.webp)
.svg)
Previous Post
The determinants of GPU cloud rental pricing
GPU rental prices have climbed back above $2 per GPU-hour and on-demand capacity is hard to find. How much that affects you depends on the provider you picked, and on choices it made long before you signed on about hardware, location, and how it secured the chips, power, and facilities behind your rate.
June 9, 2026
GPU rental prices and available capacity
Start with what shows up on your invoice. After fallingfrom $7-10 per GPU-hour at launch down to $1-2 through 2024 and 2025,2H100 rates have turned back up and now sit above $2.3 One-yearrental contracts rose about 40% in five months, from $1.70 in October 2025 to$2.35 by March 2026, and on-demand capacity is sold out across most GPU types.1When you try to spin up a large block of GPUs on demand and can’t, it’s usuallybecause the operators holding that capacity aren’t releasing it back into thepool, even as prices rise.1
Capacity has been the binding constraint across theindustry. Nebius, one of the larger GPU clouds, reported selling out across GPUgenerations while customers prepaid to lock in future supply.10 Someof the squeeze is supply, since chip production has been shifting towardNvidia’s newer Blackwell line, which leaves fewer H100s to go around even astotal compute grows, and the advanced packaging that assembles these chips hasbeen a bottleneck the whole way up.4 The practical effect is thatcheap, abundant GPUs aren’t something to count on, and whether a provider canhold your price steady depends on how well it secured supply ahead of time.
Power and cooling demands of current GPU hardware
The hardware you’ll want to run next raises the bar on theprovider. Nvidia’s current top-end system, the GB300 NVL72, packs 72 BlackwellGPUs into a single rack that draws about 132-140 kW, with short peaks up to 155kW.5 A normal server rack runs around 8 kW, so one of these pullsroughly 17 times the power, and only about 1% of data centers can handle racksabove 100 kW at all.6 It also can’t be cooled by air. It needsliquid piped directly to the chips.6
A provider can list the latest hardware without being ableto run it well. Drop these systems into a building that wasn’t designed for thepower and cooling and they don’t run at full performance, if they run at all.The providers that get full performance out of current-generation hardware arethe ones whose facilities were built or upgraded for it, and that’s notsomething a pricing page tells you.
Power cost and grid access by region
The scarce resource in AI infrastructure has shifted from the GPUs themselves to the power needed to run them, which is also the single largest ongoing cost of operating a data center. Data center electricity demand is set to roughly double from about 485 TWh in 2025 to 950 TWh by 2030, with AI-specific demand tripling over that period, while the largest tech companies spent more than $400 billion on infrastructure in 2025 and are expected to raise that about 75% this year.¹¹ A single cluster of a few thousand GPUs needs several megawatts of continuous power, and the largest campuses run far higher.⁹ Providers build where they can get cheap, reliable power at scale, and that choice flows straight into the rate you pay.
The US is the largest market, accounting for about 45% of global data center electricity in 2024, with national demand projected to grow roughly 130% by 2030.¹² The constraint there is rarely generation. The hard part is getting power delivered, because connecting a large new load to the grid can take years, and that has pushed some developers to build on-site gas or batteries to bridge the wait.¹¹, ⁸ Much of the capacity concentrates in a few hubs like Northern Virginia, straining the local grid.
The Gulf is building fastest, on cheap energy and sovereign capital. Power runs around $0.11 per kWh in the UAE and roughly half that in Saudi Arabia, well below typical US industrial rates,¹³ and state investment funds can commit to gigawatt-scale campuses without the financing cycles that slow projects elsewhere. Across the region, securing power has become a harder problem than raising capital.
Asia spans the widest range. Japan and Singapore sit at the expensive, grid-constrained end, while several Southeast Asian markets offer cheaper power, where data center electricity runs from roughly $60 per MWh in the cheapest markets to $178 in the most expensive.⁷ Cheaper power tends to come with less grid reliability, so a lower rate can carry more risk of outages. Europe is tightening too. The data center projects now proposed in the UK alone would need about 50 GW, more than the country's entire peak demand today.¹⁴
Wherever a provider sits, that geography decides how much power it can get and what it pays for it, long before any of it reaches your bill.
Securing chips and power ahead of demand
A provider's price and availability are only as stable as the supply beneath them. A cloud that secured power and chips early and runs facilities built for current hardware can hold steadier rates through a tight market than one that has to buy capacity reactively as demand spikes.
Sources
- SemiAnalysis, "The Great GPU Shortage – Rental Capacity," April 2026. https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
- Silicon Data, "H100 Rental Price Over Time (2023–2025)." https://www.silicondata.com/blog/h100-rental-price-over-time
- Silicon Data, "H100 Price Spike: Understanding the 10% Surge in GPU Rental Costs," January 2026. https://www.silicondata.com/blog/h100-price-spike
- Kavout, "Why are NVIDIA H100 GPU rental prices surging by 40%." https://www.kavout.com/market-lens/why-are-nvidia-h100-gpu-rental-prices-surging-by-40
- Lenovo Press, "Lenovo NVIDIA GB300 NVL72 Rack Scale AI." https://lenovopress.lenovo.com/lp2357.pdf
- Barrack AI, "B300 Draws 1,400W Per GPU. Most Data Centers Aren't Ready." https://blog.barrack.ai/nvidia-b300-1400w-data-center-requirements/
- Wood Mackenzie, "Southeast Asian data-centre power demand is set to explode," December 2025. https://www.woodmac.com/news/opinion/southeast-asian-data-centre-power-demand-is-set-to-explode/
- Ropes & Gray, "Data Center Investment in 2026." https://www.ropesgray.com/en/insights/viewpoints/102mvfl/data-center-investment-in-2026-ai-demand-power-constraints-and-private-equity
- SemiAnalysis, "100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability." https://newsletter.semianalysis.com/p/100000-h100-clusters-power-network
- Futurum, "Nebius Q1 FY 2026 Earnings Show AI Cloud Capacity Scaling," 2026. https://futurumgroup.com/insights/nebius-q1-fy-2026-earnings-show-ai-cloud-capacity-scaling/
- IEA, "Key Questions on Energy and AI," April 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
- Brookings, "Global energy demands within the AI regulatory landscape," April 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
- DatacenterKnowledge, "UAE Data Centers: Powering the Middle East's AI Revolution," February 2026. https://datacenterknowledge.com/build-design/uae-data-centers-powering-the-middle-east-s-ai-and-cloud-revolution
- Engineering & Technology (IET), "IEA warns AI data centre electricity use will triple by 2030," April 2026. https://eandt.theiet.org/2026/04/22/iea-warns-ai-data-centre-electricity-use-will-triple-2030
