H100 rental rates are up roughly 40% since October 2025. A100s — 2020-generation hardware — are up around 14% year-to-date. Forward GPU capacity through mid-2026 is reportedly booked.
Most people read that as proof AI demand is real. Read for capital, it is an incomplete signal. The rental price chart shows demand. It does not show the financing structure sitting underneath it — and that structure is where the risk now lives.
Edition 10 closed by asking what kind of asset you are actually buying in the intermediation layer — a standalone business, or a captive feature. This edition takes the same standalone-versus-captive lens one layer down, to the compute asset itself.
The Signal
Rising rental rates are not, by themselves, evidence that the hardware is appreciating. They are evidence that deployable, powered capacity is scarce. Blackwell availability is reportedly constrained into mid-2026, leaving H100 as the practical option for many operators, and older GPUs are retaining monetisable utility longer than the standard depreciation curve assumed — A100 and H100 indices are both trending higher year-to-date.
The scarcity runs deeper than GPU supply alone. One clear market signal of how scarce pre-energised grid connections have become is that AI cloud operators are acquiring or partnering with former Bitcoin mining infrastructure partly to accelerate access to power capacity and reduce interconnection and development bottlenecks. CoreWeave's acquisition of Core Scientific — giving it approximately 1.3GW of gross power capacity across an existing data centre footprint [REPORTED — CoreWeave / Core Scientific Form 8-K] — underscores how decisively the binding constraint has shifted toward power access. The GPU rental price chart does not capture the power access constraint sitting underneath it.
The rental price chart shows demand. It does not show the financing structure underneath it.
The Engineering Read — Technical vs Economic Obsolescence
The core distinction this edition turns on is that technical obsolescence and economic obsolescence are not the same thing. Technical obsolescence is hardware-driven: a newer GPU generation arrives with superior performance characteristics, making the prior generation less competitive for frontier workloads. Economic obsolescence is demand-driven: the prior generation loses its ability to generate rental income at rates sufficient to service the debt underwritten against it.
A GPU generation can lose frontier-performance relevance while still retaining substantial cash-flow utility if compute scarcity persists. The depreciation clock lenders use and the cash-flow clock operators experience are running at different speeds. When they converge at refinancing time, the gap becomes the equity problem.
Agentic load · a utilisation extender, not just a demand story
Agentic AI may be extending compute utilisation by making workloads more persistent and infrastructure-like, rather than bursty and episodic
Agentic pipelines push CPU-to-GPU ratios from 1:8 toward 1:1 or 1:2 [DIRECTIONAL — TrendForce]
Every gigawatt of agentic capacity is estimated to require approximately four times the CPU cores of a traditional AI training cluster [DIRECTIONAL — Tom's Hardware / CreditSights]
Implication: if economic life exceeds technical life, infrastructure investors and lenders conflating the two are applying the wrong depreciation clock to their underwriting [HYPOTHESIS]
Neocloud leverage · the risk layer below the demand signal
CoreWeave reported US$21.6B in total indebtedness as of December 2025, with an additional $8.5B delayed-draw term loan executed in March 2026 [REPORTED]
The critical refinancing variable is not GPU age alone — it is whether contracted cash flows remain sufficient to support debt-service coverage after collateral marks reset lower [HYPOTHESIS]
Implication: elevated rental rates tell you demand is real; they do not tell you whether DSCR holds at the point collateral is remarked and the loan comes up for renewal
The Capital Allocation Read
1
The underwriting question has shifted — but it has not been resolved.
The rental market is currently answering the asset-life question favourably — older hardware retaining monetisable utility longer than the standard depreciation curve assumed
But the question has not been resolved. It has been deferred by an unusually strong demand environment and a supply constraint that may not persist
The correct underwriting frame is now: how long can each generation of compute remain monetisable — and is demand durable enough to service the debt sitting underneath it?
The market is already beginning to test this frame. The reported Apollo–Blackstone financing supporting Google's TPUs leased to Anthropic separates ownership of the compute assets from their use, with institutional lenders underwriting lease cash flows rather than hardware resale value alone — but the senior debt reportedly required a vendor residual-value backstop to price, and the unbacked tranche cleared at a materially higher coupon. That spread is the tell: institutional capital is testing whether long-duration compute cash flows are durable, not yet asserting that they are [REPORTED]
2
Technical obsolescence and economic obsolescence are not the same thing.
Technical obsolescence is hardware-driven: a newer GPU generation arrives with superior performance, making the prior generation less competitive for frontier workloads
Economic obsolescence is demand-driven: the prior generation loses its ability to generate rental income at rates sufficient to service the debt underwritten against it
The market may increasingly bifurcate between these two timelines — a GPU generation can lose frontier-performance relevance while still retaining substantial cash-flow utility if compute scarcity persists
Implication: infrastructure investors and lenders who conflate technical and economic obsolescence are applying the wrong depreciation clock to their underwriting [HYPOTHESIS]
3
The GPU debt cliff is a synchronisation risk, not a single-event risk.
A large cohort of neocloud operators entered the market in the same 2023–2024 window, on similar debt structures, underwritten by lenders sharing the same AI demand thesis
Their refinancing windows cluster in the same 2026–2028 period — at the same moment Blackwell and Vera Rubin are compressing H100 and A100 collateral values
Implication: synchronised refinancing pressure across an entire asset class — a structural feature of how the build-out was financed [HYPOTHESIS]
4
For APAC data centre operators, the neocloud credit question is a tenant risk question.
APAC colocation operators increasingly count neocloud operators among their anchor tenants
The occupancy assumptions embedded in APAC data centre underwriting inherit the neocloud financing risk whether or not the operator has direct visibility
Implication: APAC data centre underwriting now requires a view on tenant capital structure, not just tenant demand outlook — confusing them is a material underwriting gap
Investment Lens
Constraint Technical vs economic obsolescence mismatch
Impact Lenders applying technical depreciation schedules may mark collateral down before cash flows deteriorate — or hold too long if scarcity masks economic decline
Capital response Underwriting frameworks separating technical useful life from contracted cash-flow duration
Winning asset Long-tenor offtake agreements with creditworthy counterparties that outlast the hardware generation cycle
Constraint DSCR compression at refinancing
Impact Collateral remarking compresses advance rates even when rental cash flows remain intact
Capital response Equity buffers sized against collateral compression; staggered debt maturities; contracted backlog as cash-flow security
Winning asset Operators with balance sheet strength to absorb fleet refresh without forced refinancing
Constraint Synchronised refinancing window (2026–2028)
Impact Neocloud cohort hits debt maturity simultaneously, amplifying any credit tightening or collateral repricing
Capital response Lenders likely to differentiate between operators with contracted hyperscaler offtake vs spot rental market dependence
Winning asset Platforms that entered earlier, locked in longer-tenor contracts, carry lower leverage relative to contracted cash flows
Constraint Tenant credit opacity (APAC)
Impact Colocation occupancy assumptions inherit neocloud refinancing risk without direct visibility
Capital response Tenant capital-structure diligence as a condition of underwriting occupancy, not just demand
Winning asset Operators with diversified, creditworthy tenant books rather than concentrated neocloud exposure
5
AI infrastructure risk has migrated — from demand risk to power access, supply allocation, and asset-life underwriting risk.
Current rental pricing and forward booking signals suggest demand is real — for now. But that answer is contingent on scarcity conditions that may not persist
Implication: capital allocated using a demand-risk framework may now be misaligned with where the actual structural risk lives [HYPOTHESIS]
What This Means for the Broader Market
Elevated rental rates are a necessary condition for the asset-life thesis. They are not sufficient proof of it. The scarcity signal is real. The hoarding risk is real. The refinancing window is real. But capital allocated on a demand-risk framework may now be misaligned with where the structural risk actually lives.
The discipline this edition argues for is to stop underwriting AI infrastructure on demand alone and start underwriting the asset-life assumption directly — how long each generation of compute remains monetisable, and whether the cash flows last long enough to clear the debt sitting underneath them. Compute scarcity is the signal. Capital structure is the risk.
If the depreciation clock and the cash-flow clock are diverging — if a compute asset can keep generating contracted cash flows well past its frontier-performance life — then the next question is not technical but financial: how does capital respond when an asset it once treated as fast-depreciating hardware starts to behave like something with durable, contractable cash flows? Recent financing structures suggest the market may already be testing an answer.
Edition 12 picks that up.
Key Sources
Market data and indices
→ SemiAnalysis GPU Rental Price Index — H100 1-year contract pricing, October 2025 to March 2026 [OBSERVED]
→ Silicon Data SDA100RT / SDH100RT Index — A100 and H100 rental price indices, YTD May 2026 [OBSERVED]
→ TrendForce — CPU-to-GPU ratio projections for agentic AI deployments [DIRECTIONAL]
→ Tom's Hardware / CreditSights — agentic capacity CPU core requirements per gigawatt [DIRECTIONAL]
Compute and capital structures
→ CoreWeave Form 10-K (2025) — total indebtedness US$21.6B as of 31 December 2025 [REPORTED]
→ Reuters — CoreWeave $8.5B delayed-draw term loan, March 2026 [REPORTED]
→ CoreWeave / Core Scientific Form 8-K — approximately 1.3GW gross power capacity [REPORTED]
→ Apollo / Blackstone Anthropic TPU financing — SPV lease structure and Broadcom residual-value support, finalised 5 June 2026 (FT / Bloomberg / Investing.com, as reported) [REPORTED]
Verification notes
[REPORTED] CoreWeave debt figures and Core Scientific acquisition sourced from public SEC filings.
[REPORTED] Apollo / Blackstone Anthropic TPU financing terms sourced from press coverage; not independently verified against transaction documents; cited as early evidence of an underwriting shift, not a settled structure.
[OBSERVED] GPU rental index movements (SemiAnalysis, Silicon Data); market data, not independently verified against primary operator reporting.
[DIRECTIONAL] Agentic compute multipliers and CPU-to-GPU ratio projections; directionally credible but not settled benchmarks.
[HYPOTHESIS] Refinancing synchronisation thesis and the technical-vs-economic obsolescence distinction; analytical inference based on publicly available capital-structure data.
No investment advice intended or implied.