Who are you paying now?

Three-year picture

The honest caveat: open models you can run yourself (Llama, Qwen, Kimi, DeepSeek, via Ollama or vLLM) are not frontier models for every task. The pattern that works is moving the private, repetitive, high-volume work onto your own box and keeping the API for what only the frontier models do well.

“But won't it be obsolete in a year?”

It is the first thing everyone asks, so here is the straight answer.

The GPU is a component, not the machine. Our systems are built to be re-carded. The chassis, motherboard, CPU, memory, storage, power supply and rails outlive several GPU generations, so when the cards age you replace the cards and keep everything around them. After two years of renting you own nothing. After two years of owning you have a platform and a decision about which cards go in it next.

Renting does not dodge obsolescence either. Cloud providers are still charging for A100s at rates set years ago. You are not getting free upgrades; you are renting aging silicon at a price set by scarcity.

Old cards keep working. The frontier moves, usefulness does not evaporate. The cheapest practical way into local inference right now is a used datacenter card from 2017. Meanwhile quantization and smaller architectures keep pushing the hardware needed for a given quality level down. The same box tends to get more capable as the software improves, not less.

And the part nobody puts in a spreadsheet: people under-use GPUs they rent, because the meter is running. Owning takes cost anxiety out of experimentation. In a field that changes every couple of months, being free to just try things is worth more than the hardware.

Email yourself these numbers

We will send your break-even and the three-year table to your inbox, with the assumptions spelled out so you can check our work or forward it to whoever signs off. Reply to it and we will size the thing against your real workload, including telling you if owning does not make sense for you.

One email with your numbers. We do not add you to a list, and we do not share your address. If you would rather just talk to a person, contact us here.

Where these numbers come from

Cloud rates, verified 2026-08-10, from each vendor's published pricing page, per GPU per hour, US regions, on-demand:

Hyperscaler is AWS on-demand in us-east-1 (p5.48xlarge, p4de.24xlarge, g6e.12xlarge, divided by GPU count). Specialist is RunPod Secure Cloud, cross-checked against Lambda and CoreWeave. Marketplace is RunPod Community Cloud. Vast.ai listings can go lower still but are individual hosts rather than a rate card, so we do not quote them. AWS one-year Savings Plans cut roughly 37-41% off their on-demand rate; if you are committing anyway, use the specialist column as a fairer stand-in.

What we leave out, in both directions: cloud storage and egress (AWS charges about $0.09/GB out; RunPod, Lambda and CoreWeave do not charge egress), your existing rack and network, and the resale value of hardware you own at year three. We also do not model the thing most buyers care about most and cannot price: your data never leaving the building.

Hardware prices are our live published starting prices. Configure any of them at eracks.com/products/ai-rackmount-servers/.