Enterprises are locking in H200 capacity now because on-demand access has effectively sold out across most providers, contract terms are getting shorter and stricter with each renewal cycle, and every recent H200 GPU price update has moved in one direction. Waiting does more than just increase the risk of a higher rate. It risks not having capacity when you actually need it.
Here’s what’s actually driving that urgency, and what to check before signing a multi-year commitment yourself.
Why is on-demand H200 capacity disappearing?
On-demand H200 capacity is disappearing because providers are renewing existing reserved contracts rather than releasing it back to the open market. As of early 2026, on-demand access is effectively sold out across hyperscalers and specialized clouds alike, and teams without an existing reservation are finding it genuinely difficult to secure compute on short notice.
This is not a temporary dip. Hyperscalers alone have committed over $450 billion to AI infrastructure in 2026, and that spending is largely going toward locking in supply for their own customers through reserved capacity, rather than expanding the pool of on-demand availability that everyone else draws from.
How much have H200 rates moved this year?
H200 rates have moved sharply enough to justify the urgency on their own. AWS raised its H200 Capacity Block pricing by roughly 15% in a single adjustment in January 2026. Comparable H100 one-year rental rates climbed about 40%, from roughly $1.70 an hour to $2.35, between October 2025 and March 2026. Neither of these was an isolated spike. They reflect the same underlying supply pressure that’s been driving GPU pricing all year.
If your organization is still budgeting off rates from even two or three quarters ago, it’s worth checking current pricing before finalizing a capacity plan built on outdated numbers.
Why are contract terms getting worse, not just prices?
The Contract terms for on-demand H200 GPU are tightening because scarcity has shifted negotiating power entirely toward providers. The short end of the market, monthly or short-term reservations, has largely disappeared for premium GPUs like the H100 and H200. Providers are now broadly unwilling to sign anything under six months, and many are pushing enterprises toward three-year terms specifically because they can.
That trend compounds the pricing problem. Every quarter an enterprise delays locking in capacity, the available contract structures get less flexible, not just more expensive, which is exactly the kind of dynamic that rewards moving early rather than waiting for conditions to improve.
What does locking in capacity save?
Locking in capacity saves a meaningful amount, though the exact number depends heavily on provider and term length. Hyperscaler reserved instances and committed-use discounts run 30% to 72% below on-demand rates for one- to three-year commitments. Specialized providers like CoreWeave offer smaller but still real discounts, typically 15% to 30%, aimed at teams running large, predictable training clusters.
Reserved capacity also does something on-demand pricing can’t: it guarantees access during a shortage. A contractual reservation takes priority over on-demand traffic when supply tightens further, which matters more than the discount itself for enterprises that can’t afford a training run to stall because capacity simply wasn’t available that week.
Is locking in reserved capacity always the right move?
Locking in reserved capacity isn’t automatically the right move, and treating it as one is a real risk. The economics only work above a real utilization threshold, generally 60% to 70% or higher. A team that reserves capacity and runs it at half that rate is effectively paying close to double for every productive hour, which erases most or all of the discount reserved pricing was supposed to provide.
Bursty or still-evolving workloads are a poor match for a hard multi-year commitment. Some providers now support secondary trading of reserved contracts, letting a team resell or reallocate capacity it no longer needs, which reduces that risk somewhat, but it doesn’t eliminate the need to model utilization honestly before signing.
What should enterprises do before committing?
Enterprises should model actual utilization first, not projected best-case usage, before committing to a term. Beyond that math, two practical details matter: lead times on tighter-supply parts like the H200 now run four to eight weeks even after a contract is signed, so “locking in now” doesn’t mean instant access, and shorter, flexible structures are worth exploring for workloads that don’t yet have a stable, predictable baseline.
Before signing a multi-year term, it’s worth comparing current terms and rates directly rather than assuming the first offer reflects the best available structure.
Conclusion
Enterprises are locking in H200 capacity now because the alternative, waiting, doesn’t lead to better pricing or easier access later. Rates have moved sharply in a single direction all year, contract terms keep getting less flexible, and on-demand availability keeps shrinking as more of it gets absorbed into long-term reservations. The right response isn’t blind urgency, though. It’s modeling utilization honestly, understanding lead times, and locking in only what the workload can actually justify.
Frequently asked questions
1. Why is H200 on-demand capacity so hard to find in 2026?
Providers are renewing existing reserved contracts instead of releasing capacity back to the on-demand market, and demand from AI infrastructure spending has outpaced available supply.
2. How much can reserved H200 contracts actually save?
Hyperscaler reserved instances and committed-use discounts typically run 30% to 72% below on-demand rates for one- to three-year terms, while specialized providers offer smaller discounts in the 15% to 30% range.
3. What utilization do you need to justify a reserved contract?
Generally 60% to 70% or higher. Below that threshold, the effective cost per productive hour can exceed what on-demand pricing would have cost for the same workload.
4. How long does it take to actually get reserved H200 capacity after signing?
Tighter-supply parts like the H200 typically take four to eight weeks to provision after contract signing, sometimes longer for rack-scale systems, so reservations need to be planned well ahead of when capacity is actually needed.
