If you reserve GPUs in a neocloud data center still under construction, Nvidia's support for that neocloud starts only once the building is financed and running. Until then, the risk of a late or failed build sits with the neocloud and, through your contract, with you. Nvidia's third-quarter fiscal 2027 CFO commentary, filed November 17, says its neocloud commitments "take effect only if the neoclouds secure financing to purchase the data center infrastructure products and the related compute capacity is deployed, made operational, and accepted by us." Its guarantees work the same way, building by building.
The practical consequence: the backstop many buyers assumed sat behind CoreWeave, Nebius, Nscale and the rest does not cover construction. Your delivery dates, slip remedies and prepayment terms have to do that job.
What Nvidia Actually Disclosed
Nvidia disclosed $32 billion of neocloud agreements and $118 billion of maximum guarantee exposure, and every dollar of both depends on conditions being met first. The figures come from the same CFO commentary, as of October 25, 2026.
The $32 billion line, previously labelled "AI cloud agreements," shows no amount falling in the rest of fiscal 2027. CNBC reported the commitments run through 2032. Under those agreements Nvidia earns revenue on the upfront sale of its infrastructure and, "if certain criteria are met," shares in the revenue neoclouds earn "from their third-party customers." The filing also says that if the contingencies "are not timely resolved, the amount and timing of our commitments may be affected, and certain commitments could be reduced or eliminated."
The $118 billion of guarantees splits three ways in the filing's table:
- $105 billion of SB Energy credit support for about 4.25 gigawatts across nine buildings that will host Nvidia infrastructure under 20-year leases to OpenAI.
- $9.5 billion of power and systems credit support.
- $3.5 billion of land, power and shell guarantees for neoclouds.
Those guarantees "take effect phase by phase, with each one becoming effective only when its related building is completed and ready for service." Nvidia says it has made no payments under any of them.
The filing also contains a sentence any capacity buyer should read twice. Neoclouds and frontier labs are seeing extraordinary demand, it says, "yet many are growing faster than their balance sheets and long-term credit profiles can support." That is the supplier describing its own customers, the same companies you are signing with.
What Kress Said About Delayed Builds
Colette Kress, who is running Nvidia as interim CEO, said on the November 17 earnings call that a guarantee lapses if the build it supports is delayed. According to TechCrunch's account of the call, she said that if a project is "delayed or they have a challenge in terms of building that, our guarantee doesn't exist anymore." She added that Nvidia is "taking a very strong look in terms of the types of companies that we would want to help in terms of the financing."
Kress described this as a clarification of existing terms, not a new policy. That framing deserves weight. Nvidia is not withdrawing from neocloud financing, and TechCrunch notes it still intends to back land, power and shell, only more selectively. CNBC quoted her saying Nvidia will be "rigorously evaluating future investments, commitments, and guarantees."
The timing still matters. This is Nvidia's first report since Jensen Huang resigned as CEO; board chair Harvey Jones told CNBC the board "does not intend to reinstate Jensen as CEO." TechCrunch cites reports that Kress had pushed back on Huang's dealmaking. A supplier under new leadership that says it will be more selective about financing is a supplier whose next round of neocloud support is less certain than the last.
The market read the disclosure as reassuring, and that reading is fair. CoreWeave, Nebius and Iren each rose 8% to 10% after hours, per CNBC. For Nvidia's shareholders, contingent exposure is the safer structure. For a company prepaying a neocloud for capacity that does not yet exist, it means the neocloud's largest partner protected itself first.
Why Construction Risk Is the Risk That Bites
Neocloud capacity has slipped repeatedly over the past year, and the cause is usually the building or the power connection, not the GPUs. Three recent cases show the pattern.
CoreWeave, November 2025. CoreWeave cut its full-year revenue guidance to $5.05 to $5.15 billion from $5.15 to $5.35 billion because a third-party developer was late delivering a powered shell for one data center, as reported at the time. The company said its backlog was unaffected because customers agreed to adjust delivery schedules. Read that from the customer side: the customers absorbed the slip.
Nebius, November 2026. Nebius's Vineland, New Jersey site, which serves a Microsoft contract worth up to $19.4 billion, will deliver just over half its capacity by year-end and the rest in the second quarter of 2027, TechCrunch reported. A Nebius executive said the third-party operator "decided to wind down the temporary generation faster than it was expected." Nebius also expects prepayments plus asset-backed financing to cover up to 85% of its planned $40 to $45 billion of 2027 capex. Customer cash, paid before delivery, is funding the build.
Nscale, October 2026. Microsoft terminated its contract for Nscale's planned Loughton, England site after the power connection was pushed out, Data Center Knowledge reported, citing reporting that the grid operator could not supply enough power until the early to mid-2030s. Nscale's September IPO filing reported $103.4 billion in total contract value against $2.6 billion of active contract value as of August 31, 2026, and no binding financing commitments for the GPUs its Anthropic agreements need.
In each case the large buyer had leverage. Microsoft's Nebius contract, as Nebius summarised it to TechCrunch, lets Microsoft terminate affected service after a grace period if dates are missed and no alternative capacity is offered, and carries financial penalties for late delivery. A mid-size enterprise signing a standard order form rarely gets either term unless it asks.
What This Means for Your GPU Reservation
Treat a reservation in an unbuilt facility as a construction contract with a compute service attached, and price the delivery risk yourself. The Nvidia disclosure removes the assumption that a supplier with a $115.8 billion quarter, per its press release, stands behind your provider while the shell is going up. It stands behind the provider once the capacity is accepted and running.
That changes three negotiating positions.
First, prepayment. A neocloud that funds construction with customer prepayments is asking you to be a lender without a lender's protections. If you prepay, ask for the money to sit in escrow or be released against milestones (powered shell, racks energised, cluster accepted), and for a refund right if the final milestone slips past a fixed date.
Second, delivery remedies. Ask for the Microsoft structure: a named delivery date per site, a grace period measured in weeks, service credits or penalties for each week late, and a termination right without a fee if the provider cannot offer equivalent capacity elsewhere. Equivalent should mean the same GPU generation, interconnect and region, written into the contract.
Third, site disclosure. You need to know whether your capacity sits in a building the neocloud owns or in a third-party developer's shell, because the CoreWeave, Nebius and Nscale delays all ran through third parties. Ask which site, who builds and powers it, and whether the grid connection is signed or still applied for.
None of this argues against neoclouds. Price is still the reason most buyers look at them (our GPU cloud comparison covers where each one wins), and providers such as Nebius AI Cloud, Lambda and Crusoe Cloud have capacity running today. The risk sits in capacity that is promised for a future date in a building that is not finished.
What to Do Before You Sign or Renew
The work is contract review and site diligence, and most of it can start this week.
This Week:
- List every reserved or committed GPU contract you hold with a neocloud, the site behind each, and whether that site is live, under construction or not yet started. Your account team can answer; get it in writing.
- For each contract tied to a site under construction, pull the delivery date, the remedy for missing it, and how much you have prepaid. If the remedy is "best efforts," flag it.
- Ask your provider directly whether the facility depends on a third-party developer, temporary generation or a grid connection that is not yet energised.
This Month:
- Draft a standard rider for new capacity deals: milestone-based release of prepayments, weekly late-delivery credits, a no-fee exit after a stated grace period, and a defined "equivalent capacity" substitute.
- Size a fallback. Price the same workload on hyperscaler reserved capacity or provisioned throughput so you know what a three to six month slip would cost you.
- Have finance read the provider's latest filing or financing disclosures for the gap between contracted and active capacity. Nscale's was under 3% of contract value active.
Before Renewal:
- Split large commitments across at least two sites or two providers, with neither depending on the same unfinished building.
- Tie any multi-year term to capacity that is already accepted and running, and buy future-dated capacity on shorter options instead of firm commits.
The Bottom Line
Nvidia's disclosure matches how colocation risk has always worked: the party with the most leverage writes the conditions, and everyone further down the chain inherits what is left. We saw the same gap in colocation liability caps on GPU loss and in installed-versus-energised hyperscaler capacity. Nvidia has now said in a filing that its support for a neocloud starts at acceptance. Your contract should say what you get, and when, if the building is late.
Ask your provider which building your GPUs are going into, and when it gets power.
Continue Reading
- GPU Clouds Compared: Nebius on Price, CoreWeave at 3AM
- BIS Requires Licenses for AI Chips Shipped to Five Countries
- Microsoft's GPUs Sit Unplugged. Buy Delivery, Not Capex.
- When the Colo's Chillers Fail, the GPU Loss Is Yours
- OpenAI's 3-Year Compute Commit: Discount or Trap?
- Equinix Set a Date, Not a Price. Cap Your Commit Term.
