Six alternative branches. These are subjective probabilities from the seed, not observed frequencies or investment recommendations.
Why this trend
It is the transmission channel for everything else in the sector. Capability, adoption and pricing all resolve into whether roughly $8.2T of projected 2026-2032 US data-centre investment, financed at 70-80% asset-level leverage with ~$660B of hyperscaler lease commitments off balance sheet, can service itself. The FOMC named this structure a financial-stability risk; no other trend in this sector reaches the macro brief.
Load-bearing assumption
Accelerator-bearing data-centre assets retain enough residual value at year five to support the 70-80% asset-level leverage at which they were financed.
Preconditions
- Capability
clear — the assets exist and are utilised
- Economics
fails — no frontier lab discloses a gross margin and no independent price-per-token index covering 2026 exists
- Supply
partial — accelerators are available; interconnection, turbines and transformers are not
- Demand
partial — contracted backlog is real ($104B at CoreWeave) but concentrated in a handful of counterparties
- Permission
fails — FERC Section 206 dockets EL26-67 to EL26-72 are open and large-load rules are unwritten
- Capital
clear for now — ~$300B of AI-related IG issuance projected for 2026, but at a 3.50-3.75% policy rate with three FOMC dissents in favour of a hike
Structured finance digests
Capex growth decelerates from 2027 as public markets keep treating guidance raises as a cost rather than a signal, but the debt is serviced. Spreads on data-centre ABS widen modestly, lease commitments keep growing more slowly than in 2026, and two to four neoclouds are absorbed into hyperscalers or larger neoclouds rather than defaulting. Capacity gets absorbed roughly on schedule and the residual-value question is deferred rather than answered.
Mechanism
Hyperscalers moderate the rate of commitment in response to equity-market punishment, which slows the rate at which new leverage is added. Existing contracted backlog converts to revenue fast enough to cover interest. Prior-generation accelerators continue to find tenants, as CoreWeave's A100 contract running to 2029 suggests, so depreciation schedules hold and lenders do not re-underwrite residuals.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- Big-4 aggregate capex guidance for 2027 comes in below 1.3x the 2026 figure
- Cloud revenue growth at the four largest buyers stays above 25%
- No hyperscaler takes an impairment on data-centre assets
- Early Indicators
- Indicator
Alphabet and Meta initial 2027 capex guidance each implying below 40% year-on-year growth (forecast F-01-01)
- Source
S-01-03 Hyperscaler investor relations, quarterly
- Would Be Visible By
2027-02
- Indicator
Combined lease commitments in the four hyperscalers' annual filings growing more slowly than the 2025-2026 pace
- Source
S-01-01 SEC EDGAR, 10-K leases footnote, annual
- Would Be Visible By
2027-03
- Indicator
Nvidia Data Center revenue growth decelerating but staying positive sequentially
- Source
S-01-02 NVIDIA quarterly results, quarterly
- Would Be Visible By
2027-02
- Affected Industries
- 01
- 03
- 05
- 07
- 18
- What Businesses Should Do
Underwrite compute contracts on the counterparty's cash flow, not on its backlog. If you are a tenant, negotiate for assignability; if you are a landlord, cap single-tenant concentration. Treat the 2027 capex guidance print in February 2027 as the year's single most informative data point.
- Precedence Note
Distinguished from downside by spread behaviour, not by capex level: base is widening inside 150bp over a quarter; downside is more than 150bp without a corresponding move in IG credit generally.
- Would Change Our Mind
Two consecutive quarters in which data-centre ABS spreads widen more than 150bp while IG credit is flat.
- Assumptions
- Text
Prior-generation accelerators retain contracted demand through year five, as CoreWeave's disclosed A100 contract to 2029 implies.
- Confidence
medium
- Load Bearing
true
- Basis
T-01-03 evidence, CoreWeave Q2 2026 disclosure (S-01-01)
- If Wrong
Depreciation schedules shorten, GAAP losses deepen, and probability mass moves from base to downside.
- Text
Equity-market punishment of capex raises persists for at least two more earnings cycles.
- Confidence
medium
- Load Bearing
false
- Basis
T-01-01, Alphabet -5% on 2026-07-22 (S-01-03)
- If Wrong
Capex reaccelerates and mass moves to upside and to the failure branch simultaneously.
Cash flow catches the coupon
Inference demand broadens beyond coding and agent workloads, utilisation stays above 90%, and operating cash flow covers interest and maintenance capex at the neocloud layer as well as the hyperscaler layer. Debt is refinanced at tighter spreads in 2028 and the off-balance-sheet lease stock is termed out. The structure stops being a systemic question and becomes an ordinary infrastructure asset class.
Mechanism
Census BTOS small-firm AI use finally moves, which broadens the revenue base beyond the narrow set of buyers financing the build. Token volume growth outruns price decline, so revenue per installed megawatt rises. Lenders observe two years of stable utilisation and residual values, reprice the risk, and the refinancing cliff flattens.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- Adoption broadens beyond the current buyer set
- Serving cost per token falls at least as fast as price
- Policy rates do not rise further
- Early Indicators
- Indicator
Census BTOS current AI use in the smallest published employment-size class breaking above 25% (forecast F-01-06)
- Source
S-01-08 Census BTOS, biweekly
- Would Be Visible By
2027-12
- Indicator
A neocloud reporting positive GAAP net income for a quarter (forecast F-01-05)
- Source
S-01-01 SEC EDGAR 10-Q, quarterly
- Would Be Visible By
2028-03
- Indicator
Data-centre ABS issued at spreads tighter than the prior comparable deal, disclosed in the offering documents
- Source
S-01-22 academic and working-paper analysis of AI infrastructure finance
- Would Be Visible By
2028-06
- Affected Industries
- 01
- 03
- 05
- 07
- 11
- What Businesses Should Do
If you are a lender or a lessor, this is the branch in which pricing risk early is rewarded. Watch the BTOS small-firm series rather than vendor adoption claims - it is free, Tier A and currently flat.
- Precedence Note
Requires utilisation and cash flow, not announcements. New capacity commitments alone leave the outcome in base.
- Would Change Our Mind
BTOS small-firm AI use staying flat through all 2027 releases while capex reaccelerates.
- Assumptions
- Text
The demand curve for tokens is elastic enough to absorb the new capacity at prices that service the debt.
- Confidence
low
- Load Bearing
true
- Basis
Dossier 01 §11 lists this explicitly as a 'nobody can know' item
- If Wrong
Upside collapses into base; nothing moves to downside on this assumption alone.
- Text
Refinancing in 2028 happens at or below current rates.
- Confidence
low
- Load Bearing
false
- Basis
Macro brief, FOMC held 3.50-3.75% on a 9-3 vote with three dissents in favour of a hike, 2026-07-29
- If Wrong
The branch survives on operating cash flow alone, with a longer payback.
Spreads widen, equity absorbs it
A cloud revenue miss at one hyperscaler triggers repricing of data-centre credit. Private-credit marks fall, ABS spreads widen, and single-tenant facilities leased to unrated model labs cannot refinance on the original terms. The assets keep operating; the equity in the financing vehicles is impaired and several neoclouds are recapitalised or sold at distressed prices. Capex continues at a lower level.
Mechanism
The duration mismatch bites: long-dated debt against compute contracts of shorter tenor. A single quarter of decelerating cloud revenue causes lenders to re-underwrite residual values on accelerators, which raises required equity on new deals and lowers marks on existing ones. Counterparty concentration - facilities financed against one lease from an unrated, unprofitable lab - converts a demand wobble into a credit event at the vehicle level rather than at the operating-company level.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- A visible deceleration in cloud revenue at one of the four largest buyers
- Continued reliance on structured rather than corporate financing
- No policy easing that offsets the spread move
- Early Indicators
- Indicator
Data-centre ABS spreads widening more than 150bp over a quarter without a corresponding move in IG credit generally
- Source
S-01-13 Federal Reserve Bank of Dallas economic research, weekly
- Would Be Visible By
2027-09
- Indicator
A neocloud disclosing a covenant amendment or a backlog cancellation in a 10-Q
- Source
S-01-01 SEC EDGAR, quarterly
- Would Be Visible By
2027-11
- Indicator
Project withdrawals and moratoria rising in GW-denominated trade coverage
- Source
S-01-12 DatacenterDynamics, daily
- Would Be Visible By
2027-06
- Affected Industries
- 01
- 03
- 05
- 07
- 11
- 18
- What Businesses Should Do
Stress-test every compute commitment against a counterparty downgrade, not against a price decline. If you hold data-centre credit, the exposure that matters is single-tenant, single-counterparty, unrated - not aggregate leverage.
- Precedence Note
Distinguished from failure by whether the assets keep operating. Downside impairs the financing equity; failure strands the capacity itself.
- Would Change Our Mind
Two quarters of spread widening followed by a successful refinancing at par of a single-tenant lab-leased facility.
- Assumptions
- Text
Single-tenant facilities leased to unrated model labs cannot be re-let quickly at comparable rents.
- Confidence
medium
- Load Bearing
true
- Basis
T-01-02 and T-01-06 risk registers; no public dataset of re-letting outcomes exists
- If Wrong
The credit event is contained at one vehicle and mass returns to base.
- Text
At least one of the four largest cloud buyers reports a revenue deceleration large enough to be read as a signal.
- Confidence
medium
- Load Bearing
false
- Basis
S-01-03 quarterly disclosure; Azure grew 43% and Microsoft commercial RPO $678B (+84%) at the last print
- If Wrong
The trigger does not fire and the branch is deferred rather than removed.
Efficiency collapses the compute bill
An architectural or post-training advance reduces compute per unit of capability by an order of magnitude and is published openly, most plausibly by a lab optimising under export-control scarcity. Contracted capacity becomes long-dated excess, accelerator gross margins compress first, and the financing structures written against a rising compute price find themselves written against a falling one. Commercially this is good for buyers of intelligence and bad for owners of the assets that produce it.
Mechanism
A published efficiency result is replicated within weeks because the weights or the method are open. Serving cost per token falls faster than price, which is the opposite of the current margin concern, but total compute demanded for a given capability falls with it. Neoclouds whose per-megawatt yield was underwritten at 2026 token prices see revenue per megawatt fall while their debt service does not.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- An open-weight model reaching within 5 points of the top Artificial Analysis Intelligence Index score at under one-tenth the blended price
- At least two independent groups replicating the efficiency result
- No offsetting increase in demanded capability that absorbs the saving
- Early Indicators
- Indicator
An open-weight model within 5 points of the Intelligence Index leader at under one-tenth the blended price
- Source
S-01-10 Artificial Analysis Intelligence Index, irregular
- Would Be Visible By
2027-12
- Indicator
Chinese open-weight models holding four of the top five positions by routed token volume for two consecutive quarters
- Source
S-01-11 OpenRouter LLM rankings, daily
- Would Be Visible By
2027-06
- Indicator
Nvidia gross margin falling below 70% in any quarter
- Source
S-01-02 NVIDIA quarterly results, quarterly
- Would Be Visible By
2028-02
- Affected Industries
- 01
- 03
- 02
- 11
- What Businesses Should Do
Do not sign multi-year compute commitments at fixed per-token prices without a most-favoured-pricing or benchmark-reset clause. Two independent parties must show the result before you re-plan; one lab is a paper, two is a shift.
- Precedence Note
Requires an independent replication and a price effect. A single published result with no price move leaves the outcome in base.
- Would Change Our Mind
An efficiency result that is published, replicated, and followed by rising rather than falling aggregate compute demand - which would move mass to upside instead.
- Assumptions
- Text
Efficiency gains reduce total compute demanded rather than being absorbed by demand for more capability.
- Confidence
low
- Load Bearing
true
- Basis
Jevons-type absorption has been the pattern in this sector since 2023; this branch assumes it breaks
- If Wrong
The efficiency gain is absorbed, the branch collapses into upside, and asset values are unaffected.
- Text
Independent capability measurement remains funded well enough to detect the result.
- Confidence
low
- Load Bearing
false
- Basis
Epoch AI's inference price dataset last updated 2025-03-12; Artificial Analysis pricing endpoint returned HTTP 429 during research
- If Wrong
The branch occurs but is not observable on the timescale needed to act.
Lease and useful-life accounting is forced open
Disclosure rather than prohibition is the binding regulatory event. SEC comment letters, or a standard-setter project, force explicit disclosure of accelerator useful lives, signed-but-not-commenced lease commitments and single-counterparty concentration. The numbers are worse than the estimates, the sector re-rates on information rather than on events, and the off-balance-sheet stock moves on balance sheet.
Mechanism
The SEC's Division of Corporation Finance issues comment letters on useful-life assumptions and lease-footnote completeness, which become public on EDGAR. Peers conform pre-emptively. Analysts reconcile the disclosed totals against the ~$970B Moody's estimate and find a gap in either direction, which itself becomes the story. Ratings agencies incorporate the newly disclosed commitments into leverage metrics.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- At least one SEC comment letter on data-centre useful lives or lease disclosure becoming public
- No offsetting deregulatory action on disclosure
- Continued growth in signed-but-not-commenced leases
- Early Indicators
- Indicator
SEC UPLOAD/CORRESP correspondence on EDGAR addressing accelerator useful lives or lease-commitment disclosure at any hyperscaler
- Source
S-01-01 SEC EDGAR, realtime
- Would Be Visible By
2027-12
- Indicator
Combined disclosed lease commitments across the four hyperscalers exceeding $1.20 trillion (forecast F-01-08)
- Source
S-01-01 SEC EDGAR 10-K leases footnote, annual
- Would Be Visible By
2027-03
- Indicator
Any of the four changing its stated useful life for server and network equipment
- Source
S-01-03 hyperscaler investor relations, annual
- Would Be Visible By
2027-12
- Affected Industries
- 01
- 07
- 05
- 11
- What Businesses Should Do
Build the lease-commitment and useful-life comparison now, from the footnotes, before it is done for you. The disclosure that matters is in the 10-K leases footnote and the property-and-equipment accounting policy, not in the earnings release.
- Precedence Note
Regulatory outranks base when a disclosure requirement changes, even if the quantities look base-like. A voluntary disclosure change without a regulator behind it stays in base.
- Would Change Our Mind
Two consecutive annual filing cycles with no comment-letter activity and no change in useful-life disclosure.
- Assumptions
- Text
The SEC continues to issue and publish comment letters on accounting estimates at this scale.
- Confidence
medium
- Load Bearing
true
- Basis
S-01-01; SEC correspondence is published on EDGAR after a delay
- If Wrong
The disclosure change happens later or not at all and mass returns to base.
- Text
Disclosed totals differ materially from the Moody's ~$970B estimate.
- Confidence
low
- Load Bearing
false
- Basis
Estimate versus sum-of-filings comparability is untested
- If Wrong
Disclosure confirms the estimate and the branch has no price effect.
Fibre, again
The thesis breaks rather than slows. A refinancing window closes with a large single-tenant portfolio unplaced; the equity in the structures is wiped out and a portion of the debt with it. Capacity is eventually used, years later, by buyers who did not pay for it. The holders are private-credit funds, insurance balance sheets and the lease counterparties, and the recovery period for the affected assets is measured in years rather than quarters.
Mechanism
The duration mismatch that broke the 1999-2001 telecom buildout operates identically here: long-dated debt against short-tenor revenue contracts and hardware of uncertain economic life. A model lab counterparty fails to raise its next round; its leases are rejected or renegotiated; the vehicles that financed its facilities default; lenders stop underwriting new single-tenant capacity, which removes the marginal bid for capacity under construction and converts a credit event into a construction halt.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- At least one unrated model-lab counterparty missing a funding round or restructuring a compute commitment
- Refinancing required within the window rather than deferrable
- No state or hyperscaler balance sheet stepping in as lessee of last resort
- Early Indicators
- Indicator
A disclosed default, covenant breach or lease rejection at any data-centre financing vehicle with a named AI counterparty
- Source
S-01-01 SEC EDGAR, realtime
- Would Be Visible By
2028-12
- Indicator
Announced data-centre project cancellations or moratoria rising sharply in GW-denominated trade coverage
- Source
S-01-12 DatacenterDynamics, daily
- Would Be Visible By
2028-06
- Indicator
FOMC Financial Stability Report escalating AI infrastructure financing from a flagged risk to a named vulnerability
- Source
S-01-14 Federal Reserve, semiannual
- Would Be Visible By
2028-06
- Affected Industries
- 01
- 03
- 05
- 07
- 11
- 18
- What Businesses Should Do
Know, for every compute contract you rely on, who owns the building and who owns the debt on it. In the 1999-2001 analogue the operating companies survived and the financing vehicles did not, and tenants of the wrong vehicle lost their capacity anyway.
- Precedence Note
Failure outranks all other branches on occurrence. Downside becomes failure at the point where capacity under construction stops rather than slows.
- Would Change Our Mind
A large single-tenant portfolio refinancing at par in a wide-spread environment, which would show the residual-value fear is unfounded.
- Assumptions
- Text
No lender of last resort - state, sovereign fund or hyperscaler - steps in to take over distressed single-tenant capacity.
- Confidence
low
- Load Bearing
true
- Basis
T-01-12 records state capital entering compute provision directly; the Pentagon was reported on 2026-09-11 in talks to lend Fluidstack $5B (single-source signal)
- If Wrong
Failure is truncated into downside: the equity is still impaired but capacity is not stranded.
- Text
The 1999-2001 fibre analogue transfers, because the financing structure rather than the technology is what broke then.
- Confidence
medium
- Load Bearing
false
- Basis
Dossier 01 §7, item 2
- If Wrong
The analogue misleads and the branch's probability is overstated.
Additional scenario notes
- Probabilities Sum
1
- What Must Be True To Grow
- Accelerator residual values at year five support 70-80% asset-level leverage
- Contracted backlog converts to cash at a rate that covers interest through the refinancing window
- Interconnection and large-load rules resolve in a way that lets contracted megawatts become energised megawatts
- The buyer base broadens beyond the handful of counterparties currently financing the build
- What Could Stop It
- Capital cost — sticky 3-4% inflation with three FOMC dissents in favour of a hike makes refinancing the binding constraint
- Demand deflation — the pipeline contains optioned and duplicate capacity, as Exelon's disclosed 22% pipeline materialisation shows on the power side
- Measurement revision — the ~$970B Moody's lease estimate and the sum of filings may not measure the same thing
- Physical constraint — interconnection, turbines and transformers meter the build regardless of financing
- Substitution — an efficiency advance that collapses compute demanded per unit of capability
- Uncertain Assumptions
- That gross margin on inference at any frontier lab supports the contracted prices — no lab discloses it and no current independent price-per-token index exists (resolution gap, see F-01-03)
- That contracted GPU capacity is actually utilised at the rates implied by backlog figures — utilisation is not disclosed by anyone
- That the real economic life of an H100/A100-class asset is at or above the depreciation schedule applied to it
- That the ~$660B of off-balance-sheet lease commitments is the correct figure rather than a floor
- Authored
2026-09-15