What are the best Crusoe Cloud alternatives for AI teams?

For a small AI team, the best Crusoe Cloud alternatives are the ones that sell you one H100 at a time, because Crusoe's smallest H100 virtual machine has eight GPUs. Nodus lists the NVIDIA H100 SXM at $2.60 per GPU-hour, runs a one-GPU job as a Job that stops billing when it ends, and enforces a dollar cap on every run. RunPod, Lambda and Modal also sell single H100s, while Crusoe stays the better fit for large reserved clusters, the newest NVIDIA and AMD systems, and free data transfer.

We build Nodus, so weigh this accordingly. Every fact about another company links to its own pricing page or docs, and every cost below is written out so you can redo it with today's prices.

Who are Crusoe's main competitors?

Crusoe competes in two groups. The first is other clouds that sell large NVIDIA clusters to AI companies, such as CoreWeave, Lambda and Nebius. The second is platforms built around smaller, shorter workloads. RunPod sells per-second GPU pods and Modal runs serverless Python functions. Nodus runs training, fine-tuning, inference and agent jobs on the cheapest capacity that will finish them.

Crusoe's own pages lean toward the first group. Its pricing page leads with NVIDIA Vera Rubin, GB300 NVL72 and AMD MI355X systems and points you to sales to "lock in guaranteed resources at our lowest rates". Its reservations docs use 128 A100 GPUs and 256 H100 GPUs as examples. That suits a team buying at that scale.

If you are three people fine-tuning one model, the questions below matter more.

Can you rent a single H100 on Crusoe?

No. Crusoe's VM specifications list one H100 type, h100-80gb-sxm-ib.8x, with eight NVIDIA H100 80GB SXM5 GPUs. Its H200, B200 and B300 VMs are also eight-GPU machines. The single-GPU sizes on that page are the NVIDIA L40S 48GB and the A100 80GB PCIe.

That makes the smallest H100 purchase on Crusoe 8 x $3.90 = $31.20 per hour at its current on-demand price. The pricing page shows the H100 moving to $4.50 per GPU-hour on December 1, 2026, which makes the same VM $36.00 per hour.

Here is what that means for a common small-team job: a LoRA fine-tune that needs one H100 for 6 hours. We priced it from each provider's public list prices, GPU compute only.

Prices checked November 8, 2026: Nodus pricing, Crusoe pricing, RunPod pricing, Modal pricing and Lambda pricing. Modal lists $0.001097 per second, multiplied by 21,600 seconds. Lambda's page says its listed prices take effect November 16, 2026.

A one-GPU job on Crusoe pays for eight H100s, so the per-hour rate stops being the number that matters.

You can get closer on Crusoe by dropping to a single-GPU type. Its pricing page lists the A100 80GB PCIe at $2.00 per hour and the L40S at $1.50. For comparison, Nodus lists an A100 80GB at $1.09 and an L40S at $0.793. Whether your job fits on those cards is a memory question, and our RTX 5090 vs A100 vs H100 guide walks through it.

What does a full 8 x H100 run cost on each platform?

When you actually need eight GPUs, Crusoe's price is competitive. We priced one training run of 8 x NVIDIA H100 for 10 hours, which is 80 GPU-hours, with storage and taxes left out everywhere.

Same sources and check date as the table above. Modal's per-second rate is multiplied by 288,000 GPU-seconds.

At full-machine scale Crusoe's current H100 rate is below Modal and RunPod Secure Cloud, and Nodus is $104 cheaper for the run.

Crusoe also bills on-demand VMs per second, according to its VM billing docs. A stopped on-demand VM is not charged for compute, though its 128 GB OS disk still costs $0.08 per GiB per month, about $10.24 a month, until you delete it. That is a fair setup. The difference on Nodus is that there is no VM to stop: a Job that has Succeeded, Failed or been Cancelled has nothing running or billed, per the Jobs guide.

Can a small team use spot GPUs or reservations on Crusoe?

Yes, but both go through people. Crusoe's spot docs say you get spot instances by contacting Crusoe sales or your support representative, and that they can be interrupted when Crusoe needs the capacity back. Reservations are signed with sales too. The reservations docs add that unused reserved capacity does not roll over, and the VM billing docs say a reserved VM is billed for the whole commitment whatever its state.

Quotas follow the same pattern. Crusoe's quotas page says you contact support to raise a quota, and that a quota is not a guarantee of capacity.

On Nodus, interruptible capacity is a flag on the run, not a conversation. The attempts and recovery docs say that on a reclaim notice Nodus asks your job for an urgent checkpoint while it prepares a replacement machine. A Checkpointed run loses at most the work since its last committed checkpoint. Our spot vs on-demand guide covers when that trade is worth it.

How do you stop a forgotten GPU from running up a bill?

Use a limit the platform enforces before and during the run, not an alert you read the next morning. Crusoe's budget alerts send email, console, Slack or webhook notifications, and are evaluated daily shortly after midnight UTC. Its billing dashboard also updates daily. We did not find a hard spending cap in those docs.

A day between checks is real money at eight-GPU granularity. An 8 x H100 VM left running after a job costs $31.20 per hour, so one full day is 24 x $31.20 = $748.80 before the next evaluation. Crusoe's billing FAQ says a running VM is billed whether or not it is processing work.

Nodus checks spending limits before paid work starts and every few minutes while it runs, according to the budgets guide. A Job that reaches its --max-cost saves its state and becomes Suspended with reason MaxCostReached. With --max-cost 20, the most a forgotten run can spend is $20, whatever GPUs it holds. For a small team where one person launches most runs, that cap is cheap insurance.

An alert tells you what happened yesterday; a cap decides what can happen today.

Which Crusoe alternative should a small AI team pick?

Pick by the smallest unit of work you run most often.

• Your typical job needs one to four GPUs for hours, not weeks: Nodus. You pay for the GPUs the job uses, each run carries a dollar cap, and checkpoints are restored on a new machine if capacity is reclaimed.
• You want a fixed-price pod you can SSH into and keep: RunPod. Its Community Cloud H100 SXM is $2.69 per hour. Our RunPod alternatives guide goes further.
• Your work is short Python functions behind web endpoints: Modal, billed per second.
• You want a single H100 or a whole 8-GPU machine from a self-serve console: Lambda. See Nodus vs Lambda for that trade.
• You are signing for hundreds of GPUs for months: talk to Crusoe and its peers. Our Nodus vs CoreWeave post covers that end of the market.

Crusoe is not a cloud account you can connect to Nodus. The clouds Nodus supports are listed in the cloud accounts guide.

Where is Crusoe the better choice?

Crusoe wins when you need large, connected clusters and the newest hardware. Its 8 x H100 VM comes with a 3,200 Gbps InfiniBand network, and its B300 VM lists 6,400 Gbps. Its pricing page also offers GB200 and GB300 NVL72 systems, Vera Rubin and AMD MI355X. Crusoe Managed Kubernetes costs $0.10 per cluster hour, and its GPU cluster quickstart covers Kubernetes or Slurm.

Data transfer is free. Crusoe's pricing FAQ says it does not charge for ingress or egress, inside a VPC or to the public internet. If you move large datasets and checkpoints every day, that can outweigh an hourly difference.

Energy can decide it too. Crusoe's locations page lists data centers in Iceland and Norway on renewable grids, and most of its US sites are matched with renewable energy certificates or power purchase agreements.

It also has an option for teams that do not want to touch GPUs at all. Crusoe's Serverless Fine-Tuning charges per training token, where tokens are dataset size times epochs. At $0.40 per million tokens for Llama 3.1 8B Instruct, three epochs over a 20-million-token dataset is 60 million tokens, or $24.00. If your model is on its supported list and a LoRA adapter is what you need, that is hard to beat for simplicity.

How do you move a small training job to Nodus?

Write checkpoints to /nodus/state and run your existing command through the Nodus CLI. Install it and sign in:

pip install nodus-compute nodus login 

Then dry-run the one-GPU job to see the estimate before anything is billed. Every flag is in the nodus run reference, and python train.py is your own command:

nodus run --name lora-h100 --gpu H100 --image nodus/pytorch \ --checkpoint /nodus/state --continuity Checkpointed --interruptible \ --max-cost 20 --dry-run -- python train.py 

Remove --dry-run to submit. The --max-cost 20 cap sits above the $15.60 list-price estimate for 6 H100 hours. If the run would spend more, it checkpoints and suspends instead of billing past the cap.

FAQ

Is Crusoe cheaper than Nodus for H100s?

Not at list price. Crusoe lists the H100 at $3.90 per GPU-hour on demand, rising to $4.50 on December 1, 2026, against $2.60 on Nodus when we checked. Crusoe's reserved and spot rates are negotiated with sales and are not on its pricing page.

Does Crusoe bill per second?

Yes. Its VM docs say on-demand VMs are charged per second while in the running state, including time spent running lifecycle scripts at startup or shutdown.

Does Crusoe have an MCP server for coding agents?

Yes. Crusoe's docs index lists a Crusoe Cloud MCP server that lets AI assistants query and inspect your infrastructure. The Nodus MCP server lets an agent read logs, estimate what a Job would cost and, after you confirm, create it.

Can I use Crusoe and Nodus together?

Yes, as two separate accounts. One workable split is a Crusoe reservation for one long, large run and Nodus for the smaller fine-tunes, evals and experiments around it. They bill separately, and Nodus does not run work in a Crusoe account.

Originally published on the Nodus blog.