Shaduf.Research preview
Is AI a Bubble?/Capacity & recovery
Skip to content

Is AI a Bubble? · Topic guide

Capacity & recovery

A functioning asset, a protected invoice and a rewarding investment are different achievements.

Updated as new research is ready

Substitution evidence: . Historical deployment measurements and conditional contract/fleet cases remain separate. Complete new investigation.

Document-work capacity interpretation: . A historical timed batch and explicit one-year assumptions do not refresh earlier fleet, facility or customer-acceptance observations.

Evidence dates and measurement boundaries

Maintained synthesis including 29 September 2026 evidence on older-design contracts and reused facilities. Earlier physical, financing, ownership and payer findings retain their own dates; the Austin budget has a March 2025 historical anchor.

Holder-side context added from the 2 October 2026 investigation; the held-position snapshot is June 30. No older capacity measurement or continuation input is refreshed by that addition.

Connected financing context: . The source portfolio, tenant capital, fee credits and maturity calendar are not a new current asset appraisal.

Completion-to-cash evidence: , retrieval-day boundary. June accounts, contractual milestones and conditional resource tests remain distinct; no October cash balance or compliance certificate is asserted.

Useful capacity can support payment without settling the return

Useful capacity can continue after a first use or hardware advantage, without establishing full original recovery or a justified ownership price. The earlier older-design contract and reused-facility evidence strengthens selective persistence. The new deployed-service comparison also shows that useful work can move to another architecture; contract timing and whole paid replicas decide whether the buyer’s bill falls. Maintenance, rent, customer credits and financing still determine what reaches investors. A strong payer’s valid invoice can protect a lender while the owner earns too little.

The useful comparison is not “announced versus fake.” Follow the equipment and facility through delivery, acceptance, service, collection, full cost and replacement. This guide joins the physical recovery investigation, the financing evidence, the ownership-price test and the payer-budget constraints. Their financial and contractual dates remain distinct.

The latest customer-capture findings strengthen the possibility of ongoing commercial demand and shared customer/provider gains. They do not establish a future five-year rental price, a measured GPU utilization curve or a complete infrastructure cash margin. Those still need evidence at the relevant service and investment boundary.

The older-capacity investigation adds new evidence for the alternatives. Its historical cost budget is not an appraisal to insert into the earlier IREN models.

The workload can move before the bill falls

A feasible alternative adds a real choice, not a universal instruction to switch. Loka’s December 18, 2024 account compares optimized NVIDIA-based G5 and Inferentia2 for fine-tuned SDXL interior-design images. It reports 17.5% lower component cost and 0.31 seconds longer per image. The ten-million-image saving is a projection; accepted output, private contract and full migration cost were not established. Tomofun’s May 6, 2026 real-time migration reports a differently defined 83% deployment-cost reduction, not the same workload or cost denominator. Deployments and source limits.

A capacity promise and a spending promise are different purchases. A Savings Plan creates an hourly floor under specified eligibility rules; it does not reserve the machine. Consumed reserved capacity must not be charged twice. The full comparison includes extra storage, network, validation, review, support, retries and interruption where they differ. A rented instance already charges for its supplier-owned capital. Complete service and purchasing terms.

Historical cost scaffold + conditional contract model

Waiting can preserve part of the saving

Waiting can preserve part of the savingIn an assumed twelve-month, ten-million-image comparison with three months of old payments left, switching with a stranded family commitment costs 88,168 dollars versus 82,017 to stay. Waiting or eligible portable spending costs 71,252 before other differences.CONDITIONAL 12-MONTH REFERENCE · USD10m projected images; three months locked; equal accepted service.No extra migration/service cost or other use of the old allowance.Stay with the reference service$82,017Switch with no old commitment$67,664Switch; old-family bill remains$88,168Switch within an eligible portable plan$71,252Wait three months, then switch$71,252Switching early can cost $6,151 more than staying.Waiting retains $10,765 of the reference advantage.Dark: modeled cloud payments. Pale: separate old-family obligation.These are reconstructed scenarios, not reported customer bills.

On a narrow screen, scroll horizontally to inspect all labels.

Loka’s December 2024 rounded 17.5% reduction and $14,353 projection anchor the scale. The horizon, commitment and equal-service conditions are assumptions. Published current terms do not identify Loka’s contract. Portable and waiting costs coincide only under these conditions. Actual extra costs reduce the saving. Definitions, exceptions and equations.

The chart allocates a rounded historical cost projection over an assumed twelve months. Three months of remaining, nonreturnable, no-upfront payments are a scenario, not evidence of Loka’s contract. No other eligible use and zero extra migration/service costs isolate the mechanism. Only 2.1 stranded months exhaust the reference 17.5% component saving. Waiting avoids the second-service bill during the lock; actual returns, negotiated terms or separately evidenced reuse change the comparison. Exceptions, upfront cash and the waiting option.

Whole replicas can independently change the result. Under the separate assumed 12-second GPU mean, 0.31-second increment, equal acceptance and one paid standby each, a reference load of 1.80 needs two active machines on each side. At 1.98, the alternative needs a third active machine while the reference still needs two. Its model fleet bill moves from 80.42% to 107.23% of the reference, even though the per-active-work component remains cheaper. These loads are not observed utilization and the assumed topology is not a certified service design. Whole paid capacity, not fractional machines.

Preserving gross revenue after a 17.5% effective price cut requires 21.21% more paid and collected output in the equal-quality reference. That is a requirement, not a demand response. The supplied serial example needs 24.34% more active resource before integer rounding and standby. Expansion and renewal must still be funded.

The best feasible incumbent may therefore remain cheaper as a complete service: usable software, response time, available machines, resilience and existing commitments matter. An alternative can still win at renewal or for suitable work. Neither outcome supplies a new appraisal of the older capacity studied below.

An older design can win a new contract; the machines’ age still needs evidence

CoreWeave’s corrected 11 August 2026 call reports a recently signed A100 contract extending into 2029. A buyer can still choose an older-design service, so a new chip generation need not extinguish all earlier capacity value. The evidence is a company-reported contract, not inspected lifetime cash.

A100’s 2020 introduction does not date installation of the specific units. Price, scale, the customer’s identity, repairs and any in-term equipment replacement were not disclosed. “Nine years after introduction” must not become “nine years of measured profit from these machines.” Contract and service comparison.

Evidence and interpretation

Four clocks—not one lifetime

Scroll horizontally to inspect the figure.

Four clocks—not one lifetimeDesign introduction, equipment installation, customer protection and financing each have different dates. The reported A100 contract establishes only some of them.DESIGN VINTAGEA100 introduced in 2020A product generation—not the age of the rented machines.MACHINE AGEInstallation dates not disclosedNot a nine-year record of these units’ performance or profit.CUSTOMER CONTRACTContract reported in August 2026 → into 2029Price, scale, customer and full service costs undisclosed.FINANCING CLOCKApply the particular debt termsA new service contract does not establish that a loan is paid off.
CoreWeave’s corrected 11 August 2026 call reports the A100 contract. It does not identify a particular machine’s installation date or disclose a lifetime cash return. CW CALL

Accepted, busy and paid are not the same state

For IREN's Horizon project, the August disclosure established acceptance of the first planned 50-MW phase, not completion of all four planned phases or a September meter reading. Contracted capacity, energized IT capacity, physical workload and paid reservation can each describe different quantities. A paid reservation may be economically secure while the buyer has unused capacity. Acceptance and measurement distinctions.

Availability asks whether service can be delivered. Compute efficiency asks how much useful processing is obtained while equipment operates. Electrical load measures power use. None is automatically the share of commercially unsold capacity. Similarly, a benchmark may describe throughput at a quality and latency target without establishing paid demand or investment return.

A megawatt of IT load is not necessarily comparable with another company's operator-defined “active,” “contracted” or “billing” megawatt. Gross hosting revenue can include reimbursed power. A facility-only landlord and an integrated compute seller do not have the same investment or expense base. Facility and service comparisons.

The new document workload adds a different capacity test. The source reports a 10,000-page batch in 2,207 seconds on one H100. Applying an assumed 75% productive-time ceiling over 8,760 hours gives 107.168m submitted pages per active GPU-year. This is not sustained production telemetry, customer acceptance or observed demand. Document mix, host limits, outages and peaks can reduce it. A passive hot spare is paid for but is not simultaneously counted as active throughput. The timed source and hybrid-price boundary.

Accepted-work economics · evidence and assumptions distinguished

Preserving revenue is not preserving room to run the business

On a narrow screen, scroll within the figure to inspect the labels.

Preserving revenue is not preserving room to run the businessA thirty percent price reduction requires 114.286 million inputs to preserve the starting $80,000 revenue, beyond a one-active-GPU assumed limit of 107.168 million. Two paid active GPUs require 151.703 million inputs to preserve the $53,415.70 remaining-cost budget before new unpriced costs. The capacity limit is a research assumption, not telemetry.30% PRICE-CUT TEST · PAID VOLUME IS ASSUMED OR SOLVED, NOT OBSERVEDStarting case: 80m inputs at $0.001One active GPU; $53,415.70 remains before other costs.At $0.0007: 114.286m to preserve revenueExceeds the one-active-GPU scenario capacity of 107.168m.Two active GPUs: 151.703m to preserve the budgetPays the second rental; holds other costs unchanged.0100m200m inputsDashed marks: stipulated ceilings for one active GPU (top rows) and two (bottom).Historical timed work; 75% productive-time ceiling; full collection; no spare orextra review charged here. More support, review or development raises the requirement.
Conditional one-year operator at the inspected $2.99/hour continuous rental offer; the incomplete $391.90 development reference is charged once. The separate favorable 64m-to-80m path also fits the assumed one-active capacity. Both paths, their costs and the limits on scaling.

The positive branch does not require unlimited volume: 64m inputs increasing to 80m while price falls 10% fits the stipulated one-active capacity and leaves $8,000 more for other costs at full collection. The gain survives only if incremental review, storage, development and support fit inside it. With a larger price cut, preserving receipts alone can require more capacity; preserving the remaining-cost budget after paying for that capacity is a higher hurdle. Both complete cases.

A reported reduction in retries is another legitimate efficiency mechanism, but it is not applied again to a timed batch that may already incorporate the improvement. Nor does better throughput automatically lower fixed annual rent. The benefit can be more sellable work, fewer paid nodes or shorter rental time, depending on the arrangement. None of this assigns a current resale value to the earlier GPUs or connects the new workload to the financed halls below. Reliability engineering and double-counting limits.

Physical progress does not supply every financing certificate

Reported progress is meaningful without being universal completion. Core Scientific’s June accounts show $465.3m of construction entering service and a remaining construction-in-progress stock. An accounting commissioning amount is not acceptance of every financed hall or a measured GPU load. Final Commencement under Finance I requires all relevant rent-start conditions, not just an initial billable phase. The observed progress; the financing milestone.

Denton’s September issuer disclosure identifies existing and additional electrical capacity outside Batch Zero, a favorable distinction from a blanket new-load pause. A separate Building A fit-out registration lists $110m estimated cost and May 31, 2027 planned completion. That accessibility-review record is not tenant acceptance, an extra Finance I bill or proof that the whole campus is delayed. The link to a particular financed contractual phase remains unknown. Electrical classifications and contracted IT capacity are also different measures. What each Denton record establishes.

The new investigation therefore strengthens a conditional productive-completion path, not a binary “finished or fictitious” claim. Physical rights, delivery, acceptance and usable funds need to line up. The earlier service-end extension and full-cost recovery tests remain necessary after that transition.

Visual explanation

Paying back the equipment is not paying for the whole investment

Scroll the figure horizontally to inspect all labels.

Full capital budget and distinct return testsThe original IREN budget comprises 5.8 billion equipment and ancillaries plus 2.8 to 3.2 billion facilities. Original investment recovery, debt repayment and value to a new shareholder require different cash tests.IREN / Microsoft Horizon plan · original November 2025 budgetEquipment and ancillaries: $5.8bnFacilities: $2.8–3.2bn200 MW of IT capacityCombined original investment: $8.6–9.0bnA dated budget, not final as-built costs or today’s remaining cash bill.Three separate testsOriginal recoveryCreditor repaymentOwnership priceAll original costs + timeActual protected paymentsRemaining owner cash
The facility component includes its disclosed subcomponents; do not add them again. Capacity investigation; ownership analysis.

Charge the facility and the continuing service

The original IREN plan separately budgeted $5.8 billion for equipment and ancillaries and $2.8–3.2 billion for facilities: $8.6–9.0 billion combined. These are November 2025 estimates, not verified as-built costs or today's remaining expenditure. The facility subcomponents are already included and are not added twice. Owning the building removes a landlord's invoice, not the capital cost of providing it. Original cost perimeter.

A project earnings measure after power and direct maintenance but before other costs is not an all-cost cash margin. The recovery investigation deliberately shows both a favorable 85% project-measure-as-cash case and lower 80% and 75% cash-share sensitivities. The latter are assumptions, not measured tax or overhead allocations.

At an assumed 10% hurdle, synchronized five-year service and the $8.8 billion midpoint investment, the 85% case needs approximately $2.81 billion of net value at the end; the 80% case needs $3.42 billion. These are required continuation values, not appraisals, realized losses or actual project returns. They identify what another use or a sale would have to support after relevant costs. Model, exact figures and assumptions.

A protected fee total still has a cash clock

Customer acceptance and supplier payment need not occur together. Equipment can be shipped and payable before the customer accepts the service. A customer's advance can help finance construction, but the same advance is credited against later service; it is not a second sale. Conditional debt availability, funded notes and released cash likewise describe different stages. Customer, supplier and lender timing.

The IREN–Microsoft statement of work extends service end dates after qualifying delayed starts to preserve fees. A faithful timing stress shifts service rather than automatically deleting a final year. Time still costs money: capital can be tied up earlier, other cash costs may persist, and an extended service period does not itself amend a lender maturity.

The later contract inspection also found that Exhibit F disapplies the generic purchase-order termination-for-convenience provision. That prevents the boilerplate from establishing a cheap exit from this specific contract. It does not remove every supplier-performance remedy or establish the terms of all other purchases. Contract-specific flexibility finding.

This is why a cash-rich payer is not enough for a funding conclusion. The Spending & obligations guide distinguishes resources, actual payment schedules and expenditure that can legally change. An affordable investment ceiling is not proof of a permissible minimum-spending plan.

Qualifying equipment finance can arrive before acceptance

The IREN terms provide a real bridge between supplier payment and acceptance. One route can qualify when at least 97% of tranche servers are delivered no later than 45 days before the applicable delivery-delay window expires, with adjustments and all other credit-event conditions. Another follows equity/eligible-advance purchase and subsequent acceptance. These are operative possibilities, not observed tranche status. The document and scoped logic test.

IREN agreement · scoped routes, not a complete eligibility engine

Funding can cross the supplier–acceptance gap

Scroll horizontally on a narrow screen to inspect all labels.

Funding can cross the supplier–acceptance gapOne route permits funding before acceptance at 97 percent delivery and a deadline at least 45 days before window expiry, subject to other terms. An alternative follows equity-funded purchase and acceptance. A separate first-tranche advance exception operates before any acceptance. All have other conditions.DELIVERY ROUTE · acceptance need not come firstAt least 97% delivered ≥45 days before the applicable window expires.Adjusted deadlines, other credit-event conditions and release rules apply.A qualifying credit event can precede customer acceptance.A commitment, funded note and released supplier cash are still different.ALTERNATIVE · fund the purchase first, then acceptanceEquity and eligible advances fund the purchase; the customer accepts.Financing can then qualify through the separate reimbursement route.SEPARATE FIRST-TRANCHE ADVANCE EXCEPTIONBefore any acceptance, specified upfront funds can buy first-trancheservers and related equipment without those utilization conditions.No route erases escrow, sizing, certificates, liquidity or spare requirements.These are documented possibilities—not an actual tranche-status report.
The 97% route, reimbursement route and advance exception are distinct provisions. The spare-covenant cure is separate again; a 60-day qualifying cure requires $50m protected funds, not a generic delay fee. Common Terms Agreement.

The separate first-tranche exception matters: before any acceptance, specified upfront funds can buy the first tranche’s servers and related equipment without those utilization conditions. It is not unrestricted advance use. Funded notes can still require escrow release; undrawn commitments are not supplier cash.

Performance support survives early finance. The parent liquidity requirement steps from $200m through $150m, $100m and $50m as specified acceptances occur, subject to the final acceptance/termination terms. Qualifying assets can include securities and available parent revolving commitments; the amount is not automatically trapped bank cash, and a step-down is not a receipt. A separate 60-day extended spare-inventory cure requires $50m of additional protected funds only under its eligible-source and no-customer-credit conditions. That is not a generic construction-delay fee. Separate liquidity, escrow and spare conditions.

Valid early funding can reduce the equity bridge; missed eligibility or a continuing support requirement can tie capital up longer. Neither finding replaces the original facility-plus-equipment recovery model or creates a current asset appraisal.

A service can outlast a loan—but its later cash is not costless

The connected cash investigation documents six Core Scientific facilities across five sites, with 588.5 MW in the April detailed rows and a rounded 590-MW headline. These are source-defined IT capacities, not measured October occupancy. The host’s fee and construction perimeter is not the complete GPU-and-service investment borne by CoreWeave. The specific assets and cost allocation.

Customer contributions can finance construction while reducing later fees through credits. Ownership and replacement duties differ across asset categories. Austin’s headlease and shorter customer term also differ from the longer non-Austin contracts and their unexercised extensions. “The building survives” cannot establish an unrestricted, transferable stream of cash for every future lender.

The May notes mature on May 15, 2031; the June filing assumes amortization begins in November 2029. A long service relationship can support another financing, but does not put every future receipt before that maturity. The new test retains cash, pays the old claims and tests a finite next loan instead of assuming a sale value or perpetual service. Dates and debt; the finite favorable case.

Under the supplied comparison, delaying amortization a year preserves $379.5 million of principal payments inside the borrower before the final date, while adding interest. If that saved cash is retained, the net new funding requirement rises by about $36.764 million—not the whole deferred principal. This is conditional liquidity relief, not an extension of maturity or proof that the cash can be distributed. The deferral comparison.

These findings complement the existing cost and useful-life diagrams. They do not calibrate IREN’s required residuals, value an A100 fleet or turn Austin’s historical expense budget into a current collateral price. Every later service path must include the costs, rights and credits needed to earn its receipts.

A performing contract can be better protection than resale

The CoreWeave DDTL 4.0 case supplies a serious favorable repayment mechanism. The inspected DBRS analysis identifies Meta take-or-pay support and expects full amortization from contracted cash rather than a dependence on selling old hardware. Delivery, availability and service costs still matter. This is attributed underwriting, not current observed compliance or a certificate of every invoice. Financing analysis.

That protection can keep lenders paid even when a buyer underuses capacity or equity earns less than hoped. It is not a general guarantee of every CoreWeave loan. Other structures can require re-contracting before the financing ends. A collateral-recovery stress is a backup case—not an appraisal or the expected route for a performing, amortizing facility.

The later CoreWeave notes closing remains evidence of financing access: $4.2 billion face was issued on September 22, with approximately $3.5708 billion left after purchaser discounts and capped-call cost but before other expenses. It is financing with new claims, not customer income or today's unspent balance. Closing evidence and ownership consequences.

The holder adds another layer of financing. The October 2 holder investigation identifies GECC’s June positions in different CoreWeave loans, a corporate bond and a private fund. Those positions do not inherit DDTL 4.0’s Meta-specific terms. A service can remain useful while the investor holding its loan faces a collateral or funding constraint of its own. The named holdings and their protection.

The older A100 design and Austin’s historical expense-budget evidence remain favorable context—not appraisals of the selected loans’ collateral. No assumed GECC recovery percentage has been fed back into the IREN continuation or ownership calculations. What the holder evidence changes, without changing the recovery inputs.

Another use needs its actual costs—not automatically a whole new fleet

The favorable continuation case is not a free terminal asset. The choice includes keeping and maintaining suitable existing service, refreshing equipment, adapting the site or transferring the remaining rights. A whole-generation replacement is one scenario, not an automatic requirement whenever a first customer term ends. A sale and continued earnings from the same sold assets cannot both be collected. The keep, replace and transfer tests.

The original recovery study's explicitly assumed replacement case needs about 85.95% paid utilization at its selected price/cost scale. At a hypothetical 20% lower price and unchanged unit variable costs, it would require 117.21%, beyond the modeled capacity. That is not a prediction of repricing: it identifies where “more volume” alone cannot repair the model. Higher throughput, better prices, lower costs or different investment could change it. Finite replacement cases.

The weaker case still sells useful service and produces positive operating cash; it fails the stipulated whole-investment hurdle. Continuing to operate an already-paid asset can therefore be rational even when the original owners never earn their required return. Buying it at a new, lower price is another question.

The later ownership investigation asks that forward question. It charges remaining investment rather than all sunk construction again, and separately identifies the value required from the rest of IREN's portfolio. Its residual requirement is not a finding that the other portfolio is worthless. Its optional renewal assumptions also differ from the earlier original-recovery test. Forward cash and whole-company claim.

A reused facility still pays rent—and its cost budget is not today’s value

Core Scientific’s Austin operation used a facility formerly occupied by Hewlett Packard, under an eight-year headlease. It is observed re-tenanting and continuing service, not completion of Core Scientific’s own second hosting contract. The separate mining-to-colocation conversions preserved useful sites while some old capital was demolished or impaired. Neither case supplies a general GPU resale curve.

At December 2024, Austin’s $157.206m remaining minimum customer payments less approximately $91.2m headlease payments left about $66.0m before other costs across the remaining term. A separate, shorter April 2025–2029 test uses first-quarter direct expense after removing matching power and depreciation. The annualized $21.812m expense remains a calibration—not measured cash. Matched accounts and schedule boundaries.

At a 10% assumed hurdle, flat expenses and zero net cash correction leave $10.427m of present-value room; 3% annual expense growth leaves $5.695m. The anchor is 31 March 2025, assembled using later amended disclosures. Actual cash adjustments, refit costs and financial claims still matter. It is not a current appraisal, full contract value or owner-cash result. The finite budget and omitted later fees.

Do not add the full headlease again when rent is already in direct expenses, or scale Austin’s budget to an owned Childress building by megawatts. Different rights, service costs and dates prevent that shortcut. The claims on continued value.

What would establish a stronger recovery case?

The strongest favorable evidence would join accepted service, net cash collected after credits, final facility/equipment costs, complete ongoing costs and economically viable renewal or transfer. Another announcement of megawatts supplies only part of that chain. A lower benchmark cost per token can help customers without establishing what an existing owner will earn.

The decisive remaining limits include actual cohort collections, current reserve and debt schedules, as-built costs, replacement needs and independently supportable transferable value. A budget is not a final bill; a rating is not actual compliance; a required residual is not an asset price. The Evidence & tests guide explains how these observations differ.

Read the dated capacity investigation for the complete calculations and supplied recovery figure, the financing investigation for payment priority and loss paths, and the ownership investigation for the price-to-remaining-cash bridge. None has been moved into History. All downloads remain in the complete research catalogue.

For the new evidence and full boundaries, read Older capacity can earn again—but it does not become free (29 September). Its source-linked historical budget complements rather than replaces the preserved first-term recovery and forward-ownership investigations.

Search published pools, pages, reports, and evidence.