.webp)
.svg)
.svg)
Previous Post
Renting versus buying datacenter GPUs under an annual hardware release cadence
NVIDIA now ships a more capable datacenter GPU tier roughly once a year, after running closer to a two-year cycle for most of the prior decade. For a team choosing whether to rent GPUs or buy them, the practical effect shows up in the length of time a card holds its value before a faster, and often cheaper, successor is available. The sections below cover the current release cadence, what it does to the rental and resale value of the prior tier, and a way to think through the rent-versus-buy decision when the next tier is rarely more than a year out.
Eva H.
June 10, 2026
The shift to a yearly release cadence
NVIDIA's datacenter roadmap moved from a roughly two-year release rhythm to a yearly one, a change the company set out on its updated roadmap in 2023 and has followed since.¹ Its recent flagship tiers have arrived about a year apart, announced at successive GTC conferences each March. The Hopper H100 was announced at GTC in March 2022 and reached general availability later that year.³ The Blackwell B200 was announced at GTC in March 2024, with GB200 NVL72 systems shipping to cloud providers from late 2024 and reaching fuller production through mid-2025.⁶ Blackwell Ultra, the B300, was announced at GTC in March 2025, with GB300 NVL72 racks reaching volume shipments around September 2025⁵ and the single B300 GPU arriving in early 2026.⁴ The next architecture, Vera Rubin, is expected to begin shipping in the second half of 2026, with broader production into 2027.²
There is also a mid-cycle refresh inside each generation. Hopper gained the H200, a memory upgrade, in 2024, and Blackwell gained the B300. Within about 12-18 months, a buyer can see two steps of improvement on the same architecture rather than one.
Resale and rental value of the prior tier
When you own a card, you hold its depreciation. As each newer tier ships, the open-market rental and resale value of the previous tier tends to soften, because buyers have a faster option and renters can choose it. The H100 shows the pattern. On-demand H100 rates averaged roughly $3.11 per GPU-hour in early 2026, with one cross-provider tracker putting the average nearer $3.45 and spot instances reported as low as about $1.25.⁸ The full spread across providers still ran about $0.70-14.90, depending on tier and provider.⁷ The H200, a year newer, now lists from about $0.49 per GPU-hour in some places,⁷ already under much of the H100 on-demand range, with 76% more memory.³ Some smaller clouds are now in a position where reselling H100s could mean taking them below purchase cost.
The softening is not uniform. Contracted and reserved capacity has held its price, and whole clusters have stayed hard to find. Through early 2026, industry reporting described H100 contracts being renewed at close to the rates signed 2-3 years earlier, several providers sold out of Hopper capacity, and B200 spot instances bid up toward $14 per GPU-hour during shortages.¹⁰ So the prior tier softens on the open, on-demand market while reserved and frontier capacity stays scarce and firm.
The rent-versus-buy decision under a yearly cadence
Buying tends to make sense when utilization is high and sustained, often cited as continuous use over roughly 18 months or more, where the hourly cost of renting, paid for every hour the hardware runs, adds up past the capital cost of owning.⁹ Renting tends to make sense for variable or exploratory work, and under a yearly cadence it carries a second benefit, in that the provider holds the depreciation and the redeployment risk on aging cards rather than the renter.
The capital side is not small. A single H100 has sold for roughly $25,000-40,000, and an eight-GPU H100 server for roughly $250,000-400,000 depending on configuration.⁹ Renting the same class of card has run about $1.25-3.45 per GPU-hour through the first half of 2026, depending on whether the workload can tolerate interruption.⁷ Whether owning pays back depends mostly on how many hours the card actually runs.
A short way to size the decision holds these variables together.
Sustained utilization. How many hours per week the GPUs will actually run, and for how many months. Low or bursty utilization favors renting.
Holding period against the cadence. If the holding period you need runs longer than the roughly 12-month gap to the next tier, owned hardware spends part of its life as the prior generation.
Capital against operating cost. Owning is capital up front plus power, cooling, and operations. Renting is operating cost that tracks usage.
Access to the newest tier. Renting gives a path onto a new tier as it ships. Owning commits to one generation until you resell or redeploy.
Resale and redeployment risk. Under a yearly cadence, the open-market value of an owned card tends to fall as the next tier arrives, and the owner carries that loss.
Switching cost. Moving between providers or generations has its own friction, which offsets some of renting's flexibility.
Selecting a provider under a yearly cadence
Under a yearly cadence, a renter has a standing question about any provider, whether it keeps moving its fleet onto current tiers and offers a path off aging hardware rather than holding customers on a single generation. A fleet that spans more than one vendor and more than one generation, with pricing shown per card and per hour, gives a renter room to match the workload to the cheapest card that fits rather than the only card on offer.
Under a roughly yearly cadence, the holding period that justifies buying keeps getting shorter. For teams below continuous, high utilization, renting works as much as a hedge against depreciation as a convenience, because whoever owns the card carries the cost of a tier aging out.
Sources
- NVIDIA datacenter release cadence and Blackwell architecture. Wikipedia, "Blackwell (microarchitecture)." https://en.wikipedia.org/wiki/Blackwell_(microarchitecture)
- Blackwell Ultra and Vera Rubin announcement and timing. CNBC, March 18, 2025. https://www.cnbc.com/2025/03/18/nvidia-announces-blackwell-ultra-and-vera-rubin-ai-chips-.html
- H100 announcement and general availability, H200 memory specifications. AI Wiki, "NVIDIA H100." https://aiwiki.ai/wiki/nvidia_h100
- B300 specifications and early-2026 availability. Spheron, "NVIDIA B300 (Blackwell Ultra) Specs, Pricing, and Benchmarks (2026)." https://www.spheron.network/blog/nvidia-b300-blackwell-ultra-guide
- GB300 NVL72 volume shipments, September 2025. Supermicro investor relations. https://ir.supermicro.com/news/news-details/2025/Supermicro-Begins-Volume-Shipments-of-NVIDIA-Blackwell-Ultra-Systems-and-Rack-Plug-and-Play-Data-Center-Scale-Solutions/default.aspx
- GB200 NVL72 shipment timing. IntuitionLabs, "NVIDIA Data Center GPU Specs: A Complete Comparison Guide." https://intuitionlabs.ai/articles/nvidia-data-center-gpu-specs
- H100 and H200 cloud rental pricing and spread, 2026. getdeploying, "H100 Cloud Pricing: Compare 44+ Providers." https://getdeploying.com/gpus/nvidia-h100
- H100 average and spot rates, early 2026. IntuitionLabs, "NVIDIA AI GPU Prices: H100 and H200 Cost Guide." https://intuitionlabs.ai/articles/h100-rental-prices-cloud-comparison
- H100 purchase and server costs, rent-versus-buy guidance. CloudZero, "H100 GPU Cost In 2026." https://www.cloudzero.com/blog/h100-gpu-cost/
- Capacity tightness, contract renewals, and B200 spot spikes. SemiAnalysis, "The Great GPU Shortage, Rental Capacity," April 2026. https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
Tell Us What You’re Building. We’ll Show You How It Runs.
