Layer 2.5 · head to head
Replicate vs Modal
Ship a feature in an afternoon, or own the deployment model.
| Replicate | Modal | |
|---|---|---|
| Can you leave | Sticky. Leaving means rewriting against a different interface, not changing a hostname. | Sticky. Leaving means rewriting against a different interface, not changing a hostname. |
| Counterparty | Low. Models are containerised with an open packaging format, so the exit is unusually clean for this layer. | Low financially; high in code. Modal's programming model is genuinely delightful and genuinely non-portable — that is the trade. |
| What it is | You never see a GPU. You send a request and pay per token or per second of execution. | You never see a GPU. You send a request and pay per token or per second of execution. |
| How you reach it | You call an endpoint. There is no infrastructure to see, and no infrastructure to move. | You supply code; scaling and idle time are the provider's problem. |
| Accelerators | A100, H100, L40S, T4 behind hosted models | H100, A100, L40S, T4, and others |
| Regions | US | US, EU |
| Pricing model | Per second of prediction time, per hardware class. | Per second of GPU time, scale to zero, no idle charge. |
| Getting started | API key. | Self-serve, generous free tier historically. |
| Capacity | Managed. | Managed pool. |
| Ownership | Private. | Private. |
Replicate
For: Product teams shipping a model-backed feature who do not want to think about GPUs at all.
The catch: Cold starts on unpopular models are long. The convenience premium over renting the same GPU directly is large and worth calculating once.
Economics: The most expensive way to buy a GPU-second and often the cheapest way to ship a feature. Both are true; which one matters depends on your volume.
Modal
For: Python teams who want infrastructure to disappear and are happy writing to a framework to get it.
The catch: Your deployment becomes Modal-shaped. Migrating off is a rewrite, not a redeploy. Price that in on day one, not year two.
Economics: Per-second billing with no idle cost is the cheapest possible shape for bursty inference. It is the wrong shape for a job that runs continuously.
Neither table row is a price
Deliberately. Published on-demand rates at this layer move weekly, and essentially nobody signing a real contract pays them — every serious buyer pays less than every list figure either of these companies publishes. Quoting one here would date this page within a month.
The GPU rental price index carries dated, sourced figures instead, and the durable finding there is the spread: the identical H100 rents from roughly $1.38 to $12.29 an hour depending only on who you rent it from.
The layers underneath both
Whichever you pick is renting you chips in a building that needs power. In 2026 that is the constraint that binds: Microsoft has disclosed an Azure backlog it cannot fill for want of megawatts rather than accelerators, and the US interconnection queue exceeds 2,600 GW with roughly 80% of projects withdrawing before they energise.
Layer 2 — Silicon · Layer 1 — Energy · The interconnection queue