Pricing, Supply & Architecture
The compute market has bifurcated and matured. Specialised neoclouds win on price and developer experience; hyperscalers win on compliance and ecosystem. On-demand H100 rates span a wide, context-dependent range, and next-generation supply remains tight against multi-quarter backlogs.
Reserve capacity for production, exploit spot for fault-tolerant batch, and design for 800 VDC and liquid cooling ahead of the Vera Rubin platform. Evaluate on goodput — not headline GPU-hour rates.
9–21×
H100 price spread
Context-dependent · varies by SKU, term & provider
-64%
Median H100 correction
From 2024–25 peaks
+40%
Workhorse demand surge
Some H100 rates since late 2025
40–70%
Neocloud savings
vs. hyperscalers
Tight
Next-gen GPU supply
Reported multi-quarter backlogs (SemiAnalysis)
36–52 wks
CoWoS / HBM lead times
Structural packaging constraint
The context-dependent price spread
On-demand price range by GPU
$ / GPU-hour · min to max, median marked in tooltip
Pricing by procurement tier
$ / H100-hour · low to high range
CoreWeave rack-scale rates
$ / hour on-demand
Reserved vs. spot vs. on-demand
On-demand
Best for experimentation & bursty inference. H100 $2.00–$4.00/hr (stabilized).
Reserved / committed
30–60% discounts; 1-yr reservations save 16–39%; CoreWeave up to 47% on 3-yr. Most enterprise volume flows via private bilateral agreements.
Spot
40–65% off on-demand (sometimes <$1.00/hr for H100) but subject to preemption — unsuitable for production inference.
Silicon supply & architecture
Turn these signals into a placement decision
Discuss power, location, cluster architecture and TCO with a CUDO AI infrastructure specialist.