In March 2025, CoreWeave's IPO S-1 disclosed that NVIDIA held an equity stake in the business (around 1.21%) and served as both a major customer and a capacity backstop. The implication for anyone renting GPUs from a vendor-financed neocloud is that they are paying a rate shaped not just by market supply and demand, but by the margin structure of a bilateral financing deal between two vendors. That deal is sitting inside every H100 invoice from the affected providers.
This post breaks down how the backstop model works, what the unit economics look like from the inside, and what it means practically when you are choosing where to run AI workloads.
What the Backstop Model Actually Is
NVIDIA's original involvement with neoclouds started as demand guarantees. The structure was straightforward: NVIDIA would commit to purchasing back unsold GPU capacity from a neocloud if customer demand fell short. This protected the neocloud from the downside of building out a cluster that nobody rented, and it let NVIDIA push data center GPU adoption faster than natural demand would allow.
The model has since evolved in two directions. First, NVIDIA started taking equity stakes rather than just providing demand floors. CoreWeave's IPO S-1 makes this explicit: NVIDIA is both the primary hardware supplier and a party with equity in the business it supplies. Second, through a separate cloud revenue-share program, some neoclouds now pay NVIDIA a percentage of cloud revenue as a return on the financing it provided, beyond the one-time hardware transaction.
Nebius, which relisted on Nasdaq in 2024 after spinning out of Yandex's cloud infrastructure division, also relies on continued large-scale NVIDIA GPU sourcing as the core of its capital model. The specific equity percentages and revenue-share terms in these arrangements are not always publicly disclosed with precision, but the structure is: NVIDIA sells the GPUs, extends financing or guarantees, and receives a return tied to how well those GPUs are rented out.
This is different from a normal hardware purchase. When a data center buys servers from Dell, Dell does not receive a share of the revenue those servers generate. With NVIDIA and certain neoclouds, the vendor relationship does not end at the hardware transaction.
Neocloud Unit Economics
The math that determines whether a vendor-financed neocloud is profitable at any given moment is worth understanding, because it explains a lot of pricing behavior.
Consider a 1,024-GPU H100 SXM5 cluster. Here is a rough cost model based on industry figures:
| Cost Component | Monthly Cost Estimate |
|---|---|
| GPU capex amortization (3-year, ~$30k/GPU) | $853,000 |
| Power (1,024 GPUs at ~700W, $0.08/kWh) | $38,000 |
| Colocation and networking | $75,000 |
| Financing overhead and interest | $120,000 |
| NVIDIA revenue-share obligation (est.) | $50,000-100,000 |
| Total monthly cost | ~$1,136,000 - $1,186,000 |
Revenue at different utilization levels, assuming a competitive market rate of ~$2.02/hr per GPU:
| Utilization | Revenue/Month | Profit/Loss |
|---|---|---|
| 50% | $754,000 | ~-$432,000 |
| 70% | $1,056,000 | ~-$130,000 to -$80,000 |
| 75% | $1,131,000 | ~break-even |
| 90% | $1,357,000 | ~+$170,000 to +$220,000 |
At 70% utilization, the cluster is running near a loss, roughly -$80,000 to -$130,000/month depending on financing overhead. At 90%, it generates meaningful profit of around $170,000 to $220,000/month. The revenue-share obligation to NVIDIA sits inside the cost stack regardless of utilization, so when clusters run light, it is a fixed drag that shrinks the cushion.
This math matters to you as a buyer for one reason: a neocloud with a cluster running at 55-65% utilization has a strong financial incentive to push on-demand prices up, enforce minimum commitments, or steer customers toward long-term reserved contracts. The pricing you see is partly a reflection of how full their clusters are.
For context on how GPU supply constraints interact with this utilization pressure, see the GPU shortage in 2026 analysis, which covers why new capacity is typically committed before it reaches the spot market.
Circular Financing 101
The pattern has a name that fits: circular financing. NVIDIA sells GPUs to CoreWeave on vendor terms. CoreWeave rents those GPUs to AI teams. CoreWeave pays NVIDIA a portion of cloud revenue as a return on the financing. Some of the money that flows from AI teams to CoreWeave flows back to NVIDIA.
The same dollar cycles. NVIDIA manufactures the hardware, finances the buyer, and then participates in the revenue that hardware generates. From a pure business model perspective, it makes NVIDIA's economics much stronger than a simple hardware seller: they capture margin at the manufacturing stage, at the financing stage, and at the ongoing revenue stage.
Nebius fits the same picture. After relisting in 2024, Nebius's growth model depends on continued access to large-scale NVIDIA GPU allocations at favorable terms. Their ability to compete in the European market is tied directly to their relationship with NVIDIA's supply chain. Any shift in that relationship would immediately change their capacity and pricing.
CoreWeave's case is the most documented because of the IPO S-1 disclosures. The filing showed NVIDIA as both a top customer (a major portion of CoreWeave's early revenue came from NVIDIA-backed compute demand) and a significant shareholder. That dual role, customer and equity owner, is the clearest illustration of the circular structure.
The io-fund and MLQ communities have both analyzed this pattern in detail. The conclusion is consistent: the financing relationships between NVIDIA and the largest neoclouds are not arm's-length supplier deals. They are co-investment structures where NVIDIA has a stake in how much the GPU rental market grows, and the neoclouds have pricing obligations that reflect that arrangement.
What This Means for AI Teams Renting Compute
The circular structure creates three practical effects for teams that rent from vendor-financed neoclouds.
Pricing pressure from both directions. When utilization is high, neoclouds can compete aggressively on price because the cluster is generating surplus revenue above the financing floor. When utilization drops, they need to hold or raise prices to service the cost stack. AI teams sometimes notice this as "pricing volatility" or opaque rate changes. It is not random. It reflects the utilization-driven profitability math described above.
Lock-in serves the cluster, not the customer. CoreWeave's 60% discount for 3-year commitments is real. It is also a mechanism that converts utilization uncertainty (from the provider's perspective) into guaranteed revenue. The discount is compensation for accepting a switching cost that protects the cluster's economics. If you take the discount, you are essentially contributing your demand to their utilization floor. That is a fair trade if you genuinely need 3 years of stable compute. It is not a fair trade if your workload demand might change significantly in 12 months.
Provider financial stress flows downstream. A neocloud carrying debt-financed clusters at below-break-even utilization is under financial pressure. That pressure shows up as capacity rationing, lower responsiveness for non-enterprise customers, and sudden pricing changes. The CoreWeave vs Spheron comparison documents specific pricing differentials that trace back to this cost structure. And the broader CoreWeave alternatives analysis shows how this has pushed the market toward providers without these obligations.
Independent Alternatives: What to Look For
The characteristics that protect you from circular-financing pricing dynamics are specific and checkable.
1. Multi-provider aggregation
Providers that aggregate capacity from independent data centers are not themselves carrying NVIDIA-vendor-financed capex. They did not borrow from NVIDIA to build out. Their pricing reflects competition between the data centers they partner with, not the terms of a bilateral revenue-share agreement with their hardware supplier.
Spheron aggregates bare-metal GPU capacity from 5+ independent data center providers. Its pricing is set by competition between those suppliers. There is no NVIDIA revenue-share obligation in the cost structure.
2. Spot availability as a signal
A healthy spot market from a provider means they have excess capacity priced at market rates rather than rates calibrated to service a financing obligation. Absence of spot pricing, or spot pricing that is unusually close to on-demand, is a signal that available capacity is already committed to covering fixed costs.
3. No long-term commitment required for competitive rates
If reaching a price close to market rate requires a multi-year contract, the provider is pricing their financing obligations into the standard on-demand rate and offering discounts to convert your demand into a utilization guarantee. Look for providers where the lowest rate is also the pay-as-you-go rate.
4. Transparent pricing from market competition
Providers whose prices derive from competition between multiple supply sources have different incentives than providers whose prices are constrained by a vendor financing structure. You can usually tell the difference: providers with circular financing obligations tend to cluster on-demand rates in a narrower band, move prices less frequently, and offer steeper discounts for long-term commitment than for short-term. Providers driven by supply competition show more price movement, cleaner spot availability, and no meaningful premium for month-to-month flexibility.
For a full cross-provider pricing comparison, see GPU cloud pricing comparison 2026. For the broader picture of where neoclouds fit in the stack alongside inference platforms and MLOps tools, see AI infrastructure companies in 2026.
Current On-Demand and Spot Rates on Spheron
Here is pricing from the Spheron API fetched on 04 Jul 2026. On-demand rates are derived from DEDICATED offers only. Spot rates are from SPOT offers using spot_price / gpuCount.
| GPU Model | Spheron On-Demand (/hr) | Spheron Spot (/hr) | CoreWeave On-Demand (est.) |
|---|---|---|---|
| H100 SXM5 | $2.54 | $1.43 | ~$6.16 |
| H100 PCIe | $2.01 | N/A | $4.76 |
| H200 SXM5 | $3.70 | $3.31 | varies |
| A100 80GB SXM4 | $1.69 | $0.79 | $2.21 |
| B200 SXM6 | $7.50 | $2.74 | limited availability |
Spheron's H100 SXM5 on-demand rate is $2.54/hr versus CoreWeave's H100 SXM5 rate of approximately $6.16/hr. The gap at spot is even larger: $1.43/hr versus CoreWeave's spot rates, which are not publicly listed but estimated at $2.00-2.50/hr by industry trackers.
Pricing fluctuates based on GPU availability. The prices above are based on 04 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
For a full H100 on-demand and spot breakdown including other providers, see H100 on Spheron.
Buyer Checklist Before Committing to a GPU Provider
Six questions to answer before signing up with any neocloud:
- Utilization transparency. Ask the provider for their average cluster utilization floor and what happens to pricing if it drops below that floor. Legitimate providers can answer this. Providers with constrained financing structures typically won't.
- Spot market availability. Check whether the provider publishes spot pricing. A healthy spot market is evidence of excess capacity priced at fair market value. A provider with no spot market, or spot prices very close to on-demand, has all capacity committed to covering fixed costs.
- Exit terms. Before any commitment, read the cancellation clauses, notice periods, and minimum terms. Commitments that offer meaningful discounts but impose 90+ day exit periods are built around the provider's utilization math, not your flexibility.
- Billing granularity. Per-minute billing versus per-hour versus daily minimum dramatically affects cost on short workloads. A 20-minute job on a per-hour provider costs the same as a 60-minute job. On per-minute billing, it costs one-third as much.
- Vendor financing disclosure. Ask directly whether the provider has equity or revenue-share arrangements with NVIDIA or any other GPU manufacturer. This is a material disclosure that affects how pricing is set. Providers without such arrangements can say so clearly.
- Stranded capacity risk. Providers with large forward-contracted clusters that miss their utilization target tend to respond with capacity rationing and rate changes for non-enterprise customers. Look at how long the provider has been operational and whether they have navigated demand cycles without major pricing restructuring.
The Bottom Line
Vendor-financed neoclouds are not inherently bad. They took NVIDIA's initial risk of pushing data center GPU supply into the market, and they built meaningful infrastructure. The problem is that the financing structure they agreed to is now embedded in their pricing, and when you rent from them, you are paying for it.
The price gap between a vendor-financed neocloud and an independent aggregator like Spheron is not purely a result of different margin targets. It reflects a structural difference in what the pricing has to cover: NVIDIA revenue-share obligations, debt service on vendor-financed capex, and the utilization floor required to stay solvent.
Teams that understand this structure can price-shop accordingly. You're not comparing apples to apples when you put a vendor-financed neocloud rate next to an independent aggregator's rate. The independent aggregator's price genuinely reflects the market cost of GPU capacity. The neocloud's price reflects market cost plus the terms of a bilateral financing arrangement you were not part of.
Spheron aggregates GPU supply from independent data centers, with no NVIDIA vendor-financing obligations baked into its prices. Check live H100 and B200 rates to see the difference.
Frequently Asked Questions
NVIDIA provides capital or guaranteed purchase commitments to neocloud providers (CoreWeave, Nebius, etc.) in exchange for a share of cloud revenue. This means NVIDIA both sells the GPUs and participates in the revenue generated by renting them back to end users, creating a circular financing structure.
A 1,024-GPU H100 cluster at competitive market rates requires around 75% utilization to cover costs. Near that threshold, at 70% utilization, the cluster runs at a loss of roughly $80k to $130k/month. At 90% utilization, the same cluster can generate $170k to $220k/month in profit. The math is highly sensitive to GPU rental rates, power, and colocation costs.
Yes. Neoclouds financed by NVIDIA have locked-in cost structures tied to vendor revenue-share terms. Their on-demand prices carry a margin that covers both debt service and NVIDIA's revenue cut. Independent aggregators without vendor-financed capex can price closer to market because they are not servicing those obligations.
Circular financing is the pattern where NVIDIA provides capital or commitments to neoclouds, which buy NVIDIA GPUs and rent them out, with NVIDIA receiving a revenue share. The same money flows from NVIDIA to the neocloud and partially back to NVIDIA. CoreWeave's IPO S-1 and Nebius disclosures illustrate this dynamic explicitly.
Use providers without long-term vendor financing agreements. Look for per-minute billing with no minimums, no contractual commitment to a single provider, transparent pricing derived from market competition between multiple supply sources, and no NVIDIA revenue-share clauses in their infrastructure agreements.