The right Replicate alternative depends on which half of Replicate you use: calling hosted models by API, or running your own training. For training and fine-tuning runs that have to finish, Nodus lists the NVIDIA H100 at $2.60/hr on its pricing page, against Replicate's $5.49, and restores your latest checkpoint on a new machine if one is lost. For serving models, Modal, RunPod Serverless, Hugging Face Inference Endpoints, Together AI and Baseten are the usual picks, and Nodus inference is still in Beta.
We build Nodus, so read this with that in mind. Every competitor fact below links to that company's own pricing page or docs, the arithmetic is shown once, and one section covers where Replicate is the better choice.
Which Replicate alternatives are worth a look?
Six platforms come up again and again, and they are not interchangeable. Some are mainly model APIs, some host your own container, and one is built around long jobs.
| Best for | H100 per hour (listed) | How billing works | Training | When a machine is lost | |
|---|---|---|---|---|---|
| Replicate | Calling public models; packaging your own with Cog | $5.49 | Per second; private models also pay setup and idle time | Trainings for supported models, billed per second | Not stated on the pages we checked |
| Nodus | Training and fine-tuning runs that must finish | $2.60 (H100 SXM) | Prepaid credits; per-run spend caps | Any script with nodus run; managed TrainingJobs (Beta) |
Urgent checkpoint, then a replacement restores it |
| Modal | Python functions that scale up and down | $3.95 (H100 SXM5) | Per second; $30/month free credit on Starter | Your own code | Function restarts on the same input; you write checkpoints |
| RunPod Serverless | Custom inference workers | $5.39 (H100 PRO) | Per second; flex workers scale to zero | Inference product (Pods cover long-running jobs) | Not stated on the pricing page |
| Hugging Face Inference Endpoints | Deploying Hub models on dedicated instances | $10.00 (on GCP) | By the minute; optional scale to zero | Inference only | Not applicable |
| Together AI | Open-model APIs and token-priced fine-tuning | $3.99 on demand, $1.99 preemptible | Per token for fine-tuning, with a per-job minimum | LoRA and full fine-tuning | Not stated on the pricing page |
| Baseten | Dedicated deployments packaged with Truss | $6.50 ($0.10833/min) | Per minute; no charge for idle time | Training product, same GPU rates | Not stated on the pricing page |
Sources: Replicate pricing, Replicate billing docs, Modal pricing, Modal preemption docs, RunPod pricing, RunPod Serverless pricing docs, Hugging Face Inference Endpoints pricing, Together AI pricing, Baseten pricing, the Nodus pricing page and the Nodus recovery docs. Prices checked October 30, 2026.
Which alternative costs the least per GPU-hour?
On listed H100 rates, Nodus is lowest at $2.60. Replicate lists a single H100 at $0.001525 per second, which is $5.49 an hour. Modal lists its H100 at $0.001097 per second, or $3.95 an hour.
Together AI lists on-demand H100s at $3.99 per GPU-hour, and $1.99 for preemptible capacity. RunPod's Serverless table lists the 80 GB H100 PRO tier at $5.39. Baseten's per-minute rate of $0.10833 works out to $6.50 an hour. Hugging Face lists a single H100 endpoint on Google Cloud at $10 an hour.
Replicate sits in the upper middle; only Baseten and Hugging Face list a higher H100 rate.
These rates are not all the same product. Replicate's H100 comes with 13 vCPUs and 144 GB of memory, and Hugging Face prices a whole managed endpoint. The Nodus pricing page lists CPU and memory as separate lines, so a heavy data loader adds to the Nodus figure. Price your real job, not one table row.
What does a 10-hour fine-tune on one H100 cost on each?
We priced the same job on the platforms that run training: one H100 for 10 hours. Each rate is the listed rate above, times 10.
| Platform | Rate used | 1 x H100 for 10 hours |
|---|---|---|
| Nodus | $2.60/hr | $26.00 |
| Modal | $0.001097/s ($3.95/hr) | $39.49 |
| Together AI, on demand | $3.99/hr | $39.90 |
| Replicate | $0.001525/s ($5.49/hr) | $54.90 |
| Baseten | $0.10833/min ($6.50/hr) | $65.00 |
The per-second and per-minute rates were multiplied out exactly: 36,000 seconds at $0.001097 is $39.49, and 600 minutes at $0.10833 is $65.00. Nodus comes in $28.90 below Replicate on this job, which is 53% less.
At list prices, the same 10 GPU-hours cost about half as much on Nodus as on Replicate.
Replicate does bill trainings per second for the hardware they use. Its multi-GPU H100 machines are listed too, but the pricing page notes that additional multi-GPU H100 capacity comes with committed spend contracts.
Together AI also offers a different model: you pay per training token, not per hour. Its pricing page lists supervised fine-tuning of Llama 3.1 8B from $0.34 per million tokens, with a $4.00 minimum per job. For a small dataset on a model Together supports, that can beat any hourly rate. Our Together AI alternatives guide covers when it does.
What happens if the machine is lost halfway through training?
This is where the platforms differ most, and where the bill can double. A 2-minute training job rarely meets a failure. A 10-hour one has many more chances to.
Modal is clear about this. Its docs say all Modal Functions are preemptible by default and are restarted on the same input. Modal asks long training functions to save checkpoints themselves, and says its non-preemptible option is not supported for GPU Functions.
On Nodus the recovery loop is built in. When interruptible capacity is reclaimed, Nodus asks your run for an urgent checkpoint and prepares a replacement machine at the same time. The new attempt restores your latest committed checkpoint, so a checkpointed run loses at most the work since its last save. Your script writes its state under /nodus/state and loads it on start, as the checkpoints guide shows.
Here is the arithmetic for the 10-hour job, lost at hour 8, at $2.60 an hour. Without a checkpoint, you repeat 8 hours, which is $20.80 of extra compute. With a checkpoint every 30 minutes, you repeat at most half an hour, or $1.30.
Checkpoints that restore automatically turn a lost machine into a rounding error.
Our decision rule: if losing the machine at hour 8 would cost you a night or a deadline, pick a platform that restores checkpoints for you. Pair it with a spend cap. On Nodus, --max-cost caps one run, and a Budget with action: Block refuses new work once it is used up.
Can Nodus replace Replicate for inference?
Partly, and it is still Beta. Nodus inference is an OpenAI-compatible API over a catalog of open models such as GPT-OSS, Whisper and Gemma. Closed models like GPT, Gemini and Claude are not in the public catalog. A named InferenceEndpoint adds its own URL, rate limits and a spending cap.
It is not a one-click home for a custom Cog model. To serve your own weights on Nodus today, you write the serving code as a Function and deploy it as an App. If you want a closer match to Replicate's model, these fit better:
- RunPod Serverless, if you already have a Cog image. RunPod publishes a guide for moving a Cog model from Replicate to a Serverless worker. Flex workers scale to zero and bill per second, with a default 5-second idle timeout.
- Hugging Face Inference Endpoints, if the model lives on the Hub. Billing is by the minute, and endpoints can scale to zero after 15 idle minutes, with a cold start afterwards.
- Baseten, for a dedicated deployment. It packages models with Truss, its open-source format, and says you do not pay for idle time.
- Modal, if you want to write the server in plain Python and let it scale. Our serverless GPU platforms guide compares the serving side in more depth.
Where is Replicate the better choice?
Replicate wins in several specific cases, and we would tell you to stay.
- You want to call a public model with no setup. On public models you pay only for active time; setup and idle time are free.
- You run short, packaged trainings. Replicate's FLUX guide says a typical fine-tune runs on 8 x H100 for about 2 minutes and costs about $1.46. Nothing here beats that for a quick image LoRA.
- You already package models with Cog, Replicate's open-source packaging tool. Replicate documents pushing a Cog model as a cloud API, so there is nothing new to learn.
- You want a fast-booting fine-tune. Replicate bills those only for active time, because they share a pool running the base model.
The trade-off shows up with private models and deployments. Replicate bills those for setup, idle and active time while instances are online. An H100 private model left online for 24 hours costs $131.76 at $5.49 an hour, whether or not requests arrive.
How do you move a Replicate training job to Nodus?
If your trainer is a Python script, the move is one command from its directory. Install the CLI, sign in and submit the run:
pip install nodus-compute
nodus login
nodus run --gpu H100:8 --max-cost 40 -d -- torchrun train.py
That nodus run line is the example from the nodus run reference: 8 H100s, a $40 spend cap, and -d to return once the Job exists. Add --dry-run first to see the estimate without creating the Job, then follow it with nodus logs -f, as the quickstart shows.
If you used Replicate only to call a model, point the OpenAI SDK at Nodus. This is the snippet from the inference docs:
from openai import OpenAI
client = OpenAI(base_url="https://inference.nodus-compute.ai/v1", api_key="YOUR_NODUS_API_KEY")
reply = client.chat.completions.create(
model="nodus/gpt-oss-120b",
messages=[{"role": "user", "content": "What is the capital of France?"}],
)
print(reply.choices[0].message.content)
If you want a managed fine-tuning flow instead, the training guide shows TrainingJobs for SFT, DPO and GRPO, and our Modal alternatives for training and RunPod alternatives guides cover the related choices.
FAQ
Is Replicate good for fine-tuning LLMs?
It is strongest for short trainings of models that already support it, such as FLUX image LoRAs. For long LLM runs you write yourself, an hourly platform with automatic checkpoint restore is usually cheaper and safer.
Can I connect my Replicate account to Nodus?
No. Replicate is not one of the clouds listed on the Nodus cloud accounts page. You run the job on Nodus directly instead.
Which Replicate alternative has a free tier?
Modal's Starter plan includes $30 of compute a month, and Baseten says new accounts get credits to experiment. Replicate itself lets you run select models free before asking for billing.
Is there a Replicate alternative that bills per token?
Yes. Together AI prices fine-tuning per training token with a per-job minimum, and model APIs on Together, Nodus and Replicate bill per token for language models. Token pricing suits small datasets; hourly pricing suits long custom runs.
Originally published on the Nodus blog.
Top comments (0)