CoreWeave’s $6.3B Backstop Meets 15% Bond Yields

In September 2025, Nvidia agreed to buy CoreWeave cloud capacity that other customers leave unsold. By October 2026, CoreWeave’s bonds yielded more than 15%. A fleet could have a buyer of last resort and still be expensive to finance.

The original cloud bargain made availability flexible. In 2018, Google offered preemptible GPUs at a 50% discount, but limited jobs to 24 hours and could shut them down with no more than 30 seconds’ warning. A customer bought a cheaper hour by accepting that the hour might end. By 2023, Google was packaging H100 GPUs into its A3 supercomputer VM for high-performance training, while reporting described traditional cloud infrastructure as poorly suited to large-scale AI. A training cluster needed more than an available chip: its machines and network had to work together for the duration of the job.

AI infrastructure is being remade from on-demand hardware rental into controlled-access capacity. Operators that can finance, energize and coordinate clusters can sell a reliable place in the schedule; operators that merely acquire GPUs remain exposed to the price of a rental hour. Nvidia’s CoreWeave agreement makes the distinction visible because it gives capacity left unsold to other customers a contractual buyer, without settling what that capacity costs to build or operate.

Contracts made the fleet financeable without proving reliability

Blackstone led a $2.3 billion private-credit loan to CoreWeave in 2023, secured against Nvidia chips. Reporting on the loan describes it as a blueprint for a much larger chip-backed debt market. CoreWeave later moved $2.6 billion of AI-data-center construction debt off its balance sheet through special-purpose vehicles. In each case, a lender had to look beyond the resale value of hardware toward the customer payments that hardware might produce. That is the work of AI infrastructure finance: turning a utilization expectation into an asset someone will fund.

CoreWeave told investors in 2024 that it had $17 billion in signed contracts, including approximately $10 billion from Microsoft covering 2023 through 2030. On September 15, 2025, Nvidia committed $6.3 billion to buy CoreWeave cloud capacity not sold to other customers through April 13, 2032. The agreement put the chip supplier on the demand side of the business it helped equip. Nvidia later said it would rent unused GPUs from younger cloud providers in exchange for a share of their revenues—a different arrangement, but one that also assigned residual-demand risk to the vendor.

Neither a signed backlog nor a vendor backstop is collected revenue. The available account of Nvidia’s order does not establish its drawdown schedule or the price CoreWeave receives for each unused unit. It does establish a consequential change in the business: CoreWeave can seek financing with a named counterparty for capacity that might otherwise sit unplaced. The related market for long-duration compute commitments begins with that separation between a rack’s possible output and a contract for its use.

Customers began buying admission to the queue

Amazon encountered the allocation problem inside its own retail division. After GPU shortages delayed projects, the company created Project Greenland in July 2024 to manage who received GPUs and how those resources were used. Amazon later said it had ample capacity; the earlier shortage had already forced it to make approval and scheduling explicit. Even a company that owns the fleet must decide which workload gets in first.

OpenAI carried the same decision to customers in May 2026 with Guaranteed Capacity, offering assured access to its compute through one- to three-year commitments. OpenAI sells a different service from a bare GPU cloud, but both must reserve scarce resources against future requests. A reservation protects one customer by limiting what the operator can promise another.

CoreWeave reported $2.58 billion in second-quarter 2026 revenue, up 112% year over year, alongside a $104 billion revenue backlog and 1.5 gigawatts of contracted power. CoreWeave’s buyers have committed to long-term access. The figures do not say how much of that power was delivering billable compute at any particular moment. In the CoreWeave coverage record, enterprise framing rose from 20.0% to 45.7% between the compared periods while funding framing fell from 60.0% to 30.4%. That measures a change in what reporting emphasized, not a conversion rate from contracts to operating clusters.

The service boundary extends beyond the GPU

Luiz Barroso and Urs Hölzle described the data center as a computer because servers, storage, networking, power and cooling must function as one system. For an AI customer, the useful unit is not a chip in inventory but a cluster that can accept a workload, keep it isolated, survive component failures and return work on schedule. Reliability has to be delivered through the service interface rather than left for each renter to assemble.

CoreWeave, Akamai and Crusoe agreed in October 2026 to offer Perplexity’s open-source SPACE agent sandbox to customers. The partnership joins capacity providers to an application-facing isolation layer; it does not, by itself, prove that any one provider has solved cluster reliability. It does show where providers are trying to meet customers: at the point where a workload enters the system, not at a catalogue of installed accelerators.

That distinction also leaves room for competitors. A cloud that runs a customer’s job predictably can earn more than one that lists the same GPUs. But orchestration can sit above multiple fleets, and a customer can compare providers if their isolation, scheduling and recovery arrangements fail to differ in practice. Owning the warehouse does not automatically mean owning its control room.

A planned megawatt cannot admit a workload

Crusoe raised $3.9 billion at a roughly $30.9 billion post-money valuation while pursuing factory-built data centers. Factory construction is an answer to a concrete delivery problem: customers cannot use a reserved cluster until the site is built, connected and equipped. Sources also reported a potential Meta commitment for roughly 1.6 gigawatts across Crusoe sites in Texas and Missouri, although the supplied relationship record marks that arrangement as rumored rather than confirmed.

Crusoe’s Wyoming project supplies the other side of the construction case. The company paused the proposed data center after reportedly failing to secure customers including Google amid concerns about cost and timetable. Financing and a construction plan could not substitute for a customer willing to accept the site on those terms.

In 2025, U.S. data-center capacity that was built, underway, planned or stalled exceeded 80 gigawatts. Those categories cannot be added up as operating supply. A site’s capacity lag runs through permits, construction, power delivery and cluster commissioning before a customer can start work. The place-bound AI capacity market prices those handoffs as surely as it prices the GPUs that eventually occupy the racks.

The buyer of last resort does not insure the entire business

CoreWeave’s growth is substantial, and its 2026 revenue and backlog argue against treating contracted demand as fiction. Its financing costs argue against treating the same contracts as utility-grade cash flow. CoreWeave bond yields exceeded 15% in October 2026, and its stock had fallen more than 60% from its June 2025 peak. Nvidia’s order addresses unsold capacity; it does not establish that every site will be commissioned on time, that operating costs will stay fixed, or that debt will remain cheap.

CoreWeave’s S-1 showed that its two largest customers supplied about 77% of 2024 revenue, with Microsoft alone accounting for 62%. A large customer can fill a fleet and still retain considerable bargaining power when the next contract comes due. At the adjacent data-center-leasing layer, Applied Digital attributed part of a quarterly revenue shortfall to customers delaying renewals: it reported $53 million against a $63 million estimate. A lease has a term; a building has a longer life.

Customers can also change how much hardware they need. OpenAI engineers reportedly found a way to more than halve inference cost, while Cerebras, Groq and large technology companies pursued alternatives for inference workloads. Barclays projected $208.2 billion in inference capital expenditure for 2026, but a large market does not guarantee that any fixed GPU fleet captures it at its original economics. Cheaper computation helps customers first; it helps a lessor only if new usage fills the hours that efficiency frees.

Nvidia can sign for an empty slot in CoreWeave’s schedule. The operator must still build the room, close the breaker and keep the machines working after the customer enters.