Why would a customer sign through 2029 for an Nvidia GPU introduced in 2020?
CoreWeave reportedly signed exactly that kind of contract for the A100. The customer, pricing, utilization, and operating costs were not disclosed, so the agreement does not prove profitability. It does reveal that technical obsolescence and economic obsolescence are not the same.
A Six-Year-Old GPU Can Still Earn Scarcity Rent
Older GPUs can remain useful for inference and less demanding workloads. That makes the 5 to 6 year useful lives used by CoreWeave and Nebius more defensible than the bear thesis assumes.
The useful life of AI hardware may depend less on when the next chip arrives and more on whether customers can find productive workloads for the existing one. A six-year-old GPU with paying demand is not economically obsolete.
Customers Are Paying Up and Paying Upfront
CoreWeave’s backlog rose 46% year over year to roughly $104 billion. The company reportedly raised prices by about 25% in July.
Nebius provided more detail on what scarcity pricing looks like. It signed four customer agreements averaging more than $1 billion in total contract value, while total contract value won during Q2 increased 4 times.
Mid-term agreements lasting 1 to 3 years reportedly price at $20 million to $25 million per megawatt, with upfront payments covering 50% to 60% of the associated capital expenditure. Short-duration agreements lasting up to 6 months reportedly command $40 million to $50 million per megawatt, with some contracts even higher.
Those figures cannot be blended casually. A short-term annualized rate is not necessarily representative of multiyear economics. Still, customers are doing more than expressing interest. Around 70% of Nebius’s Q2 deals included prepayments, and management expects more than $9 billion of prepayments in 2026 against over $40 billion of commitments.
Nebius’s first capacity auction reportedly cleared 15% above its previous highest Blackwell price. Management said it could sell all of its 2027 capacity on those terms today, but is retaining some capacity for immediate customer needs.
That decision makes sense while immediate availability earns a premium. It becomes risky if supply arrives faster than demand and that premium disappears.
Services Could Make Compute Less Commoditized
CoreWeave’s managed inference annual recurring revenue increased from roughly $1 million to more than $100 million in one quarter, with over $250 million expected by year end.
If inference adds software, orchestration, and integration around the hardware, CoreWeave may be developing beyond the role of a commodity GPU landlord. The current numbers have not settled that question, but additional services could make the customer relationship harder to replace than a simple capacity rental.
Power Determines Which Megawatts Have Value
The limiting factor is increasingly power, not simply access to GPUs.
Nebius raised its 2026 contracted power guidance to 5 gigawatts. At its completed Vineland facility in New Jersey, it switched the power source to Bloom Energy fuel cells. Management expects no significant impact on the project timeline.
On-site generation may help capacity bypass the power bottleneck. But the facility’s capacity, cost, fuel requirements, commercial terms, and approval timeline were not disclosed. Without those details, its economics cannot be estimated confidently.
Not all contracted megawatts are equal. Their value depends on when they become operational, how they are powered, what they cost to build, which contracts are attached, and how expansion is financed.
Financing Decides Whether Scarcity Reaches Shareholders
CoreWeave’s average debt cost fell about 300 basis points year over year, reportedly saving roughly $1.1 billion annually in interest. For a capital-intensive operator, financing cost is part of unit economics.
Customer prepayments can improve that equation. When customers cover 50% to 60% of associated capital expenditure upfront, they absorb part of the financing burden before revenue begins.
But demand does not automatically produce strong shareholder returns. Nebius reported $5.66 billion of capital expenditures and negative $0.12 in EPS. Its reported 22-month payback period offers some apparent cushion, but the calculation methodology and applicable contract cohort were not disclosed.
The broader counterargument is more severe. Combined hyperscaler free cash flow has fallen to less than one-quarter of its level two years ago. JPMorgan estimates that $4.1 trillion of $5.5 trillion in AI capital expenditure will be debt-financed. Meanwhile, token prices are falling, cheaper models are emerging, circular financing is a concern, and future supply could weaken current contract pricing.
Customers will pay a premium for immediately available AI compute. They will prepay substantial portions of capital expenditure, bid above previous Blackwell pricing, and sign contracts for GPUs introduced six years ago.
The unresolved question is how much of that scarcity rent remains for shareholders after the industry finances the power and capacity required to eliminate the scarcity.