Silicon Lien · Research · Data & trackers
What Is a Financed GPU Actually Earning? — July 2026
Dated note · August 11, 2026
A GPU can be active without producing customer revenue, and it can earn contracted revenue while sitting idle. A lender therefore cannot infer current earnings from a utilization percentage alone. The useful record connects the financed devices to contracts, invoices and cash receipts.
Summary
A public rental rate and assumed availability and utilization figures can produce an annualized billing illustration. Establishing actual earnings requires separate availability, activity, usage-billed and fixed-payment measures, borrower billing records and cohort-level allocation. Telemetry supplies operating evidence; it does not prove billability or collection by itself.
A simple billing illustration
Let’s consider one NVIDIA H100 SXM 80GB data-center GPU. This model appears across the rental sources we track and has dated first-party price references in public archives.
The illustration assumes a $2.97 advertised hourly rate, 100% service availability, 70% usage-billed utilization and no fixed-capacity payment:
Usage billings = hourly rate × calendar GPU-hours × service availability × usage-billed utilization
$2.97 × 8,760 hours × 100% × 70% = approximately $18,200
This answers one narrow question: what would one GPU bill in a year if the stated rate, availability and usage held constant? It does not establish the borrower’s realized price, the amount collected or the cash remaining after costs.
Start with the price actually realized
Our GPU rental data page provides external reference points. On July 30, our captures showed $2.69 for RunPod’s lowest H100 SXM on-demand offer, $2.97 for the daily median of four Vast.ai platform-median captures and $6.16 for CoreWeave’s catalog rate. The products and statistics differ, so they frame the advertised market rather than establish one borrower’s revenue.
Public archives provide older reference points: in May 2023 CoreWeave listed a $4.76 configurable GPU-component rate and advertised HGX H100 “from $2.23/hr”; in August 2023 Lambda listed H100 SXM at $2.59. These are point-in-time observations for different configurations and terms, not a price index or depreciation curve.
The deal-specific price comes from executed contracts, discounts, credits and invoices allocated to the financed cohort. Public prices are useful tests of plausibility, not substitutes for that record.
Define what the utilization figure measures
Calendar GPU-hours are the financed device count multiplied by the hours in the period. Available GPU-hours are the subset during which those devices were in service and capable of accepting eligible work. The lender should keep four measures separate:
Service availability = available GPU-hours ÷ calendar GPU-hours
Physical activity ratio = workload-active GPU-hours ÷ available GPU-hours
Usage-billed utilization = usage-billed GPU-hours ÷ available GPU-hours
Allocated fixed-payment coverage = nonduplicated capacity-hours subject to enforceable fixed payment ÷ calendar GPU-hours
Availability must remain visible: excluding downtime from the denominator without reporting it would overstate the fleet’s earning capacity. Fixed-payment hours should be allocated once to the financed cohort; overlapping commitments should be reported as overcommitment rather than counted twice. Service credits and termination rights also affect whether apparently contracted capacity is economically covered.
Device and scheduler telemetry can evidence which GPUs existed, were available and showed workload activity. Activity counters alone do not establish useful customer work or billability. Internal or synthetic jobs can keep a GPU active without producing customer revenue, while take-or-pay capacity can earn revenue when the device is idle. NVIDIA’s DCGM documentation likewise distinguishes hardware activity from effective use.
Reconcile operations with the commercial record
Contracts establish pricing and payment obligations. Scheduler and device records show capacity and activity. Invoices show charges and credits. Cash receipts show collection. Recognized revenue, invoicing and collection are different events and should be reconciled rather than collapsed into one utilization measure.
| Question | Evidence required |
|---|---|
| Did the financed GPU exist and remain available? | Original financed-fleet list, asset identity, in-service dates, maintenance and outage records, device telemetry. |
| Was capacity used or contractually covered? | Scheduler records, usage-billed hours, fixed-capacity terms, service credits, termination rights and nonduplicated cohort allocation. |
| What price applied? | Executed contracts, rate schedules, discounts, credits and invoices. |
| Was the amount collected? | Invoice-level cash receipts, receivable aging and cohort allocation. |
Customer concentration and overlapping relationships among customers, suppliers and capital providers can affect collectibility. The lender should identify who ultimately pays and whether the customer depends on financing from another participant in the same ecosystem. Those facts belong in explicit collection and renewal scenarios; they do not negate properly reconciled current-period cash.
Sources
Data and market observations through July 30, 2026; sources verified through August 10, 2026.
- Silicon Lien daily captures — source-specific rents observed on July 30, 2026.
- CoreWeave pricing page, May 2 2023 and Lambda pricing page, Aug. 31 2023 — archived first-party rates.
- NVIDIA DCGM profiling documentation — definitions and limits of device-activity metrics.
General information from cited public sources, provided without warranty and not investment, legal, accounting, tax, appraisal, valuation, rating, or credit advice. Advertised rental rates are not realized revenue, and the annualized-billings illustration is not a forecast. Company and product names are used only for identification; no affiliation or endorsement is implied. Corrections: info@siliconlien.com.