GPU Spot Pricing Q1 2026, Where the Market Is Heading

Across the brokered desks and public marketplaces we monitor, GPU spot rates softened modestly through Q1 2026 for Hopper-class silicon, while Blackwell and Blackwell Ultra remained tight enough to clear at, or above, list. The story is not a single price curve. It is a divergence.

H100 SXM5 dropped from a January median of $3.45/GPU/hr to roughly $3.07/GPU/hr by the end of March. H200 followed a similar but shallower path. B200 held the $4.80–$5.10 band almost without movement, and B300 cleared at $5.10–$5.80 with no sustained dips below that floor.

Blackwell Ultra remains supply-limited at the silicon level. CoWoS-L packaging capacity at TSMC is the proximate bottleneck, and the order book for B300 systems through Q3 2026 is, by every signal we have, fully allocated to hyperscalers and a small number of sovereign and frontier-lab buyers.

What does reach the open market is overwhelmingly H100 and H200 capacity coming off of expiring 12–18 month reservations from the 2024–2025 cohort. That overhang is the dominant driver of Hopper softness. It is not new fabs coming online, it is older contracts rolling off.

Inference demand is the clear story of the quarter. The shift in workload mix from training to production serving (particularly serving 70B+ models with long-context features) has pulled forward demand for memory-rich SKUs (H200, B200, B300) and softened demand for memory-light Hopper.

Training demand remains structurally strong but is increasingly concentrated. The mid-market training tier (10–100 GPU clusters for fine-tuning, research, evaluation) is the segment most exposed to Hopper softening, and it is also the segment where ClusterBid sees the largest opportunity for buyers to lock in 6-month reservations at substantial discounts to spot.

Our base case for the next two quarters: H100 settles in the $2.80–$3.00 band by mid-Q2, then drifts sideways. H200 holds slightly above that on the strength of its memory advantage for inference. B200 stays range-bound. B300 prices remain stiff through Q3 and only soften meaningfully when the next wave of CoWoS-L capacity clears in Q4.

The asymmetric risk is to the downside on H100. If even one large hyperscaler or sovereign buyer accelerates a Blackwell migration and dumps a meaningful Hopper reservation back into the market, the $2.80 floor breaks quickly. We rate that scenario as plausible but not yet probable.

The widest dispersion in observed prices is not across SKUs. It is across providers. The same H100 SXM5 hour is trading at $2.65 from tier-3 colos in secondary markets and $3.85 from premium tier-1 facilities with sub-millisecond cross-region networking and a real SLA. That is a 45% spread on identical silicon.

Most of the spread is real and reflects what you actually get: power redundancy, network topology, support response times, and contract flexibility. Some of the spread is purely informational: buyers who do not know what the median is end up paying the asking price. The fix for that is straightforward and is one of the reasons ClusterBid publishes live offers.

If you are buying Hopper capacity for the next 6–12 months, this is the best window in 18 months to lock in a reservation. Spot is cheap, providers are motivated, and the negotiating leverage has shifted noticeably toward buyers.

If you are buying Blackwell, plan for tightness through at least Q3. Reservations matter more than spot in this segment, and the providers who actually have B300 inventory have meaningful pricing power. Get on a waitlist now if you do not already have one.

And if you do not know which tier you should be buying (what tier of facility, what term length, what SKU mix actually fits your workload), that is exactly the conversation we exist to have. Pricing is the easy part. Matching capacity to a real workload is the part where buyers most often leave money on the table.