Resources
GPU availability, explained.
GPU supply moves with every model launch. Here's where capacity actually comes from and how to plan around it.
Supply channels
Supply channels
- Retail and e-commerce: fastest for single consumer cards, worst for volume
- Authorized distributors and system integrators: allocation for workstation and datacenter GPUs
- Refurbished and retired farms: strong value on high-VRAM cards with health checks
- Colocation and bare-metal rental: no capex, fixed monthly cost
- Decentralized networks: Akash, Render, io.net, ICP and AO for burst capacity
What drives lead times
What drives lead times
- New architecture launches pull allocation to hyperscalers first
- VRAM capacity matters more than raw compute for agents and LLMs
- Power supplies, risers and cooling can be the real bottleneck
- Export rules affect some datacenter-class parts
How to plan
How to plan
- Define VRAM per job and concurrency first
- Price own vs rent at your expected utilization
- Keep a second source for every critical card
- Reserve decentralized capacity for peaks instead of overbuying
Want help applying this?
Tell us about your project and we'll spec it with you.