Scenarios
What could move the whole sector, for better and for worse: how much each scenario would matter, how likely it is, and which way the latest data is pushing it.
Economic signals are improving; physical and financial-structure signals are deteriorating.
Demand, rental prices, residual values and the cost of new debt all moved the right way in 2026. Power, equipment and construction, the 2027–2028 debt wall, customer concentration and circular financing moved the wrong way. The stress case is not a fall in demand but a credit accident at a large counterparty at the moment of refinancing.
| Importance | Under 25% | 25–49% | 50–74% | 75% and up |
|---|---|---|---|---|
| 5/5 | ||||
| 4/5 | ||||
| 3/5 | ||||
| 2/5 |
Likelihood across the top. Hover a code to see its title, click to jump to it.
Downside risks 15
StructuralCan break the business model
- D1
Compute demand breaks
Less likelyImportance5/5Likelihood20%2027–2028The master risk: every other risk is bearable while demand holds, none is if it breaks. Buyers stop expanding because model revenue lags the build-out, a large lab cannot fund its commitments, or capacity already bought sits idle.
Latest signal · Nov 8, 2026Clear improvement. H100 rental prices are up about 38% and A100 about 20% since the start of 2026; the median on-demand H100 reached about $3.81 per GPU-hour in mid-October, up 21% in a year. Google Cloud's backlog exceeds $460B and Microsoft's commercial RPO reached $627B.
- D2
The refinancing wall
MixedImportance5/5Likelihood30%2026–2028Debt raised in 2024 and 2025 comes due in 2026–2028. If lenders pull back when it does, after a credit event at a leveraged AI name or a fall in rental prices, companies refinance on punitive terms or not at all.
Latest signal · Nov 8, 2026Mixed, to watch closely. CoreWeave carried $35B of debt at 30 June 2026, with maturities concentrated in 2026–2028. New debt is cheaper: its $8.5B DDTL 4.0 was rated investment grade (Moody's A3), secured by assets and contracts. But in July Nvidia started backstopping lenders by committing to re-rent unused neocloud GPU capacity at a fixed rate, which extends its exposure to credit risk.
- D3
Customer concentration and counterparty credit
MixedImportance5/5Likelihood30%2026–2029A backlog is only worth something if the customer pays for five years. A handful of tenants carry most of it, and more of them are AI labs whose commitments far exceed their revenue. One default or renegotiation hits revenue and the debt raised against it at once.
Latest signal · Nov 8, 2026Improving but still high. CoreWeave's largest customer was 45% of revenue in Q1 2026, down from 72% a year earlier, and S&P moved its outlook to positive on that. But the risk is shifting: more contracts are signed with non-investment-grade AI labs whose commitments far exceed their revenue.
- D4
GPU residual values erode
Less likelyImportance4/5Likelihood30%2027–2029Neoclouds onlyOlder GPUs rent and resell for less than the models assume, because a new generation arrives faster or performs far better. This decides whether the unit economics are real or only an accounting result.
Latest signal · Nov 8, 2026A surprising improvement in 2026, but the debate is open. Silicon Data reports the A100 stopped depreciating at the end of 2025, and H100 and B200 values rose in 2026 with rental rates; CoreWeave re-signed 2022 H100 contracts at about 95% of the original price. The bear case: if the real economic life is 3 years rather than 6, the industry understates depreciation by $50–60B a year. Amazon shortened server lives in 2025 while Meta lengthened them the same quarter.
- D5
Overcapacity compresses compute prices
Less likelyImportance4/5Likelihood25%2027–2029Distinct from a fall in demand: demand can keep growing while prices collapse, if supply grows faster. Many projects started in 2025 and 2026 all deliver in 2027 and 2028.
Latest signal · Nov 8, 2026Not visible yet: spot rental prices rose through 2026 and the physical bottlenecks on power and equipment slow new supply.
ExecutionDelay the build or raise its cost
- D6
Grid interconnection delays
More likelyImportance4/5Likelihood75%2026–2029Power, not GPUs, is now the first physical constraint. Projects without secured interconnection wait years for it.
Latest signal · Nov 8, 2026Clearly worse. Securing interconnection typically takes 24 to 72 months, up to 5–7 years in constrained regions; the ERCOT queue reached about 410 GW in April 2026, 87% of it data centers, and the average wait in Virginia is seven years. About half of the 12 GW of AI capacity planned for 2026 was delayed or cancelled over grid constraints (Allianz Trade).
- D7
Construction delays and component inflation
More likelyImportance3/5Likelihood70%2026–2028Labor shortages, record equipment lead times and rising component prices push delivery back and cost per MW up.
Latest signal · Nov 8, 2026Worse. Microsoft attributes about $25B of its 2026 capex to higher component prices alone, and Meta raised its guidance for the same reasons. HBM memory, transformers, gas turbines and switchgear all have record lead times.
- D8
Rates rise
MixedImportance3/5Likelihood30%2026–2028Higher Treasury yields raise the cost of every new facility and the discount rate behind every valuation. Real, but behind the refinancing wall: the sector absorbs high rates as long as its spreads tighten.
Latest signal · Nov 8, 2026Contained so far: each generation of CoreWeave debt has cost less than the one before, which offsets high base rates.
ContextRedraw the map more than they shrink it
- D9
Hyperscaler capex slows
MixedImportance4/5Likelihood20%2027–2028The four largest buyers cut or pause their spending guidance, once free cash flow turns negative and boards push back.
Latest signal · Nov 8, 2026For now the opposite of a risk: the four plan about $725B of capex in 2026, up 77% on last year's record of about $410B. The leading signal is free cash flow, not capex, and it is weakening: Amazon's trailing FCF shrank to $1.2B, down 95% in a year.
- D10
Custom silicon displaces Nvidia fleets
More likelyImportance3/5Likelihood30%2027–2030Neoclouds onlyHyperscalers move more of their workloads to their own chips (TPU, Trainium, Meta's silicon with Broadcom), and demand for rented Nvidia capacity falls.
Latest signal · Nov 8, 2026Amazon expects Trainium to save it tens of billions of capex a year, and Meta is developing its own chips with Broadcom.
- D11
Circular financing unwinds
More likelyImportance3/5Likelihood20%2026–2029Nvidia is at once equity investor, GPU supplier and leasing customer of the neoclouds. While it turns, the loop accelerates everything; if one link breaks, it amplifies the fall.
Latest signal · Nov 8, 2026The loop is tightening: Nvidia leased back about 18,000 GPUs from Lambda for $1.5B, and since July it backstops lenders against unused neocloud capacity.
- D12
An efficiency shock cuts compute per unit of output
MixedImportance3/5Likelihood20%2026–2027A model or technique delivers the same results with far less compute, and buyers conclude they have over-ordered.
Latest signal · Nov 8, 2026Distillation and better tokens per watt reduce the compute needed per unit of value, but so far the Jevons effect dominates: cheaper inference leads to more consumption, not less.
- D13
Equity de-risking compresses multiples
MixedImportance3/5Likelihood45%Any timeInvestors rotate out of AI infrastructure regardless of fundamentals, as the sector trades as one high-beta group.
Latest signal · Nov 8, 2026Fundamentals improved in 2026, but the group still trades as one high-beta block, so a drawdown elsewhere in AI reaches it regardless.
- D14
Permits and local opposition
More likelyImportance2/5Likelihood60%2026–2029Local moratoriums and a backlash over residential power prices slow or block projects. A site-by-site risk: it redraws the map more than it cuts total volume.
Latest signal · Nov 8, 2026Rising: more local moratoriums and a growing backlash over residential electricity prices in the main hubs.
- D15
Geopolitics and export controls
MixedImportance2/5Likelihood35%Any timeChip export restrictions, tariffs on imported equipment and even temporary limits on access to some models. A tail risk, hardest on non-US neoclouds.
Latest signal · Nov 8, 2026Effects were visible this year, with chip restrictions and temporary limits on access to some models. Hard to measure, but a real tail risk for non-US neoclouds such as Nebius.
Upside scenarios 8
- U1
Hyperscaler and AI-lab capex keeps compounding
More likelyImportance5/5Likelihood65%2026–2028The largest cloud providers and AI labs keep raising their data-center budgets and sign capacity ahead of need, so contracted demand stays ahead of what can be built.
Latest signal · Nov 8, 2026Amazon, Microsoft, Alphabet and Meta plan about $725B of capex in 2026, up 77% on last year's record of about $410B.
- U2
Rates fall and credit stays open
MixedImportance4/5Likelihood45%2026–2027Policy rates ease and lenders keep funding AI infrastructure: GPU-backed loans, investment-grade project debt and convertibles stay available at narrowing spreads.
Latest signal · Nov 8, 2026Credit is open and getting cheaper for the best collateral, with CoreWeave's $8.5B DDTL 4.0 rated investment grade, but the 2026–2028 maturities still have to be refinanced.
- U3
Inference demand keeps compute scarce
More likelyImportance4/5Likelihood60%2026–2027Reasoning models and agents turn each user request into far more compute, so inference demand grows faster than supply even as training spend levels off.
Latest signal · Nov 8, 2026The median on-demand H100 price reached about $3.81 per GPU-hour in mid-October, up 21% in a year, and Google serves 16 billion tokens a minute through its API.
- U4
Energized power becomes the scarce asset
More likelyImportance4/5Likelihood70%2026–2030HPC/AI Colocation onlyGrid interconnection and generation stay the binding constraint, so sites that already have power, or a firm path to it, are worth far more than land and shells.
Latest signal · Nov 8, 2026Interconnection takes 24 to 72 months and about half of the 12 GW of AI capacity planned for 2026 was delayed or cancelled over grid constraints, which raises the value of sites already energized.
- U5
Contract quality improves
MixedImportance3/5Likelihood55%2026–2028Leases and capacity contracts move toward investment-grade tenants, take-or-pay terms, prepayments and longer durations.
Latest signal · Nov 8, 2026CoreWeave's top customer fell from 72% to 45% of revenue in a year and new debt reached investment grade, but more contracts are signed with non-investment-grade AI labs.
- U6
Older GPU generations keep earning
More likelyImportance3/5Likelihood60%2027–2029Neoclouds onlyPrior-generation chips stay in demand for inference and smaller models, so GPUs resell and re-contract at good prices well past their first term.
Latest signal · Nov 8, 2026H100 rental prices are up about 38% and A100 about 20% since the start of 2026, and CoreWeave re-signed 2022 H100 contracts at about 95% of the original price.
- U7
Efficiency gains expand demand instead of shrinking it
More likelyImportance3/5Likelihood45%2026–2027Cheaper tokens open new uses faster than they reduce the compute each one needs, so total demand rises when models get more efficient.
Latest signal · Nov 8, 2026So far the Jevons effect dominates: cheaper inference per token has come with more consumption, not less.
- U8
Demand broadens beyond a handful of labs
MixedImportance2/5Likelihood50%2027–2029Sovereign programs, enterprises and regional clouds sign their own capacity, in Europe, Australia and the Gulf as well as the United States.
Latest signal · Nov 8, 2026Customer concentration is falling at the largest neocloud, but much of the new demand still comes from a few AI labs.
Editorial judgements, not model outputs. Importance (1–5) is how much of the sector's value the scenario would move; likelihood is our estimate of the chance, in percent, that it materially plays out within its horizon; the signal says which way the latest data is moving that chance. Downside risks are grouped by tier: structural risks can break the business model, execution risks delay the build or raise its cost, context risks redraw the map. Scenarios are not mutually exclusive, so likelihoods do not add up. Figures in the signal notes come from the sources listed with each scenario. Not investment advice.