# GPU clouds & neoclouds

URL: https://justuju.org/concept/gpu-clouds/
Layer 3 · Cloud & Infrastructure

Companies that own huge GPU clusters and rent slices of them by the hour, from big general clouds to specialist GPU-only providers.

## D1 · Story

Most people building AI don't own the giant computers it runs on. They rent them, a bit like hiring a car instead of buying one.
Some rental companies sell every kind of computer service. Others, called neoclouds, mostly rent out powerful AI chips.
Like: Hiring a car instead of buying one.

## D2 · Picture

Hyperscalers are the big general clouds: GPUs sit beside storage, databases, identity and networking you may already use, with many regions and compliance certifications.
Neoclouds are specialist GPU clouds. They focus on accelerators, often offer newer chips sooner or simpler pricing, and run fewer other services.
You can rent on demand (pay by the hour, may be unavailable), reserve capacity for a fixed period (cheaper per hour, you pay even when idle), or own hardware. Renting suits bursty or uncertain workloads; owning pays off only if the GPUs stay busy for years.
Like: Hiring a car by the day, leasing one, or buying one outright. Where it breaks: Hire cars are interchangeable and plentiful. The newest GPUs are scarce, can sell out in a region, and large training jobs need thousands of them networked together in one place, which a rental counter can't promise.

## D5 · Frontier

D2's rent, reserve or own choice assumes a settled market. GPU capacity is neither priced nor valued in a settled way.
How fast GPUs lose value. One hyperscaler cut the assumed life of some servers from six to five years in 2025, citing the pace of AI hardware. Depreciation choices change the true cost of each GPU-hour (Amazon 10-Q, 2025).
Concentration at both ends. A large specialist GPU cloud disclosed that its biggest customer made up 62% of 2024 revenue and its top two about 77%, and that all its GPUs come from one vendor (CoreWeave S-1, 2025).
Falling prices. GPU compute per dollar has doubled roughly every 2.5 years, so long reservations at today's prices can look expensive within a generation (Hobbhahn and Besiroglu, 2022).

## Examples as of 2026-10-30

- Hyperscalers: AWS, Microsoft Azure, Google Cloud, Oracle Cloud
- Neoclouds: CoreWeave, Lambda, Nebius, Crusoe
- GPU marketplaces: Vast.ai, Runpod

## Sources

- The NIST Definition of Cloud Computing (SP 800-145): https://csrc.nist.gov/pubs/sp/800/145/final
- Amazon EC2 User Guide: Capacity Blocks for ML: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-capacity-blocks.html
- Amazon Form 10-Q for the quarter ended March 31, 2025 (Note 1, useful lives of servers): https://www.sec.gov/Archives/edgar/data/1018724/000101872425000036/amzn-20250331.htm
- CoreWeave Form S-1 registration statement: https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm
- Trends in GPU price-performance (Hobbhahn and Besiroglu, Epoch AI): https://epoch.ai/publications/trends-in-gpu-price-performance


---
Justuju · https://justuju.org/ · Content CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote with a link to the page.