Why would an enterprise customer execute a binding multi-year contract extending through 2029 for an Nvidia GPU architecture originally released in 2020?
CoreWeave reportedly secured that long-term commitment for the legacy Nvidia A100 platform. While specific contract margins and operating expenses remain undisclosed, the transaction illustrates a vital reality: technical obsolescence and economic obsolescence are fundamentally different concepts. As immediate compute availability commands steep premiums, hardware useful lives are extending far beyond initial depreciation schedules.
Legacy Silicon Generates Enduring Scarcity Rent
Older accelerators remain highly cost-effective for model inference, fine-tuning, and less demanding enterprise workloads. This economic utility validates the five to six year depreciation schedules adopted by infrastructure operators like CoreWeave and Nebius, countering bearish assertions that GPUs become worthless within three years.
The operational life of computational hardware depends less on when next-generation silicon ships and more on whether customers can deploy the capacity profitably. A six-year-old GPU with paying demand is not economically obsolete. CoreWeave expanded its contractual backlog by 46% year over year to approximately $104 billion, while reportedly instituting a 25% price increase across available capacity in July. When demand for immediately usable compute outpaces supply, existing silicon continues generating substantial operating cash flow.
In addition, CoreWeave’s managed inference annual recurring revenue surged from roughly $1 million to more than $100 million in a single quarter, with management guiding to over $250 million by year-end. By wrapping orchestration, software, and networking around raw accelerators, operators can build customer relationships that are far harder to displace than commodity server rentals.
Scarcity Pricing and Customer Prepayments Shift Project Risk
Nebius provided concrete evidence of how compute scarcity is reshaping commercial contracts. The company executed four customer agreements averaging over $1 billion in total contract value, quadrupling its newly won commitments quarter over quarter. Short-duration contracts lasting up to six months reportedly cleared at $40 million to $50 million per megawatt, while mid-term agreements spanning one to three years priced at $20 million to $25 million per megawatt.
Importantly, customers are absorbing meaningful project risk. Approximately 70% of Nebius’s quarterly agreements included upfront cash prepayments covering 50% to 60% of associated capital expenditure. Management anticipates securing over $9 billion in prepayments during 2026 against $40 billion in overall customer commitments.
In its initial capacity auction, bids cleared 15% above previous peak Blackwell pricing. Management noted that it could sell its entire 2027 capacity today on those terms, but is deliberately retaining capacity to serve immediate enterprise needs at a premium. Retaining uncommitted capacity allows operators to capture spot premiums, though it exposes them to risk if capacity additions eventually catch up with demand.
Power Availability Governs Realized Economic Value
Electrical power and utility interconnections have supplanted chip availability as the primary bottleneck limiting deployment. Not all contracted megawatts possess equal economic value; value depends on when power is energized, what it costs to deliver, and which commercial agreements are attached.
Nebius expanded its 2026 contracted power target to 5 gigawatts. At its completed Vineland facility in New Jersey, the company switched its power source to Bloom Energy fuel cells, allowing the campus to bypass utility grid delays without disrupting project timelines. On-site generation helps operators avoid years-long utility queues, but facility economics depend heavily on fuel supply contracts, maintenance expenses, and regulatory approvals that remain undisclosed.
Financing Costs Decide What Reaches Equity Holders
For capital-intensive infrastructure providers, financing costs dictate whether top-line expansion reaches shareholders. CoreWeave’s average debt cost declined approximately 300 basis points year over year, generating roughly $1.1 billion in annual interest savings. Customer prepayments covering more than half of hardware capital expenditures further buffer balance sheets before revenue commences.
However, demand does not automatically translate into attractive equity returns. Nebius reported $5.66 billion in capital expenditures alongside an operating loss of $0.12 per share, and while its stated 22-month payback period offers comfort, calculation methods were not detailed.
More broadly, JPMorgan estimates that $4.1 trillion of the projected $5.5 trillion in global AI capital expenditures will require debt financing. Combined hyperscaler free cash flow has dropped to less than one-quarter of its level from two years ago. The central question is how much scarcity rent will remain for equity holders after the industry finishes financing the capacity required to satisfy that demand.