NBIS — Durable Neocloud Unit-Economics Claims and Market Proxies
Research cutoff: August 2, 2026.
Currency: U.S. dollars unless stated otherwise.
This memo compares four Modeledge company dossiers—CoreWeave, Nebius, Boost Run, and SharonAI—to isolate claims that remain true across different operators, accounting presentations, contract structures, and capital stacks. “Consistently true” means supported by at least three companies or by two companies plus a directly observable market proxy. It does
not mean the companies use identical definitions.
Primary synthesis sources:
Executive summary
- The durable economic unit is the contract cohort, not the quoted GPU-hour. A useful underwriting unit combines GPU/facility capex, accepted service capacity, term price, billed hours, customer prepayment, direct operating cost, useful life, financing, and post-contract value.
- Take-or-pay converts usage risk into delivery, acceptance, counterparty, and recontracting risk. It can make billed utilization approach 100% during the committed term, but it does not prove physical GPU utilization, successful commissioning, collectability, or residual value.
- Long-duration capacity consistently clears below public on-demand pricing. Disclosed or reconstructed B300/GB300 long-term rates cluster around roughly $3.15–$3.97 per GPU-hour, versus Nebius’s $7.85 B300 on-demand price and $4.30 preemptible price. The observations are not perfectly comparable, but the discount direction is robust.
- Management payback claims cluster near two to three years; fuller reconstructions often move toward three to four years. CoreWeave stated about 2.5 years, Nebius stated 2.5–3.0 years for Hopper/H200, and SharonAI illustrated 2.1 years. Scope-corrected cases reached about 3.0 years for CoreWeave and 3.77 years for SharonAI. Blackwell/Rubin cohort payback remains largely undisclosed.
- Long-run returns are dominated by five variables: realized contract price, all-in capex scope, useful/economic life, direct facility and power cost, and the price/utilization achievable after the initial contract. Spot pricing is useful as a market-clearing and residual-value proxy, but it should not replace the contracted first-term cash flows.
1. The consensus map
| Claim that survives comparison | CoreWeave | Nebius | Boost Run | SharonAI | Confidence |
|---|
| The true unit is a contract/deployment cohort, even when billing is per GPU-hour | Five-year reserved clusters priced per GPU-hour | Accepted GPU-service tranches, active MW, GPU-hour, token/workload | Per-deployment underwriting tied to GPU, site costs, lease structure, price and term | GPU-hour, MW, cluster and contract economics | High |
| Multi-year commitments trade price for utilization/revenue certainty | Five-year rate below on-demand; take-or-pay | Microsoft/Meta fees payable irrespective of use after acceptance | One-to-three-year discounts of 10%–40%; disclosed take-or-pay orders | Five-year targets; ESDS $3.30 versus $4.50 illustration | High |
| “100% utilization” usually means billed reserved capacity, not measured physical occupancy | Explicit management framing | Accepted tranches paid irrespective of usage | Full fixed fees regardless of use in disclosed orders | Payment for reserved capacity after acceptance | High |
| Two-to-three-year payback is the management underwriting zone | 2.5 years, Dec. 2024 vintage | 2.5–3.0 years for Hopper/H200 | No numeric payback disclosed | 2.1-year illustrative B300 case | Medium-high |
| Full-scope cash economics are weaker than headline gross/EBITDA measures | Skeptical unlevered case about three years; financing and lease tail matter | 45% core adjusted EBITDA excludes D&A and financing | 85.6% Q1 gross margin becomes about −0.1% after colo and D&A | 2.1-year illustration becomes 3.77 years using later full-project capex | High |
| Depreciation life materially moves GAAP earnings but not cash returns | Six-year life is a major swing factor | Changed four years to five; Q1 D&A reduced $43.1m | Four-year computer-hardware life | Filing life two-to-five years versus six-year deck life | High |
| Clean power cost and all-in capex are poorly disclosed | No current all-in $/GPU or clean energized-MW bridge | No company-wide all-in $/GPU or pure $/MWh | Colo bundles power; no clean tariff | ESDS pass-through exists, but base tariff is undisclosed | High |
| Customer concentration and financeability are part of unit economics | Anchor customers and Nvidia support determine collateral quality | Microsoft/Meta scale, prepayments and credit reduce financing risk | Customer concentration and lease funding are material | ESDS credit/security is central despite attractive rack math | High |
| Corporate free cash flow can be deeply negative while individual cohorts are profitable | Expansion and refinancing dominate | $20–$25bn 2026 capex versus $3.0–$3.4bn revenue guidance | Lease-prepayment-inclusive Q1 FCF was negative | Growth financing and commissioning precede revenue | High |
2. The most consistently true claims
Claim 1 — Contract-cohort cash yield is the correct economic denominator
All four operators market capacity in multiple units—GPU-hour, MW, cluster, contract, and increasingly token or completed workload. None of those units alone captures investor returns. The economically complete unit is:
>
Cash contributed by one accepted deployment over its contracted and residual life ÷ all equity and debt-funded cash required to procure, commission, operate, maintain, and retire or redeploy it.CoreWeave describes a five-year fixed-price reservation; Nebius’s Microsoft and Meta agreements are accepted service tranches; Boost Run says deployment-specific GPU type, site cost, hardware/lease structure, price, and term determine underwriting; SharonAI publishes both per-MW and per-cluster cases. The common economic object is therefore a
deployment tied to a customer contract, not a fleet-wide average GPU price.
For NBIS, the missing bridge is a cohort table with GPU count/SKU, active MW, contract price, service start, customer prepayment, direct opex, capex, useful life, and terminal assumption. Without it, a provider IRR is not independently reconstructable.
Claim 2 — Take-or-pay removes usage risk only after delivery and acceptance
The phrase “100% utilization” is consistently used as a billing or reservation concept. CoreWeave described a five-year commitment as economically equivalent to 100% utilization because the customer must pay for each contracted hour. Nebius’s Microsoft and first Meta agreements require payment for accepted tranches irrespective of actual use. Boost Run’s Thinking Machines orders require full fixed fees regardless of use. SharonAI targets payment for 100% reserved capacity after acceptance.
This shifts, rather than eliminates, the risk stack:
- Before acceptance: construction, interconnection, GPU delivery, commissioning, SLA testing, and delay penalties.
- During the term: customer credit, service credits, availability, contract termination rights, power pass-throughs, and financing covenants.
- After the term: recontract price, physical utilization, remaining useful life, lease tail, and residual hardware value.
The most defensible wording is:
take-or-pay can secure billed utilization during the enforceable term, subject to delivery, acceptance, performance, and counterparty terms. It does not demonstrate physical utilization or terminal value.
Claim 3 — Duration earns a discount; scarcity earns a premium
The clearest agreement across the dossiers is that long-term capacity prices below on-demand or planning rates:
- CoreWeave says its five-year committed rate is below on-demand, although the exact contract rate is undisclosed.
- Nebius’s Microsoft price is redacted. A high-uncertainty reconstruction using the filed $17.3929bn contract value and a third-party estimate of 100,000 GB300 GPUs implies about $3.97/GPU-hour, roughly 49% below Nebius’s June 2026 B300 on-demand price of $7.85.
- Boost Run disclosed or implied B300 contract prices of about $3.15, $3.45, and at most $3.59/GPU-hour, approximately 30%–38% below its $5.10 planning start rate.
- SharonAI’s executed ESDS B300 contract is $3.30/GPU-hour, 26.7% below its earlier $4.50 illustrative rate.
The exact magnitudes should not be averaged: bundles, GPU generations, prepayments, terms, locations, SLAs, and start dates differ. The consistent claim is directional:
the customer buys price protection and dedicated availability; the provider buys contracted utilization and financeability.Claim 4 — Two-to-three-year payback is the promoter case, not a universal realized fact
Company-stated or management-framed paybacks cluster tightly:
| Company | Stated case | Important scope limitation |
|---|
| CoreWeave | About 2.5 years | December 2024 average; includes customer prepayments; uses Adjusted EBITDA to recover GPU and associated PP&E, not all debt/leases/corporate overhead |
| Nebius | 2.5–3.0 years | Hopper/H200 estimate from February 2025; Blackwell payback explicitly not yet stated |
| Boost Run | UNKNOWN | No disclosed cluster payback or IRR |
| SharonAI | 2.1 years | October 2025 illustrative B300 case using $4.50/hour, 90% utilization, 77% gross margin, and six-year life |
Fuller cross-checks are less aggressive:
- A third-party CoreWeave unlevered-FCF reconstruction produced roughly three years with zero terminal value.
- SharonAI’s later $733m ESDS project capex and exact $1.264bn contract value imply a 3.77-year sensitivity if the earlier 77% margin is retained; this is still before interest, overhead, and tax.
- Nebius’s hardware-only GB300 gross payback can appear below 1.5 years using public prices and a third-party rack cost, but the result omits facility, network, electricity, maintenance, software, financing, and decay and therefore is not a provider payback.
The consistent underwriting conclusion is a range, not a point:
two-to-three years is the stated deployment target; three-to-four years is a plausible full-scope outcome when capex, facility cost, and conservative terminal assumptions are included. New Blackwell/Rubin vintages require separate evidence.
Claim 5 — Capex scope matters at least as much as price
Headline capex per GPU is not comparable unless the scope is identical. The dossiers show the problem directly:
- SharonAI’s October illustration used $58,180 per B300 GPU, including server, networking, and setup. Its later ESDS project estimate was $89,303 per GPU, 53.5% higher, because the project includes broader storage, network, and deployment scope.
- Boost Run’s cross-filing proxy is roughly $85,500–$102,600 per planned GPU, while a third-party 576-B300 cluster benchmark is about $62,500/GPU and a greenfield all-in benchmark reaches about $133,000/GPU.
- Nebius does not disclose company-wide all-in capex per GPU. A third-party GB300 rack estimate is $41,667/GPU, but it is a rack figure—not a powered, commissioned, financed service.
- CoreWeave likewise does not disclose a current cohort-level all-in $/GPU or $/MW.
Any comparison should explicitly label whether capex includes the GPU server, network fabric, storage, spares, facility shell, electrical and cooling infrastructure, interconnection, capitalized software, construction interest, and customer-specific fit-out.
Claim 6 — Headline margin is not cohort return
Reported margins differ mostly because cost perimeters differ:
- CoreWeave targets a mid-20% contribution margin on a base five-year infrastructure contract, but the public bridge to depreciation, allocated data-center rent, interest, and unlevered cash return is incomplete.
- Nebius reported 45% adjusted EBITDA margin for its core AI cloud in Q1 2026 and 32% for the group, but group GAAP operating loss was $128.0m on $399.0m revenue and Q1 D&A was $212.0m.
- Boost Run reported 85.6% Q1 gross margin, yet subtracting separately presented colocation and D&A produces approximately −0.1% infrastructure contribution before SG&A. FY2025 similarly moves from 85.5% reported gross margin to 26.8% after those costs.
- SharonAI’s 77% gross-margin illustration supports a 2.1-year headline payback only under the narrower illustrative cost base; later capex and contract evidence materially lengthen it.
The durable rule is:
compare margin only after reconciling power, colo/rent, maintenance, network/storage, depreciation, lease cost, and financing. Adjusted EBITDA remains useful for operating traction but is not an asset IRR.
Claim 7 — Useful life is the largest accounting and residual-value swing factor
The companies use or discuss lives ranging from four to six years:
- CoreWeave uses six years for computing equipment; third-party sensitivities show sharply lower margins at four or five years.
- Nebius changed server/network life from four to five years beginning in 2026. The change reduced Q1 D&A by $43.1m and increased net income by $41.6m without improving cash economics.
- Boost Run uses a four-year computer-hardware life.
- SharonAI’s filing indicates two-to-five years for computer equipment, versus a six-year economic life in its investor illustration.
A longer accounting life raises GAAP EBIT, but only actual recontracting, maintenance, failure rates, and realized residual prices determine economic life. For a five-year initial contract, residual value should be shown as a separate sensitivity rather than embedded invisibly in the base case.
Claim 8 — Financing terms are part of the product economics
Take-or-pay contracts, prepayments, investment-grade customers, OEM credit, leases, and asset-backed debt reduce the equity check but can also create refinancing and lease-tail risk.
CoreWeave’s December 2025 active contracts carried weighted-average customer prepayments of 15%–25% of total contract value, and its March 2026 debt table implied an approximately 9.1% weighted effective cost of debt. Nebius disclosed roughly $6.96bn upfront on its Microsoft agreement and uses large customer commitments to finance expansion. Boost Run often obtains up-front cash and matches customer terms with 30–36-month GPU leases, but excluding lease prepayments can make free cash flow look materially stronger. SharonAI’s financeability depends on customer security, acceptance, and funded project structures; attractive gross rack math alone is insufficient.
The correct equity return calculation must include prepayment timing, debt amortization, lease payments, interest, hedging, and residual obligations. A high asset IRR can coexist with a fragile equity story if refinancing, customer concentration, or lease tail is ignored.
Claim 9 — Growth free cash flow and cohort profitability can diverge for years
Nebius’s 2026 capex guidance of
$20–$25bn is roughly
5.9×–8.3× its $3.0–$3.4bn 2026 revenue guidance, because spending supports later capacity. CoreWeave similarly carries large contracted-versus-active power gaps, leases, and financing needs. Boost Run’s Q1 company-style FCF margin was 39.3%, but including operating- and finance-lease prepayments produced approximately
−24.8%. SharonAI must fund and commission projects months before revenue acceptance.
Corporate FCF is therefore a poor short-run test of whether a single accepted cluster earns an attractive return. The reverse is also true: a good cohort model does not prove the corporate capital structure is safe.
3. Market proxies: what is actually available
Public market proxies are available and useful, but only for three narrow purposes:
- observing the current scarcity premium between on-demand and interruptible capacity;
- estimating the price at which post-contract or uncommitted capacity might clear; and
- stress-testing whether a residual-value assumption is plausible.
They are
not substitutes for confidential reserved-contract pricing or a complete provider margin.
B300/GB300 observations
| Observation | $/GPU-hour | Date / type | Interpretation |
|---|
| Nebius B300 on-demand | $7.85 | Effective June 1, 2026 | Scarce, flexible list price |
| Nebius B300 preemptible | $4.30 | Effective June 1, 2026 | Best public NBIS spot-like proxy |
| CoreWeave HGX B300 spot | $4.48 | Observed August 2, 2026 | Interruptible competitor proxy; on-demand was quote-only |
| Boost Run disclosed B300 contracts | $3.15–$3.59 | Nov. 2025–May 2026 | Multi-year/bundled contract observations |
| SharonAI ESDS B300 contract | $3.30 | March 31, 2026 | Five-year nominal service order; only 36 months fully non-terminable before break terms |
| Nebius Microsoft GB300 reconstruction | $3.97 | High-uncertainty sensitivity | Uses filed value and third-party 100,000-GPU estimate; not a disclosed rate |
| SharonAI B300 illustration | $4.50 | October 2025 | Planning case, not realized result |
| Boost Run B300 planning start rate | $5.10 | April 2026 prospectus | Forecast assumption, not market clearing |
Observed pattern: flexible B300 capacity is priced around $4.30–$7.85, while disclosed/reconstructed long-duration capacity clusters around roughly $3.15–$3.97. The narrowest defensible conclusion is that
current long-term B300/GB300 pricing is materially below on-demand and broadly near or below spot/preemptible pricing after allowing for bundles and contract differences.Other current spot-like observations
| GPU | Nebius preemptible | CoreWeave spot | Silicon Index | SemiAnalysis composite | Observed range |
|---|
| H100 | $2.15 | $2.46 | $2.77 | $2.82 | $2.15–$2.82 |
| H200 | $2.45 | $2.62 | $3.10 | n/a in source set | $2.45–$3.10 |
| B200 | $3.95 | $4.26 | $5.66 | $3.68 | $3.68–$5.66 |
Dates differ: Nebius and CoreWeave prices were current around June–August 2026; Silicon Index was observed July 30, 2026; the SemiAnalysis composite was April 2026. Service quality, location, network, CPU/storage bundles, availability, and minimum size differ.
For context, public on-demand prices were materially higher: Nebius listed H100/H200/B200 at
$3.85/$4.50/$7.15, while CoreWeave listed
$6.16/$6.31/$8.60 per GPU-hour. This confirms the scarcity/flexibility premium but does not disclose reserved contract margins.
Price decay is real but not monotonic
The source set contains both decay and rebound evidence:
- A J.P. Morgan market series showed neocloud H100 pricing falling from about $2.70 in November 2024 to $2.00 in December 2025, a roughly 26% decline.
- SemiAnalysis showed an H100 one-year term rebounding from about $1.70 in October 2025 to $2.35 in March 2026, roughly 38%.
- Nebius raised displayed June 2026 on-demand prices by roughly 29%–31% for B300/B200/H200/H100 and preemptible prices by roughly 27%–72% from the prior displayed levels.
- CoreWeave said its own mature-SKU average pricing was more stable than broad market series because fixed reservations and bundled service reduce exposure to spot repricing.
The long-term claim is therefore not “GPU prices always fall.” It is:
hardware generations commoditize, but supply, power, network quality, contract structure, and short-run scarcity can create large cyclical rebounds. A conservative model should stress both secular decay and temporary scarcity rather than extrapolating either in a straight line.
How NBIS should use these proxies
For Nebius, the best observable pricing ladder is:
- On-demand list: ceiling-like flexible pricing, currently $7.85 for B300.
- NBIS preemptible: $4.30 for B300; the cleanest public proxy for immediately available interruptible NBIS capacity.
- Competitor/market spot: CoreWeave B300 $4.48 and the H100/H200/B200 cross-provider bands above.
- Long-term contracted/reconstructed: roughly $3.15–$3.97 for B300/GB300 observations across Boost Run, SharonAI, and Nebius’s Microsoft sensitivity.
The correct monitoring question is not whether NBIS list price holds. It is whether the
market-clearing spot/preemptible band remains above the cash operating cost and supports an acceptable recontract return after the initial take-or-pay term.4. NBIS-specific underwriting implications
What is demonstrably attractive
- Nebius has secured large, long-duration commitments from Microsoft and Meta, including substantial upfront cash and fees payable irrespective of actual usage after acceptance.
- Core AI cloud adjusted EBITDA margin reached 45% in Q1 2026 on $389.7m revenue, indicating operating leverage in the active base.
- Management reports three to four customers competing for each GPU and sells a mix of long-term, short-term, and spot capacity.
- Nebius’s public infrastructure has a current premium-priced on-demand channel and a visible preemptible channel, creating at least some price-discovery mechanism.
- In-house data-center/hardware design and software/workload optimization can improve customer TCO and potentially support more revenue per GPU than bare-metal rental.
What is not yet demonstrated
- Blackwell/GB300 or Rubin cohort payback, IRR, and contribution margin.
- Exact active GPU count and GPU mix by site.
- Realized reserved $/GPU-hour for Microsoft and Meta.
- Mature physical utilization, billed hours, and downtime by cohort.
- All-in capex per GPU/MW and direct electricity/colo cost.
- The bridge from 45% adjusted EBITDA to unlevered cluster cash flow after maintenance capex, D&A, and financing.
- Post-contract residual value or recontract price.
A disciplined NBIS base-case framework
Until those disclosures exist, the most defensible framework is:
| Variable | Conservative treatment | Why |
|---|
| Initial contract revenue | Use filed non-cancelable consideration and service timing; do not substitute headline “commitments” for RPO | Backstops, options, LOIs, and variable consideration are not equivalent |
| Contract price | Use disclosed rate if available; otherwise keep the Microsoft $3.97 sensitivity explicitly high-uncertainty | Contract quantity and price remain redacted |
| Utilization | Treat accepted take-or-pay hours as billed utilization; model physical utilization separately only for variable cost/wear | Prevents mixing billing with telemetry |
| Capex | Include GPU/server, network, storage, spares, site fit-out, electrical/cooling, commissioning, and capitalized interest | Cross-company evidence shows 50%+ variation from scope alone |
| Useful life | Base at five years for GAAP consistency; stress four years economically | Nebius moved from four to five; peer evidence shows life is a major swing factor |
| Residual value | Set to zero in the base IRR, then add spot-linked sensitivities | Avoids making the terminal assumption carry the thesis |
| Residual price | Link to observable NBIS preemptible and cross-provider spot bands, with explicit decay/rebound cases | This is where market proxies are relevant |
| Software/token upside | Value separately until revenue and margin are disclosed | Avoids using customer TCO benefits as provider cash flow without a bridge |
| Financing | Model customer prepayments, debt/lease amortization, interest, and refinancing explicitly | Financeability is a real source of value and risk |
The long-term NBIS question
NBIS’s first-term economics may already be substantially de-risked by hyperscaler contracts and prepayments. The long-term equity question is narrower and harder:
>
After five years of hardware aging, can Nebius recontract or repurpose the same powered cluster at a price above cash operating cost while avoiding a lease/power tail and excessive refresh capex?That answer depends less on today’s $7.85 on-demand B300 sticker price than on the future equivalent of today’s $4.30 preemptible price, observed market spot bands, physical reliability, workload goodput, and the cost of the next refresh.
5. Claims that are not consistently true
The comparison rejects the following shortcuts:
- “Take-or-pay means 100% physical utilization.” It means payment for accepted reserved capacity, subject to contract terms.
- “A 75%–85% gross margin proves an attractive IRR.” Reported gross margin can exclude colocation, depreciation, leases, and financing.
- “Two-year payback is proven for current Blackwell/Rubin cohorts.” Most stated paybacks are older, illustrative, or definition-dependent.
- “Five- or six-year depreciation proves five or six years of economic value.” Accounting life is an estimate; residual cash pricing is the test.
- “Backlog or TCV is funded revenue.” RPO, TCV, optional renewals, backstops, and LOIs must be separated.
- “Spot price is the reserved contract price.” Spot/preemptible capacity is interruptible and reflects immediate availability; dedicated capacity includes different service and credit economics.
- “GPU-hour price alone determines customer value.” Goodput, network performance, recovery time, storage, and software can change workload TCO.
- “A profitable cluster guarantees positive corporate FCF.” Expansion timing and the capital structure can overwhelm cohort cash generation.
- “GPU price decay is monotonic.” The source set shows both multi-quarter decay and sharp scarcity-driven rebounds.
- “Residual value is free upside.” It requires remaining hardware life, powered space, demand, and a clearing price after the first contract.
6. The disclosures that would resolve the debate
The single most useful NBIS disclosure would be an anonymized cohort table, updated quarterly:
| Required field | Why it matters |
|---|
| GPU generation, count, and in-service date | Establishes the asset cohort |
| Active MW and revenue-producing MW | Separates contracted, energized, active, and billed capacity |
| Contracted $/GPU-hour or $/MW, term, and prepayment | Establishes realized first-term revenue and financing |
| Physical and billed utilization | Separates telemetry from take-or-pay billing |
| GPU/server, network, storage, and facility capex | Produces a consistent all-in denominator |
| Electricity, colo/rent, maintenance, and support cost | Produces direct cash contribution |
| Useful life, refresh plan, failure reserve, and salvage policy | Makes residual assumptions visible |
| Debt/lease amount, rate, amortization, and covenants | Converts asset return into equity return |
| Recontracted price and utilization for older cohorts | Tests the residual-value thesis directly |
Until then, the strongest conclusion is deliberately modest:
NBIS has credible evidence of demand, contracted utilization, prepayment-supported financeability, and improving active-base EBITDA. It does not yet provide enough public cohort data to prove Blackwell/Rubin IRRs or post-contract residual economics.Bottom line
Across CoreWeave, Nebius, Boost Run, and SharonAI, the long-run neocloud model is consistently attractive only when three conditions hold together:
- the first contract repays most or all of the fully scoped deployment cost before hardware obsolescence;
- take-or-pay protection survives delivery, acceptance, SLA, and counterparty terms; and
- residual powered capacity can clear above cash operating cost without requiring an uneconomic refresh or leaving a long lease tail.
For NBIS, current public evidence supports the first-term demand and financeability case more strongly than the terminal-value case. Spot and preemptible GPU pricing are available and should be tracked—but as a
recontracting and residual-value stress test, not as a replacement for the contracted first-term economics.