Rent GPU for AI: what it costs, who earns, and when it beats an API

Key facts

What “Rent GPU” can mean	Hourly whole-GPU cloud, token-priced inference, or a contributed GPU network
Plugsky shared GPU price	$0.30 input / $2.40 output per 1M tokens
Operator share	60% of routed customer revenue; 40% gross platform share
Minimum GPU	7 GB VRAM / unified memory, 7 GB RAM; 16 GB recommended
Shared model	Plugsky GPU 27B · 32K context · tools · JSON
Data boundary	Shared network: non-sensitive workloads only. Isolated sovereign option for sensitive data.
Payouts	Monthly Stripe Connect, $50 minimum
Product status	Live

TL;DR

• Hourly GPU rental wins for sustained, predictable workloads and training, where you can keep the hardware busy.
• Token-priced inference wins for bursty chat, RAG, and agents because you do not pay for idle GPU hours.
• Plugsky’s shared GPU network pays operators 60% of routed revenue and gives buyers an OpenAI-compatible API.
• The shared network is not for sensitive data. Use a dedicated, VPC, on-prem, or air-gapped environment for that.
• You can bring your own GPU into an isolated sovereign deployment; your hardware stays in your perimeter.

Three things people mean by “rent GPU”

The phrase is overloaded, and that causes bad comparisons. A whole-GPU cloud instance is billed by the hour or second. A token-priced API is billed by usage. A contributed GPU network is a marketplace where owners supply hardware and a platform routes work to it. They solve different problems.

Model	You pay for	Best fit	Main risk
Hourly GPU cloud	GPU time, storage, egress, ops	Training, fine-tuning, steady high utilization	Idle hours; you run serving and monitoring
Token-priced inference	Input and output tokens	Chat, RAG, agents, bursty traffic	Per-token cost at very high sustained volume
Contributed GPU network	Tokens routed to verified nodes	Cost-sensitive inference; owners monetize idle GPUs	Non-sensitive workloads only; supply varies

What whole-GPU time costs in 2026

Published hourly rates are a useful anchor. The table below uses public list prices from Runpod and Lambda, checked on 2026-10-27. Rates change with supply, commitment, region, and taxes.

GPU	VRAM	Published hourly range	Source
RTX 4090	24 GB	$0.34 – $0.89 / hr	Runpod
RTX 5090	32 GB	$0.69 – $1.19 / hr	Runpod
A100 80GB	80 GB	$1.19 – $1.79 / hr · $1.99 – $2.79 / hr	Runpod · Lambda
H100 80GB	80 GB	$1.99 – $3.99 / hr · $3.99 / hr	Runpod · Lambda
H200	141 GB	$3.59 – $5.29 / hr	Runpod
B200	180 GB	$5.98 – $7.99 / hr · $6.69 / hr	Runpod · Lambda

These are whole-GPU prices. They exclude the engineering time to serve a model, monitor it, handle failover, and keep utilization high. A GPU that sits idle still bills. That is the hidden cost that makes “$0.34/hr” look cheaper than it is for bursty workloads.

Where Plugsky fits: token-priced shared GPUs

Plugsky’s shared GPU network is built for inference, not for renting a bare-metal GPU by the hour. Buyers pay $0.30 per 1M input tokens and $2.40 per 1M output tokens. GPU owners keep 60% of the customer revenue; Plugsky keeps 40% as a gross platform share before payment processing. Stripe fees come from the platform side, not the operator.

The current shared model is Plugsky GPU 27B: a 27B, IQ2-mix, Apache-2.0 model with a 32K context window, tools, JSON mode, and streaming after completion. A worker needs at least 7 GB of VRAM or unified memory and 7 GB of system RAM; 16 GB is recommended for headroom and 32K context. The minimum throughput is 4 tokens per second, with 12 tokens per second recommended.

The trade-off is straightforward: token pricing removes the idle-hour risk for the buyer, and the operator carries the utilization risk. That is why the shared network pays 60% rather than a fixed hourly rate.

When hourly GPU rental still wins

Hourly infrastructure is the right tool when you can keep it busy. Training, fine-tuning, batch jobs, video rendering, and steady production inference can justify a dedicated GPU. You also get full control of the runtime, custom CUDA builds, and the ability to run any model you want.

Token-priced inference is the better default for interactive chat, RAG, agents, and spiky traffic. You avoid paying for idle capacity, and you do not need a serving stack, watchdog, or GPU operations team. The honest answer is to model your own utilization: if the GPU is busy most of the month, hourly can win; if it is bursty, token pricing usually wins.

When you need an isolated sovereign GPU

The shared GPU network is explicitly for non-sensitive workloads. Every payload is classified before routing, but shared consumer hardware is not the right place for regulated, confidential, or customer data. For those workloads, Plugsky offers dedicated and isolated sovereign environments:

• Dedicated GPU node — reserved capacity, not shared with other tenants.
• Private VPC — run inside your cloud account or network boundary.
• On-prem — keep inference in your own data center.
• Air-gapped — fully disconnected deployment for defense and critical infrastructure.
• Bring your own GPU — your hardware stays yours; Plugsky provides the control plane, API, metering, and support.

These deployments are scoped per environment and quoted separately. They are not the public shared network, and they are not self-serve. If you need region-locked routing, customer-managed keys, audit logs, or a DPA, start with the GPU team.

How to start

1. Rent capacity: create a Plugsky account and connect an OpenAI-compatible client to https://api.plugsky.com/v1.
2. Rent out a GPU: run the preflight, register at /nodes, install the worker, and wait for approval.
3. Need isolation: talk to the GPU team about a dedicated, VPC, on-prem, or air-gapped deployment.
4. Model the economics: use the earnings calculator and the GPU cloud cost calculator before committing.