Replicate pricing: GPU rental rates
Visit Replicate →Updated July 2026 · day 14 of the ledger · verified 2026-07-18
The cheapest GPU on Replicate right now is the Tesla T4 at $0.81/hr (serverless, verified 2026-07-18). Replicate is a model-serving platform: you run public models or deploy your own on its managed hardware, billed per second. It is not a VM cloud, so there is no SSH box to rent, and we label its rates serverless. The public hardware table prices single-GPU T4, L40S, A100 80GB and H100 tiers, which we track daily; larger multi-GPU configurations are gated behind committed-spend contracts and are not public on-demand prices, so we exclude them.
Live Replicate prices (2026-07-18)
| GPU | Tier | $/hr | Verified |
|---|---|---|---|
| Tesla T4 | serverless | $0.81 | 2026-07-18 |
| L40S | serverless | $3.51 | 2026-07-18 |
| A100 (unspecified) | serverless | $5.04 | 2026-07-18 |
| H100 (unspecified) | serverless | $5.49 | 2026-07-18 |
Cheapest listed per-GPU-hour rates verified at the source. Before storage, egress and networking.
Pricing model
Replicate bills hardware by the second, with the pricing page publishing both per-second and per-hour figures per hardware tier. Each GPU tier bundles fixed CPU and RAM (for example the H100 tier pairs the GPU with 13 vCPUs and 72GB RAM). You pay for the time your model runs on the hardware rather than renting a persistent machine, which is why we classify Replicate offers as serverless alongside platforms like fal and Modal.
Hardware tiers
Public single-GPU tiers cover Nvidia T4, L40S, A100 80GB and H100, plus CPU-only tiers. Two-GPU A100 and L40S tiers are listed at exact price multiples of the single-GPU tier. Beyond that, the page states that additional multi-GPU A100, H100 and L40S capacity is available with committed-spend contracts, so 4x and 8x configurations are not public on-demand prices and are excluded from our tracker.
Who should pick Replicate
Pick Replicate to run or deploy models without managing servers, containers or scaling, paying per second only while the hardware works. It fits teams shipping model-backed features who want managed infrastructure. It is not a fit for workloads needing root access, custom drivers, persistent local storage or raw VM rental; for those, compare the on-demand VM providers in our tables. A formal public uptime SLA was not found on the pages reviewed. Not publicly documented.
FAQ
No. Replicate is a model-serving platform billed per second. You run models on its managed hardware tiers rather than renting a virtual machine, so we label its rates serverless.
Single-GPU T4, L40S, A100 80GB and H100 tiers, each bundled with fixed CPU and RAM. Larger multi-GPU configurations require committed-spend contracts.
The pricing page states that additional multi-GPU capacity is available with committed-spend contracts, so those rows are not public on-demand prices. We only publish prices anyone can pay without a contract.
Profile verified 2026-07-18. Sources: replicate.com · replicate.com