Replicate has the shortest distance in this market between having nothing and having a working model behind an API. That is a real achievement and it is the reason the platform is worth the money for a lot of people.
The trouble starts at the second GPU.
Every multi-GPU configuration on Replicate requires a committed spend contract. Two A100s, four H100s, eight of anything: all of it is a sales conversation. So is the H200 outright. The self-serve tier that makes the platform so easy to start on tops out at exactly one card, and the moment your model needs more than that, the thing you liked about Replicate stops being available to you.
That is not a flaw so much as a boundary, and it is worth knowing where it sits before you build on top of it. What follows is where people go when they reach it. Prices are as of 18 August 2026, taken from each provider's own pricing page.
Why people leave Replicate, and it is not always the bill
Replicate sells convenience. People leave when they outgrow it, and they outgrow it in four different directions.
They hit the one-GPU ceiling. The most common one, and the least negotiable. There is no self-serve path to multi-GPU.
They want control of the environment. Replicate models are packaged with Cog, which is pleasant until you need a dependency, a CUDA version or a serving stack it does not anticipate. At that point you want your own container.
They need a machine, not an endpoint. Replicate hosts models. It does not give you somewhere to sit and develop, or a cluster to train on. If your next step is fine-tuning rather than serving, the platform does not go there.
The cost stops making sense at volume. Convenience pricing is worth paying at low volume and stops being worth it at some point that depends entirely on your traffic. This is the reason that gets written about most and it is the last one to bite.
Which of those four is yours determines the answer, and the four answers are different platforms.
Comparison table
What I was actually comparing on
Can you scale past one GPU without a contract? This is the Replicate boundary, so it is the first question to ask of anything replacing it. Most of this list clears it. Two do not.
Can you bring your own container? Cog is a packaging format. If you left because it would not accommodate your stack, a platform that swaps it for a different opinionated format has not solved anything. Standard Docker images matter here.
Does it cover the stage after the one you are on? Serving and training are different problems, and several platforms here do exactly one of them. Replatforming later is a cost nobody puts on a pricing page.
What unit are you buying? One provider on this list sells eight GPUs at a time. One bills per minute. One bills per second but charges triple for non-preemptible execution. A per-hour figure that ignores the unit is not a price.
What does it cost at your volume? Last, and deliberately so. The right question is not the hourly rate but the rate multiplied by how long your workload actually runs, which is a different ranking for a model called twice a day than for one called constantly.
Runpod
Why it works here
Runpod answers the specific thing that Replicate does not: multi-GPU is self-serve. Instant Clusters are provisioned from the console without a contract or a sales call, so the ceiling that sends people looking is not there.
It also covers the stage after serving. Pods for development and long-running jobs, Serverless for inference that scales to zero, Instant Clusters for multi-node training, all on one account with the same container images. Moving between them is a configuration change rather than a migration.
What you get
Standard Docker images, not a proprietary packaging format. Bring the container you already have or start from a community template. Deploy time is under 30 seconds.
The good
Per-second billing on Pods and Serverless with no minimum, and no charge during provisioning. No ingress or egress fees, so leaving is not expensive either. Sub-200ms cold starts via FlashBoot, which matters if your models are called infrequently enough to go cold. More than a million developers have used it.
The not so good
You are operating more of the stack than you were on Replicate. Replicate's value is that it makes decisions for you, and taking those decisions back is work that you now own. If what you wanted was a model behind a URL with no infrastructure to think about, that is a genuine trade rather than a straight upgrade.
The lineup also does not go as high as some of these. There is no GB300 or rack-scale NVLink option.
How it performs
- GPU selection: B300 through RTX A5000, plus AMD MI300X at $2.39/hr on Secure Cloud
- Pricing: from {{gpu:rtx-a5000:community}}/hr on Community Cloud. On Secure Cloud, H100 PCIe {{gpu:h100-pcie}}/hr, A100 PCIe {{gpu:a100-pcie}}/hr, L40S {{gpu:l40s}}/hr
- Multi-GPU: Instant Clusters, self-serve
- Billing: per second on Pods and Serverless, no minimum
- Compliance: SOC 2 Type II
- Coverage: 31 global regions
What it costs
On the cards Replicate also sells, Runpod's Secure Cloud rates are {{gpu:h100-pcie}}/hr for an H100 PCIe against their $5.49, {{gpu:a100-pcie}}/hr for an A100 80GB against their $5.04, and {{gpu:l40s}}/hr for an L40S against their $3.51. Storage is $0.10/GB/month for container disk and $0.07/GB/month for network storage under 1TB, $0.05 above. No ingress or egress charges.
Modal
The closest thing here to Replicate's developer experience without Replicate's ceiling. You decorate a Python function, it deploys as a serverless endpoint, and there is no orchestration layer to configure. Multi-GPU autoscaling is self-serve.
Read the multipliers before you budget. Rates are per second, working out to $3.95/hr for an H100 SXM5, $2.50 for an A100 80GB, $1.95 for an L40S and $0.59 for a T4. But region selection costs 1.5 to 1.75 times base, and non-preemptible execution costs three times base. A workload that needs a specific region and cannot tolerate preemption is not paying the headline number.
Volumes are $0.09/GiB/month with the first TiB free. Plans start free and step to $250/month for Team.
The limit is scope. Modal is serverless only. No persistent development machines, no multi-node training. If you left Replicate because you needed somewhere to fine-tune, Modal has the same gap.
Best for: teams who liked Replicate's simplicity and hit the GPU ceiling rather than the container ceiling.
Baseten
Managed production inference with the strongest operational story here. You do not pay for idle time, only for the minutes a model spends deploying, scaling or predicting, which for spiky traffic is a real difference from paying by the hour. SOC 2 Type II and HIPAA compliant.
Models are packaged with Truss, their open-source format. That is worth noting carefully: if you are leaving Replicate specifically because Cog would not accommodate your stack, you are trading one packaging format for another rather than getting standard Docker.
Baseten quotes per minute. An H100 is $0.10833/min, which is $6.50/hr. An A100 80GB is $4.00/hr, a B200 is $9.98, an L4 is $0.85 and their cheapest is a T4 at $0.63. The same rate card covers dedicated deployments and training.
They are also the only platform here that runs inside your own cloud, with self-hosted and hybrid deployments and data residency control on the Enterprise tier. If you are moving off Replicate for compliance reasons rather than scaling reasons, that is the entire decision and the hourly rate is noise.
Best for: teams who want inference operated for them, and anyone with a data residency requirement.
fal.ai
The most direct substitute if what you run on Replicate is image or video generation. fal is built around generative media and prices its model APIs the same way Replicate prices public models: per output unit rather than per GPU-hour, at rates like $0.05 per second of video or $0.03 per image.
On raw GPU rental their list prices are H100 $4.50/hr, H200 $4.50, B200 $6.25, B300 $8.50 and RTX PRO 6000 $2.99. That undercuts Replicate on the H100 by a dollar. Their pricing page also shows an "as low as" column reaching $1.89 for an H100, which is negotiated rather than self-serve.
Multi-GPU goes through sales, so if the one-GPU ceiling is your reason for leaving, fal reproduces it.
Best for: image and video workloads, where the per-output pricing model and the model catalog are the point.
Together AI
Worth considering if what you host on Replicate is an open language model rather than a media model. Together AI serves and fine-tunes open models, and their GPU clusters are self-serve.
On-demand rates are $3.99/hr for an HGX H100, $5.99 for an H200 and $8.19 for a B200, with reserved commitments bringing the H100 to $3.19 at 24 months. They publish nothing below the H100, so a small model that ran happily on a T4 has no cheap home here.
Best for: open LLM serving and fine-tuning, at a scale where an H100 is the right floor.
Fireworks
Open-model serving where most of the value is in per-token inference rather than GPU rental. If you have been calling models on Replicate and want someone else's optimization work rather than a raw GPU, this is the same shape of product one tier up.
Their on-demand GPUs are the most expensive on this list: $7.00/hr for an H100 or H200, $10.00 for a B200, $12.00 for a B300 and $18.00 for a GB300. Region-restricted deployments cost a 1.5x premium.
A price increase lands 1 September 2026 and both columns are published now. From that date the H100 and H200 go to $8.00, the B200 to $13.00, the B300 to $15.00 and the GB300 to $20.00.
Best for: open-model serving at scale on a per-token basis, not GPU rental.
Lambda
Raw GPU instances, no model-hosting layer at all. This is a bigger jump than the others: you are not replacing Replicate so much as deciding you would rather run the infrastructure.
On-demand H100 SXM is $3.99/hr, A100 SXM 80GB is $2.79, A100 40GB is $1.99, B200 SXM6 is $6.69 and a Tesla V100 is $0.79. 1-Click Clusters run B200 from $9.86/GPU/hr at 16 GPUs down to $8.87 at 256 or more. Prices exclude sales tax, VAT and GST.
Best for: teams who want instances and clusters and are content to build the serving layer themselves.
CoreWeave
Capacity at a scale most people leaving Replicate are nowhere near. CoreWeave sells eight GPUs at a time, and reading their pricing any other way will mislead you. HGX H100 is $49.24/hr for the instance, which is $6.16 per GPU. An 8x A100 is $21.60, or $2.70 per GPU. A single GH200 is $6.50.
Spot capacity runs around 40% of on-demand and egress, ingress and transfer are free. Reserved capacity goes up to 60% off.
There is no self-serve tier and no small experiment. If you are moving off Replicate because two GPUs required a phone call, note that here the phone call is the front door.
Best for: teams whose minimum unit really is eight GPUs, committed for months.
Cold starts: who is actually faster than Replicate
This is the most searched version of the question and it is built on a premise worth taking apart first.
What Replicate says about its own cold boots. Their docs describe a prediction sitting in starting for more than a few seconds as a new worker being booted. Their engineering blog puts the range plainly: a cold boot "can be as slow as a few minutes for large models." They also shipped sub-second boots for fine-tuned Llama 2 and SDXL models. So the honest version of Replicate's number is not one number, it is a range from under a second to several minutes depending on the model and how it was created.
What the alternatives publish:
These numbers are not comparable. Modal's own documentation makes the point better than I can: containers boot in about a second, but a container is not warm until your global scope and any enter methods have finished, and Modal says that step "can range from seconds to minutes." A 200ms resume and a one-second container boot are measuring different events, and neither includes loading a 40GB model off disk.
The thing that actually determines your cold start is your model size, your image size, and whether weights are baked in ahead of time. Every platform here gets dramatically faster when weights are pre-loaded rather than pulled at boot, and dramatically slower when they are not. That is a property of your deployment, not of the vendor.
And on Replicate specifically, cold boots are a configuration choice rather than a ceiling. Create a deployment, set minimum instances to one or more, and the model stays warm. Replicate also does not bill for the boot: a prediction that never starts is aborted and costs nothing, and you pay only for time the prediction actually ran. So the real question is not who boots faster. It is what it costs you to never be cold, which is a pricing question in disguise, and there the hourly rates in this article apply.
If cold starts really are your reason for leaving, the two options that change the shape of the problem rather than the size of it are Fireworks, which serves from shared always-on capacity so there is no boot to wait for, and Runpod's FlashBoot for workloads that must scale to zero between calls. Everything else is the same trade you already have.
Replicate in production
The second question people ask is whether Replicate is a production platform or a prototyping one. It is both, with specific edges that decide which.
The ceiling is the main one. Self-serve stops at a single GPU. Every multi-GPU configuration is a committed spend contract, so the platform's scaling story past one card runs through a sales team rather than a console. For a production service with growing traffic, that is a commercial dependency and not just a technical one.
Predictions time out after 30 minutes. Fine for inference, and a hard wall for anything batch-shaped. Longer timeouts require contacting support.
The billing behavior is better than its reputation. You are billed for time the prediction actually ran. A prediction that never starts costs nothing, and a canceled one bills only for what ran. Combined with deployments and minimum instances, that is a more controllable cost model than "convenience pricing" implies.
What you give up is control of the environment. Cog is the packaging format, and production teams tend to hit its edges around CUDA versions, serving stacks and dependency pinning before they hit anything else. If your platform team has opinions about the container, this is where the friction lives.
Where the alternatives are stronger for production specifically:
- Self-serve scale past one GPU – Runpod, Modal, Lambda, Together AI and Fireworks all clear this. fal does not.
- Standard Docker images – Runpod, Modal, Lambda and CoreWeave. Baseten uses Truss, its own format.
- Data residency and self-hosting – Baseten is the only one here that deploys into your own cloud.
- A training tier on the same platform – Runpod, Baseten, Together AI and Lambda. Modal has none.
- Compliance – Runpod is SOC 2 Type II. Baseten publishes SOC 2 Type II and HIPAA. Check the others against your own requirements rather than trusting a comparison table, this one included.
The summary a production team probably wants: Replicate is a strong choice up to the point where you need a second GPU, your own container, or a training step, and those three arrive together more often than not. Nothing about it is unfit for production. It is fit for a particular shape of production, and the shape is narrower than the onboarding experience suggests.
FAQ
Which inference providers offer lower cold start times than Replicate?
On published figures, Runpod claims sub-200ms via FlashBoot, Modal documents container boots of about one second, and Fireworks states there are no cold starts on its serverless tier because capacity is always on. Replicate's own range runs from under a second for certain fine-tuned models to, in their words, "a few minutes for large models." The catch is that these figures measure different events and none of them includes loading your model weights, so they are not a ranking. Your model size and whether weights are pre-loaded matter more than the vendor. On Replicate you can also remove cold boots entirely by creating a deployment with minimum instances set to one or more, and Replicate does not bill for boot time.
How does Replicate compare to other inference providers for production use?
It is production-capable within a specific shape. Self-serve stops at one GPU and every multi-GPU configuration requires a committed spend contract, predictions time out after 30 minutes, and the environment is defined by Cog rather than a standard Docker image. Against that, you are billed only for compute that actually ran, and deployments with minimum instances give you real control over latency. Runpod, Modal, Lambda, Together AI and Fireworks all scale past one GPU without a contract. Baseten is the only one here that runs inside your own cloud. Whether Replicate is the right production platform depends on whether you will hit the one-GPU ceiling, the container ceiling or the training gap, and most teams that outgrow it hit more than one at once.
Why is Replicate expensive?
It is priced for convenience rather than for compute. Their H100 is $5.49/hr, the A100 80GB is $5.04 and the L40S is $3.51. Those cover the packaging, the API layer, the autoscaling and the model catalog as well as the card. Whether that is expensive depends on how much of that work you would otherwise do yourself.
Can you run multiple GPUs on Replicate?
Not on the self-serve tier. Every multi-GPU configuration, from 2x through 8x on the A100, H100, H200 and L40S, requires a committed spend contract, as does the H200 in any configuration. An 8x H100 is listed at $43.92/hr under contract. This is the most common reason people go looking for an alternative.
What is the cheapest Replicate alternative?
Runpod at the low end, where Secure Cloud starts at {{gpu:rtx-a5000}}/hr and ten cards sit at or under {{gpu:l40s}}/hr. On the cards Replicate also sells, an H100 PCIe is {{gpu:h100-pcie}}/hr against their $5.49 and an A100 PCIe is {{gpu:a100-pcie}}/hr against their $5.04. Replicate's own floor is a T4 at $0.81, a 2018 card.
Which alternative is closest to Replicate's ease of use?
Modal, if your workload is Python and serverless. Baseten, if you want the operational side handled and are comfortable packaging with Truss. Both keep more of the "model behind a URL" experience than moving to raw GPUs does.
I only left because of Cog. What should I look at?
Anything that takes a standard Docker image, which means Runpod, Modal, Lambda or CoreWeave. Be careful with Baseten here, since Truss is a packaging format too. Swapping one opinionated format for another does not fix the problem you left over.
Does anything match Replicate's per-model pricing?
fal.ai, for generative media. They price model APIs per output unit, at rates like $0.05 per second of video or $0.03 per image, which is the same billing shape as Replicate's per-run public model pricing. Fireworks does the equivalent for language models on a per-token basis.
What should I check before switching?
Whether the alternative reproduces the limit you are leaving over. fal routes multi-GPU through sales, so it does not fix a scaling ceiling. Baseten uses its own packaging format, so it does not fix a container problem. Modal has no training tier, so it does not fix a lifecycle problem. Match the fix to the reason.
Making the right choice
Replicate is a good product with a hard edge, and where you go next depends entirely on which edge you hit.
If you hit the one-GPU ceiling, you need self-serve multi-GPU. Runpod, Modal, Lambda and Together AI all clear that. fal does not.
If Cog would not take your stack, you need standard Docker images. Runpod, Modal, Lambda and CoreWeave.
If you needed to fine-tune rather than serve, you need a platform with a training tier. Runpod, Baseten, Together AI and Lambda. Modal does not have one.
If your data cannot sit on someone else's infrastructure, Baseten is the only option here that deploys into your own cloud.
If you generate images or video, fal's per-output pricing is the closest thing to what you already had.
If the cost stopped making sense at volume, compare on your actual runtime rather than the hourly rate, and expect the answer to be Runpod or Modal depending on whether you need more than serverless.
If more than one of those is true – and past a certain size they usually all become true at once – the argument for Runpod is that Pods, Serverless and Clusters sit under one account. The next ceiling is not a migration.
