Lock GPUs with 3–5 year deals

GPU contracts are quietly turning into infrastructure mortgages. The old model was simple — buy servers, depreciate them, refresh when the next chip looked worth it. The new model is tighter and riskier. To secure enough AI capacity, customers are locking in supply years ahead, not just quarters ahead, and the bill keeps running even if demand, model design, or pricing shifts underneath them. ### What changed? The clearest signal came from memory, not GPUs themselves. Micron said it had completed agreements on price and volume for its entire calendar 2026 HBM supply, including HBM4. That matters because HBM is the bottleneck memory stacked next to AI GPUs — if HBM is spoken for, a lot of future accelerator capacity is effectively spoken for too. (investors.micron.com) ### Why does HBM tell you so much? HBM is the hard part of the AI server bill of materials. You can think of it as the premium fuel tank attached to the GPU engine — without enough of it, the chip cannot hit the performance buyers are paying for. So when Micron says 2026 HBM is already contracted and SK hynix is framing 2026 around an HBM-led supercycle, that is a sign customers are reserving AI buildouts far in advance. (investors.micron.com) ### Are these really multi-year commitments? Yes — just often one layer removed from the chipmaker press release. CoreWeave’s March 10, 2025 agreement to provide dedicated compute capacity to OpenAI was worth up to $11.9 billion, and the September 25, 2025 expansion added up to $6.5 billion, bringing total disclosed contract value to about $22.4 billion. Those are not spot rentals. They are long-duration capacity commitments built on top of pre-secured GPU and networking supply. (investors.coreweave.com) ### Why would customers sign that kind of deal? Because scarcity changed the math. If you wait for on-demand AI capacity, you may get nothing, or you may get it too late for a model launch. Prebooking buys certainty — racks, power, networking, memory, and deployment windows. NVIDIA is also leaning into this by productizing cloud and fabric layers like DGX Cloud and NVLink Fusion, which makes the stack more available but also more vertically tied to NVIDIA’s ecosystem. (nvidia.com) ### So where does the risk creep in? The catch is utilization. A 3–5 year capacity deal only looks smart if the GPUs stay full and commercially relevant. But AI demand is lumpy, model architectures change fast, inference gets more efficient, and the next chip generation can reset economics in a year. If a company overcommits, it is stuck paying for yesterday’s capacity at yesterday’s scarcity price. ### Why is vendor dependence getting worse? Because the constraint is no longer one component. It is the whole chain — GPU, HBM, packaging, networking, software, and cloud orchestration. Once a customer commits to one vendor stack, switching gets expensive fast. NVIDIA’s interconnect and software moat, plus limited HBM supply from Micron and SK hynix, means “capacity security” can turn into “ecosystem lock-in” almost by accident. (investors.micron.com) ### What should boards and CFOs actually worry about? Not just price. Flexibility. The dangerous contract is the one that assumes straight-line growth in GPU usage for four years. The safer version has ramp clauses, resale or reassignment rights, refresh options, and penalties that do not explode if workloads move. Basically, procurement is becoming capital allocation. ### Bottom line? The AI race is pushing buyers to secure compute the way airlines secure fuel or utilities secure gas — years ahead and at huge scale. That solves the shortage problem. But it creates a new one: if the market loosens before the contract does, the “guaranteed capacity” can become stranded cost.