Weekly memos on what moved in GPU compute pricing, why, and what it means. Previews are public; full memos and the archive are ComputePulse Pro.
May 31, 2026
Marketplace H100 rental prices drifted lower this week as cheap independent capacity re-entered listings, while managed-cloud and hyperscaler rate cards held flat. The divergence widened the spread between the cheapest spot capacity and enterprise-grade compute — a signal worth watching for anyone modeling training cost curves.
The Marketplace sub-index for H100 declined ~3–5% week-over-week, driven primarily by additional independent host capacity appearing in listings. Managed-cloud (neocloud) pricing was essentially unchanged, and hyperscaler rate cards did not move at all. Net effect: the headline blended index ticked down, but the move was concentrated in one tier.
The decline was composition-driven rather than a broad repricing. A handful of marketplace hosts re-listed cheap H100 capacity, pulling the marketplace median down. Managed clouds, which price on longer commitment cycles and bundled support, did not respond within the week. This is consistent with marketplace capacity acting as the fast-moving, demand-sensitive layer of the market.
Most of the marketplace move traces to independent hosts on peer-to-peer platforms. After de-duplicating by physical machine (one box listed at multiple GPU counts counts once), the effective number of distinct cheap suppliers rose modestly — enough to move the median. Managed-cloud providers (CoreWeave, Lambda, RunPod, Nebius) held their published rates.
The widening spread between marketplace and managed-cloud H100 pricing matters for anyone underwriting compute cost. Marketplace pricing is the leading indicator for short-term supply/demand, but it is not the price most enterprises actually pay. Modeling a training run off marketplace lows will understate real cost; modeling off hyperscaler rate cards will overstate it for cost-optimized buyers.
ComputePulse tracks observable public and quoted GPU rental pricing. It does not claim to observe every private contracted transaction. This week's marketplace move was verified against per-provider observation counts and de-duplicated by host to avoid a single seller skewing the index. Coverage in the hyperscaler tier for newer GPUs (B200) remains thin.
ComputePulse tracks observable public and quoted GPU rental pricing. It does not claim to observe every private contracted transaction.