, a grid connection or an industrial control system can be used by several customers and several compute generations. A financier of a concentrated GPU fleet needs a narrower set of assumptions to hold simultaneously: demand must arrive on time, the customer must pay, hardware must retain economic value and the next refinancing must be available at a tolerable rate.
The portfolio therefore avoids a universal view on AI hardware. It favours companies whose scarce capability is useful even if the first rental cohort earns modest returns. The negative positions concentrate on the parts of the stack where the market may be treating contractual volume as fully independent, durable cash generation.
Evidence
The Bank for International Settlements describes the AI infrastructure boom as being financed by borrowing that sits both on corporate balance sheets and outside them. A reported investment programme can therefore be larger than the debt line in the sponsor’s accounts.
Bank for International Settlements, Quarterly Review: financing the AI infrastructure boomBloomberg’s map of AI financing shows companies investing in, buying from and supplying one another. A payment inside that circle is real revenue for the recipient, but it is not a new dollar from an outside customer.
Bloomberg: AI circular dealsJ.P. Morgan treats the data-centre build-out as a capital-markets financing task, with project debt raised alongside borrowing by the operating companies. The existence of the financing does not establish the return on the assets being financed.
J.P. Morgan: financing AI infrastructure and data centres
The Argument
The demand calculation begins after internal transfers are removed
A customer payment can support several reported revenue lines as it moves through an application, a model provider and a cloud platform. Adding those lines produces a useful map of activity, but not a measure of final demand. The relevant economic question is whether the unique external payer base can ultimately support the return required on the capital commissioned across the stack.
This is a discipline, not a forecast of failure. Corporate software budgets, consumer subscriptions, advertising and public-sector use can expand materially. The point is that investment claims should be tested against the cash that leaves those buyers, rather than against every contractual hand-off that follows.
Evidence
Brookings estimates that roughly $9.6 trillion of commissioned AI infrastructure would require about $3.7 trillion of annual revenue by 2032 to earn a 10 percent unlevered return, assuming a 50 percent operating cash-flow margin. That figure is a revenue requirement, not a forecast of unique end-user spending.
Stijn Van Nieuwerburgh, Brookings Institution, September 2026The same author’s data-centre study separates the long-lived building and power connection from the equipment. Lenders to servers and accelerators usually finance them on shorter terms, often within five years, because that hardware depreciates and becomes obsolete faster than the site.
Stijn Van Nieuwerburgh, Columbia Business School: data-centre economicsReuters documented in July 2026 that companies were committing billions of dollars to AI infrastructure. Announced spending measures the build-out. It does not measure how much of that spending end customers have already funded with their own cash.
Reuters, 22 July 2026: companies committing billions to AI infrastructure
Backlog changes character when the payer is also financed by the same build-out
A contract is most valuable when it survives the same funding test as the asset it is meant to support.
Long-term commitments can be economically robust. They are not interchangeable. A contract with an established enterprise customer, a disclosed guarantee and a real operational dependency differs from a chain of commitments among companies that are each extending capacity ahead of their own confirmed cash receipts. The difference is not visible in a single backlog number.
For investors, the useful work is contractual rather than thematic: identify the ultimate payer, the commencement date, termination rights, collateral, renewal logic and the consequences if a provider replaces rather than expands a GPU generation. The more these terms depend on one another, the less safely they should be capitalised as separate demand.
Evidence
S&P Global Ratings lowered Oracle to BBB− and brought roughly $260 billion of future lease commitments and $13 billion of unconditional purchase obligations, mainly for data-centre power, into the credit assessment. The backlog is therefore already being read as a claim on the balance sheet.
S&P Global Ratings: Oracle credit researchReuters reported on 18 September 2026 that about $18 billion of Oracle-linked data-centre debt was under pressure in the market. A large remaining-performance-obligation figure had not, by itself, kept the related debt at par.
Reuters, 18 September 2026: pressure on Oracle-linked data-centre debtReuters described Oracle’s AI infrastructure plan as a high-stakes decision for the credit rating. Cloud revenue can grow while the lease and power commitments still change the risk borne by creditors and shareholders.
Reuters, 4 August 2026: Oracle’s ratings gamble on its AI strategy
Asset-backed finance does not remove residual-value risk
Special-purpose vehicles and asset-backed structures can match capital to a physical project and protect senior lenders through covenants, reserves and contracted payments. They do not settle the economic life of the asset. GPUs can retain useful value for longer than a sceptic expects, but that value depends on technical relevance, power costs, secondary demand and the cost of the next generation.
The distinction is especially important where a site, power connection and building have a much longer life than the first hardware installed inside them. A lender may have recourse to several assets; an equity holder still bears the outcome when the return on the combined project falls below its cost of capital. This is why an infrastructure boom can leave durable sites intact while repricing the most leveraged layer.
Evidence
Meta’s annual report for 2025 is the company’s own account of its data-centre commitments, leases and guarantees. A rating note can interpret that filing, but it does not replace it.
Meta Platforms, annual report for the year ended 31 December 2025S&P Global Ratings describes Meta’s Hyperion structure as one in which a residual-value guarantee moves part of the GPU risk back to Meta. The guarantee does not transfer every operating and market risk away from the project.
S&P Global Ratings: AI data-centre financing, including HyperionMorgan Lewis describes the prevailing project-finance structure as a ring-fenced vehicle. The building, power connection and offtake contracts support a different claim from the short-lived equipment installed inside the site.
Morgan Lewis, February 2026: project finance in the data-centre sectorClifford Chance puts the point directly for lenders: the data-centre lease is the instrument that carries the credit, because repayment follows the contracted cash flow rather than the resale value of the hardware.
Clifford Chance, July 2026: the data-centre lease is the bond
Own scarce capability; treat financed capacity as a claim on a sequence
Taiwan Semiconductor Manufacturing combines process qualification, yield learning and customer relationships that remain necessary across hardware cycles. Hitachi supplies industrial systems and grid-related capability where the customer set extends far beyond one AI programme. RELX owns embedded professional workflows whose value rests on data, distribution and customer habits rather than on rented compute.
CoreWeave, Oracle and SK hynix are not equivalent shorts and should not be treated as a blanket rejection of AI demand. They represent different tests: financing and utilisation for rental capacity, credit and off-balance-sheet commitments for a platform buyer, and the duration of a memory upcycle priced for unusually strong accelerator demand. Each position requires defined risk and a separate evidence standard.
Evidence
CoreWeave closed an $8.5 billion financing facility and described it as the first investment-grade-rated financing backed by GPUs. The rating applies to that facility. It is not a rating of the equity.
CoreWeave: $8.5 billion GPU-backed financing facilityCoreWeave’s quarterly report for the period ended 30 June 2026 is the primary filing for its debt, customers and the conversion of contracted capacity into cash. The press release for a facility does not answer those questions.
CoreWeave, quarterly report for the quarter ended 30 June 2026Fitch Ratings assigned an A (sf) rating, with a stable outlook, to loans at CoreWeave Compute Acquisition Co. VIII. A financed pool of GPUs can therefore carry an investment-grade credit opinion while the equity remains a separate, riskier claim.
Fitch Ratings, 1 April 2026: CoreWeave Compute Acquisition Co. VIII loansReuters reported on 10 September 2026 that Chinese AI-chip makers were raising prices because high-bandwidth memory remained scarce. The shortage is observable. It can also reverse when supply and customer inventories catch up, which is why the memory position is cyclical.
Reuters, 10 September 2026: high-bandwidth-memory shortage and higher chip prices
Current Market Valuation
The likely regime is uneven monetisation. AI capacity keeps growing, yet the return on the first projects spreads unevenly between the parties that own scarce production, power access, workflow distribution and financing capacity. Equity markets can initially value all of them as a single build-out trade. Credit terms, utilisation data and replacement cycles eventually force a distinction.
This does not require a collapse in AI spending. It requires a more ordinary capital cycle: some projects clear their cost of capital, some contracts are refinanced at lower returns and some equipment earns less once supply broadens. The durable long positions should be assessed on their own industrial and workflow economics, not on the continuation of a single accelerator shortage.
Evidence
The European Central Bank’s May 2026 Financial Stability Review treats private credit as a stability question because a loss can reach insurers and pension funds even where their direct exposure is limited.
European Central Bank, Financial Stability Review, May 2026In a special feature of that review, the ECB reports that euro-area pension funds held 4.4 percent of their assets in private credit in 2024, up from 3.3 percent in 2020, and that 79 percent of euro-area private-credit funds were closed-ended. Neither figure is an AI-specific exposure.
European Central Bank, Financial Stability Review special feature on private credit, May 2026The IMF’s April 2026 Global Financial Stability Report examines how losses in private credit can spread to banks, insurers and traded markets. It is a study of transmission, not a forecast that AI borrowers will default.
International Monetary Fund, Global Financial Stability Report, April 2026, chapter 1
The Investment Selection
The selection separates durable bottlenecks and workflow ownership from financed compute capacity and the most cycle-sensitive component exposure. Company ratings on the cards are independent Leeway assessments; the position notes identify the evidence needed for the thesis and are not personalised investment advice.
The Business-Rating scores the quality of the business model, independent of the share price. The Market-Fit-Rating tests eighteen fundamental figures for how well the company currently fits the market; a negative reading implies expected negative performance. The Cycle-Rating places the valuation in the stock’s own history: a higher figure means the shares are historically cheaper. The Leeway-Score combines the three in equal parts. How the ratings are calculated
The recommendations since publication
Not enough trading sessions have elapsed since publication to compute a consistent performance series.
Durable production and infrastructure bottlenecks
LongThese positions have scarce capabilities that can remain valuable across several customers and hardware generations.
Role in the thesis: Long exposure to advanced process manufacturing and qualification that remains necessary regardless of which cloud, model or rental provider earns the highest return.
Investment case: Yield learning, customer trust and scarce leading-edge capacity are durable constraints. The company can benefit from AI demand without relying on one financed compute operator.
Position and invalidation
Position: Treat TSMC as the manufacturing bottleneck in a diversified AI allocation, not as a hedge against all semiconductor cyclicality.
What to watch: Monitor leading-edge utilisation, customer concentration, capital-expenditure discipline and the pace of advanced-node capacity additions.
What would invalidate it: The case weakens if leading-edge utilisation falls for several quarters while pricing and return on invested capital deteriorate.
Principal risk: A technology-cycle correction, customer inventory adjustment, geopolitical risk or accelerated competitive progress could outweigh the relative-quality argument.
Role in the thesis: Long exposure to industrial systems, electrification and project-delivery capability that benefit from grid and data-centre investment without depending on the first GPU-rental cohort.
Investment case: The business addresses multiple infrastructure budgets. Its value lies in installed relationships and execution, rather than in the residual value of one generation of compute hardware.
Position and invalidation
Position: Use Hitachi as a diversified industrial-capability position, not as a pure data-centre proxy.
What to watch: Monitor order quality, margin conversion, grid and digital-infrastructure demand, and working-capital discipline.
What would invalidate it: The case weakens if order growth depends increasingly on low-margin projects or if cash conversion deteriorates as the backlog grows.
Principal risk: Industrial spending, project execution, Japanese currency movements and a valuation that already reflects strong demand can all limit returns.
Your knowledge edge
Understand the real drivers of our portfolio
Look behind the scenes: see not only what we buy, but why. Get the complete view of our strategy.
- Every position revealed: Access each holding in our portfolio.
- Understand the investment thesis: See the exact role each security plays in our return.
- See the connections: Analyse our long and short strategies in direct context.
Three Possible Outcomes
| Path | Weight | What happens |
|---|---|---|
| Base: demand proves real, returns separate | 50% | AI usage expands, but contracts and utilisation differentiate. Foundry, power and workflow positions retain support; some financed capacity reprices as the cost of credit and hardware replacement becomes visible. |
| Favourable: cash conversion arrives early | 25% | Enterprise and consumer spending grows quickly enough to support capacity commitments. Contract commencements, utilisation and refinancing all improve, reducing the case for negative positions and extending the HBM cycle. |
| Adverse: a financing test exposes overlap | 25% | Customers delay or resize commitments, older hardware rental economics soften and lenders demand wider spreads or stronger guarantees. The infrastructure continues to operate, but equity in the most leveraged layer absorbs the reset. |
Counter-arguments and Risks
The strongest counter-case is straightforward: AI demand may compound fast enough that current capacity remains scarce, contracted revenue converts cleanly to cash and hardware retains resale value longer than sceptics expect. In that setting, the financing structures are an efficient way to build useful infrastructure, not evidence of excess.
There is also a company-specific objection. Oracle has a large installed software base and can finance a cloud expansion through more than the AI contracts examined here. CoreWeave may diversify its customers and prove that specialised GPU operations earn durable returns. A disciplined view must test those outcomes against reported cash conversion rather than dismiss them in advance.
Evidence
Oracle reported triple-digit growth in cloud-infrastructure revenue for its first quarter. That result shows demand for the service. It does not show that the associated leases and power commitments earn their cost of capital.
Oracle: first-quarter results and cloud-infrastructure revenueReuters reported on 17 September 2026 that Nebius raised AI-cloud prices again because demand for computing power remained strong. Rising rental prices are evidence against an immediate glut in rented compute.
Reuters, 17 September 2026: Nebius raises AI-cloud prices again
Unresolved Market Factors
The central uncertainty is timing. Capacity is committed years before the cash profile of a new AI workload is fully observable. A project can therefore be rational for a customer and still deliver a mediocre return for the equity investor who financed it at the wrong point in the cycle.
The second uncertainty is asset life. Accelerators can be reused, repurposed or sold, and data-centre facilities can outlive the first hardware cohort. The analysis is not a claim that every GPU becomes worthless. It asks whether the assumptions used in financing and valuation remain prudent when utilisation, replacement cycles and credit availability are no longer uniformly favourable.
Evidence
Moody’s frames the constraint as power delivery, not only as a shortage of chips. A data centre that cannot secure electricity does not turn a financed fleet of accelerators into cash flow.
Moody’s: power without deliveryQuinn Emanuel’s note sets out the contractual and litigation risks in AI data-centre finance, including residual-value guarantees and lease obligations that rating agencies may add back when they do not appear as ordinary debt.
Quinn Emanuel: AI data-centre financing and litigation risksFitch Ratings wrote on 21 September 2026 that power constraints may push more data centres behind the meter. That arrangement adds complexity to the project. It does not remove the need for a power contract.
Fitch Ratings, 21 September 2026: power constraints and behind-the-meter data centres
The Evidence That Matters
- Deduplicated evidence of end-customer spending across enterprise software, consumer services, advertising and public-sector deployments.
- The identity, credit quality, commencement dates and cancellation rights of the counterparties behind major capacity commitments.
- Utilisation, realised rental rates and renewal terms for current and prior GPU generations.
- Credit spreads, covenant changes, equity raises and the price of refinancing for asset-backed compute structures.
- Capital expenditure, lease commitments and free-cash-flow conversion at major cloud and platform buyers.
- HBM pricing, inventories, supplier capacity additions and the pace at which accelerator demand reaches end deployment.
- Grid-connection queues, power-contract terms and the share of data-centre investment that is reusable across hardware generations.
What Would Change the View
The thesis should be reduced or abandoned if these developments persist.
- Independent end-customer spending grows fast enough to support the combined infrastructure return without relying on overlapping internal revenue claims.
- Financed compute providers convert diverse, long-term customer commitments into sustained operating cash flow while refinancing costs decline.
- Older accelerator fleets retain high utilisation and economic rental value through the next major hardware transition.
- The selected durable positions lose pricing, utilisation or workflow retention even as the broader AI infrastructure cycle remains healthy.
How this analysis is produced
The assessment is produced in several steps. Independent model families answer the same question separately and then attack the results. What you read here has survived several rounds.
- Two independent first theses. The same opening question goes to several model families that cannot see one another. Disagreements are kept, not averaged away.
- Dated evidence. Every claim that depends on facts is broken into individual search questions and answered with dated, sourced web research. Question, answer, sources and timestamp are logged and remain traceable.
- Adversarial review. Several review roles attack the thesis from different angles: one hunts for the strongest refutation, one for the awkward edge cases, one tests whether a path from thesis to share price exists at all, one checks the timeline for contradictions. Each role raises its own questions, which are again answered with evidence.
- Merge, then the next round. The surviving theses are merged into one and attacked again. The counter-position and the unresolved tension on this page come out of that step. They were not bolted on afterwards to look balanced.
- Back to the start. The process runs again until there is a clear result and a list of tradable companies with structural advantages.
Any analysis can be wrong. That is why the falsification criteria and the counter-position sit on the same page as the thesis, not in the small print.
The numbers shown against individual companies do not come from this process. The Business-Rating scores business-model quality, the Market-Fit-Rating eighteen fundamental figures against the current market, the Cycle-Rating the valuation against the stock’s own history. They are documented under the Leeway scores.
Questions Investors Ask
Does this thesis predict that AI spending will collapse?
No. It distinguishes useful and growing AI demand from the return earned by each balance sheet financing the first capacity wave. Spending can rise while returns separate sharply between foundries, infrastructure owners, workflow providers and financed rental capacity.
Why can reported AI revenue overstate final demand?
The same end-customer payment can appear as revenue at several layers when an application provider pays a model provider and the model provider pays a cloud host. Those payments are real, but they should not be added together when assessing the cash available to support all installed capital.
What would make the short positions wrong?
Sustained cash conversion, diversified counterparties, lower refinancing costs and durable economics for older hardware would materially weaken the case. The positions concern specific financing and cycle risks, not a general judgement that the companies cannot execute.
Evaluate the selected companies with the three Leeway ratings
The Business Rating assesses business-model quality, the Market-Fit Rating evaluates fundamentals in the current market regime, and the Cycle Rating places valuation in the company’s history. Use Leeway’s general equity analysis to examine each selected company.
Company Valuation and Fundamental Analysis
The data is recalculated on a weekly basis and depends on the current market value of the company and the balance sheet figures of the annual financial statements. The market value changes continuously with price changes, the balance sheets are created annually and change the valuation massively. The time of the annual financial statements and the metrics used can be viewed under "Metrics". Further information on how the analyses work can be found as tooltips directly on the analyses as well as in our explanations.
General
The information is intended to support you in your independent decision-making in implementing investment strategies and serves purely informational purposes. Past performance is not a reliable indication of future performance. No warranty can be given for the completeness, correctness and accuracy of the listed content. The information does not constitute specific investment recommendations. We neither know you nor your financial situation and do not provide investment advice. Only licensed investment advisors with knowledge of your personal circumstances may do this. PWP Leeway UG (limited liability) is not an investment advisor and does not collect any personal data for the purpose of investment optimization. PWP Leeway UG (limited liability) is a provider of investment recommendations and investment strategy recommendations. As such, it is registered with and supervised by the Federal Financial Supervisory Authority (BaFin). The legal notices on the use of the website and the General Terms and Conditions of PWP Leeway UG (limited liability) also apply.
All price data are closing prices of the respective stock exchanges. Price information and master data are provided by an external service provider. Furthermore, public trading data, such as provided by Finra.org, is used to analyze market behavior. No warranty can be given for the completeness, correctness and accuracy of the listed content.
Disclosure of interests: To the best of our knowledge, Leeway and the persons responsible for creating or publishing the respective content do not hold a net long or net short position exceeding 0.5% of the issued share capital of any issuer discussed in that content at the time of publication. There are no other interests or business relationships that could conflict with objectivity, unless expressly stated otherwise in the respective content.
