H100 rentals are back to $2.35 an hour, and your AI cost model is stale

The average on-demand price of an Nvidia H100 went from roughly $1.70 per GPU-hour in October 2025 to roughly $2.35 in March 2026, according to SemiAnalysis data reported by Crypto Briefing ^[1][2]. If you built your inference gross margin on 2025 compute prices, the input has moved about 38% against you and the wait for a new cluster is now 12 to 18 months ^[3][17].

That is a reversal, not a continuation. Through much of 2025, rental prices fell more than 60% from their 2023-2024 peaks as CoreWeave, Lambda and other specialists poured on capacity and the hyperscalers widened their own offerings ^[4][5]. Take those two figures at face value and current pricing still sits around 45% below the boom-era peak ^[18]. This is not a return to 2023 scarcity economics. It is the end of the deflation that a lot of 2025 business plans quietly assumed would continue.

The supply picture explains the turn. Crypto Briefing reports that on-demand H100 capacity is effectively sold out across both neoclouds and hyperscalers, with inference workloads and multi-agent systems pushing requirements past the installed base ^[6][7]. Clusters as small as 8 nodes, or 64 GPUs, have become hard to procure on short notice ^[8]. February 2026 alone saw 15% to 20% increases as providers repriced ^[9]. On-demand rates now span $2.19 to over $4 per GPU-hour depending on provider and whether capacity was reserved ahead of time, an 83% spread for the same silicon ^[10][19].

Blackwell is not the escape hatch. Lead times for B200 and GB200 deployments have stretched into mid-2026, and most of the available 2026 capacity is already committed ^[11].

The contract layer is where the real advantage now sits. Some operators are renewing existing H100 agreements at previously agreed legacy rates, and a few have extended commitments through 2028, a four-year term that was nearly unheard of in cloud GPU markets 18 months ago ^[12][13]. That produces a two-tier cost base: competitors in the same product category paying materially different prices for identical compute. On a modest 64-GPU cluster, the move from $1.70 to $2.35 is about $41.60 an hour, or roughly $364,000 a year at continuous utilisation ^[20]. That is the difference between two seed-stage runways.

For providers, the same squeeze reads as higher utilisation and better revenue per rack, and anyone holding a favourable long-term Nvidia supply agreement is running a spread trade: cheap wholesale, expensive retail ^[14][15]. AWS, Google Cloud and Azure can absorb temporary margin compression on balance sheet; smaller neoclouds with uncommitted capacity have more pricing power than their brand would suggest ^[16].

Two cautions. This is a single-source read on a market with no public order book, and the publisher's own headline claims a 50% surge in six months while its text cites roughly 40% between October and March ^[21]. Treat the direction as more reliable than the decimal.

What to watch: whether the early mid-2026 stabilisation reported at these higher levels actually holds once Blackwell deliveries land ^[22]; whether the reserved-to-spot spread widens past $4, which would signal that spot is now a distress channel rather than a discount one ^[10]; and whether four-year lock-ins spread from the few to the many, because that is the point at which cheap compute stops being available to late entrants at any price.