The Economics of Idle GPUs
NVIDIA signed a $6.3 billion take-or-pay contract with CoreWeave through 2032. That contract is the thesis of this article in one sentence.
The contract, and what it is
In September 2025, NVIDIA agreed to purchase unsold compute capacity from CoreWeave through April 2032, worth $6.3 billion. This section describes what the contract says and why that structure — take-or-pay, vendor financing, guarantee of collateral value — is different from insurance.
Section 2How GPU clouds are financed
GPU clouds do not buy hardware with cash. They buy it with debt facilities from private credit lenders like Blackstone, Magnetar Capital, Carlyle, and Coatue. This section describes how those facilities are structured, how depreciation cycles affect collateral value, and why lenders require a residual value guarantee from the chipmaker.
Section 3What happened to Lucent in 1999
Late 1990s: Lucent, Nortel, and Cisco financed their own customers so those customers could buy Lucent gear. The customers could not sell the bandwidth. They defaulted. Equipment flooded the secondary market. Lucent nearly went bankrupt. This section describes that history and what the structural similarity to NVIDIA's current position is.
Section 4NVIDIA's acquisitions of scheduling companies
NVIDIA acquired Run:ai for $700 million in December 2024, SchedMD (maintainers of Slurm) in December 2025, and Augtera Networks in March 2025. All three are GPU scheduling or orchestration companies. This section describes what each one does and what it means that the same chipmaker bought all three.
Section 5Signals that the market is softening
Michigan's Public Service Commission ruled that data centers must pay up-front energy costs. Microsoft slowed its data center leases in Q3 FY25. Wingspire Equipment Finance is doing rounds that treat GPUs as depreciating assets. Inference software (vLLM, quantization) is reducing compute demand per token. Hyperscalers are moving workloads to their own ASICs (Google TPU, AWS Trainium, Microsoft Maia). This section lists each of these signals and what they mean.
Section 6The counter-case
The argument against the piece's thesis is that NVIDIA is vertically integrating, not propping up weak customers. Evidence: DGX Cloud, NVIDIA's internal R&D compute needs for Nemotron and Earth-2, and the Jevons Paradox argument that cheaper compute will be consumed by new demand. This section presents that counter-case in its strongest form.
Section 7Where this leaves the market
NVIDIA must sell chips at a high rate to maintain its valuation. Its customers need credit to buy those chips. The credit requires a residual value guarantee. The guarantee puts NVIDIA on the hook for idle compute. This section describes where that dynamic leads.
This is a working outline, not the final piece. Research is done, frame is locked, draft is pending source interviews with Dylan Patel (Semianalysis), a private credit underwriter, Silicon Data, and Tom Blackwell (Nebius). Target publication: mid-May 2026.