If you use Databricks mainly to run LLM training and fine-tuning jobs, the cheapest alternative at list price is Nodus: Databricks AI Runtime lists the NVIDIA H100 at $7.00 per GPU-hour, Nodus lists it at $2.60, so one day on 8 x H100 costs $1,344.00 against $499.20. Modal, RunPod, Lambda and Amazon SageMaker AI also price that day below Databricks, and none of them requires a Databricks workspace. Stay on Databricks when your training data is governed in Unity Catalog, or when MLflow tracking and Model Serving are why you are there.
We build Nodus, so weigh our view accordingly. Every Databricks and competitor fact below links to the provider's own page, checked November 22, 2026, and the arithmetic is written out so you can redo it with your own numbers.
What does Databricks charge for LLM training now?
Databricks sells GPU training as AI Runtime, its serverless GPU compute. The AI Runtime overview describes it as serverless GPU compute for training, fine-tuning and batch inference on A10 and H100 GPUs, with 8 x B300 in Public Preview behind a waitlist. You submit work from the Databricks CLI, a notebook or a scheduled job.
The AI Runtime pricing page lists A10 On Demand at $2.50 per GPU-hour and H100 On Demand at $7.00 per GPU-hour, both marked Preview, in its default AWS view. The same page says the charge appears on your bill as DBUs on the Model Training SKU, based on GPU hours and GPU type. So "serverless GPU on Databricks" is still DBU billing, just with a per-GPU-hour list price.
The older route is gone. Databricks' Foundation Model Fine-tuning docs say that product has reached end of life, the databricks_genai package is no longer available, and AI Runtime is the recommended replacement.
Three limits from the overview matter for planning. CLI jobs can run for up to 14 days and notebook jobs for up to two days, after which Databricks asks you to checkpoint and restart. Only 8 x H100 and 8 x B300 support multi-GPU training. Capacity may be constrained in your region, and AI Runtime sometimes uses GPUs in another region, which can add egress costs.
How do Databricks and Nodus compare at a glance?
| Databricks AI Runtime | Nodus | |
|---|---|---|
| How you pay | DBUs on the Model Training SKU, inside a Databricks workspace | Prepaid credits by default; monthly usage billing for approved organizations |
| H100 list price | $7.00 per GPU-hour | $2.60 per GPU-hour |
| GPU choices | A10, H100, 8 x B300 (preview, waitlist) | 16 listed types, from an 8 GB RTX 3070 to the 288 GB B300 |
| Multi-GPU and multi-node | 8 GPUs on one node; multi-node H100 is a preview | Multi-node is Beta, 2 to 8 machines |
| When a run is interrupted | You checkpoint to Unity Catalog volumes and restart | Nodus restores /nodus/state into the next attempt |
| Where data lives | Unity Catalog volumes and tables | Nodus Volumes, or S3 through a Connection |
| Agent access | Genie Code and a databricks-ai-runtime agent skill | MCP server for Claude Code, Cursor and Codex |
Sources, checked November 22, 2026: AI Runtime overview, AI Runtime pricing, AI Runtime docs home, Nodus pricing, Nodus billing, multi-node training, checkpoints, connections and MCP.
What does the same training run cost on Databricks and its alternatives?
We priced one job everywhere: 8 x H100 for 24 hours, which is 192 GPU-hours, at each provider's listed on-demand rate. This is GPU compute only. Storage, egress and any platform fees outside the GPU rate are excluded on every row.
| Platform | Listed rate | Per H100-hour | 8 x H100 for 24 hours |
|---|---|---|---|
| Nodus | $2.60 per GPU-hour | $2.60 | $499.20 |
| Modal | $0.001097 per GPU-second | $3.95 | $758.25 |
| RunPod (Secure Cloud) | $4.39 per GPU-hour, H100 SXM | $4.39 | $842.88 |
| Lambda | $4.59 per GPU-hour, 8 x HGX H100 instance | $4.59 | $881.28 |
| Databricks AI Runtime | $7.00 per GPU-hour | $7.00 | $1,344.00 |
| Amazon SageMaker AI | $63.296 per hour, ml.p5.48xlarge training | $7.91 | $1,519.10 |
Prices checked November 22, 2026: Nodus pricing, Modal pricing, RunPod pricing, Lambda pricing, Databricks AI Runtime pricing and the AWS SageMaker price list (US East, N. Virginia). Modal's per-second price is multiplied by 3,600. The SageMaker instance price is divided by 8 GPUs.
The Databricks row is 192 x $7.00 = $1,344.00. The Nodus row is 192 x $2.60 = $499.20, which is $844.80 or 63% less. Modal is 192 x $3.9492 = $758.25, RunPod Secure Cloud is 192 x $4.39 = $842.88, Lambda is 192 x $4.59 = $881.28, and SageMaker is 24 x $63.296 = $1,519.10.

At list prices, the same 192 H100-hours cost $844.80 less on Nodus than on Databricks AI Runtime.
How much faster would Databricks need to be to cost the same?
Divide the two hourly prices, and that is the speedup Databricks has to deliver on the same GPU. Databricks says that AI Runtime has optimized data loading and asynchronous checkpointing that cut idle GPU time (why AI Runtime). That is a fair claim to test, and the test is simple.
Against Nodus, the ratio is $7.00 / $2.60 = 2.69. Your job would have to finish in 37% of the GPU-hours on Databricks to cost the same. Against Modal, it is $7.00 / $3.9492 = 1.77, so the job would need to finish in 56% of the time.
Our decision rule is to run one hour of the real job on both, with the same model, batch size and sequence length, and compare tokens per second. If Databricks is less than 2.69 times faster on the same H100, Nodus is the cheaper place for that job. We have not run this comparison ourselves on AI Runtime, so treat the ratio as the bar, not a result.
What does a run that lasts for days cost?
The gap scales with time, because both are flat hourly prices. At 8 x H100, Databricks AI Runtime costs $1,344.00 a day, Modal $758.25 and Nodus $499.20. After 7 days that is $9,408.00, $5,307.72 and $3,494.40. At the 14-day CLI limit it is $18,816.00, $10,615.45 and $6,988.80.

A two-week 8 x H100 run on Databricks costs $11,827.20 more than the same GPU-hours on Nodus.
Long runs also raise the question of interruption. The Databricks resiliency guide has you save PyTorch Distributed Checkpoint shards to a Unity Catalog volume, find the newest complete checkpoint on restart, and record your data loader position yourself.
Nodus does the copying and restoring for you. You write model, optimizer and step to /nodus/state, Nodus checkpoints that directory on a schedule and when capacity gives notice, and the next attempt starts with it restored, per the checkpoints guide. Your script still has to load what it finds. We walk through that pattern in How do you make spot GPU training survive preemption?.
Is there a cheaper Databricks option for small fine-tunes?
The A10 is the smaller Databricks option, at $2.50 per GPU-hour. It is a 24 GB card, which fits a LoRA or QLoRA fine-tune of a small model. Other 24 GB cards cost much less per hour.
Modal lists the A10 at $0.000306 per second, which is $1.10 per hour (Modal pricing). Nodus lists the 24 GB NVIDIA L4 at $0.43 per hour and the 24 GB RTX 4090 at $0.29 per hour on the pricing page. For a 10-hour fine-tune, that is $25.00 on Databricks, $11.02 on Modal, $4.30 on a Nodus L4 and $2.90 on a Nodus RTX 4090.

For small fine-tunes, a 24 GB card on Nodus costs a fifth or less of the Databricks A10 hour.
These are different GPUs, so the same break-even rule applies. The L4 and RTX 4090 do not run at the same speed as an A10, and your job may take longer or shorter on each. Measure one short run before you commit a long one.
How do you get lakehouse data into a training job outside Databricks?
Export the training set as files, then point the job at them. On Databricks, AI Runtime streams data from Unity Catalog volumes and tables. Outside it, the simplest bridge is object storage that both sides can read.
On Nodus, an S3 Connection gives a Job read access to a bucket through an IAM role that Nodus assumes for minutes at a time, with an external ID unique to that Connection (connections guide). The Job then lists the bucket prefix as an input. For smaller datasets, you can upload to a Nodus Volume instead.
The honest cost is governance. Once the files leave Unity Catalog, its access controls and lineage stop following them. If that is a problem for your data, it is a reason to stay, which is the next question.
Where is Databricks the better choice?
Databricks wins when the training job is one step in a platform you already run.
Your data is governed in Unity Catalog. AI Runtime reads Unity Catalog data through Spark Connect without running a Spark cluster, according to why AI Runtime. Copying regulated data out to train elsewhere means rebuilding those controls.
You want tracking, registry and serving in one place. The AI Runtime docs cover MLflow experiment tracking, scheduling with Declarative Automation Bundles, and registering and deploying trained models to Model Serving endpoints. That is a lot of glue you would otherwise write.
Your security team already approved Databricks. The overview says AI Runtime works with the compliance security profile, except for the FedRAMP High and DoD IL5 standards. A new vendor review can take longer than the training run you are trying to save money on.
You need very large GPU memory on one node. The 8 x B300 option gives 288 GB per GPU, which Databricks positions for the largest models and longest context lengths, though it is a waitlisted preview.
You have committed Databricks spend. AI Runtime bills as DBUs, so it lands on the Databricks bill you already have.
How do you move a training job from Databricks AI Runtime to Nodus?
The job definition maps almost one to one. This is the minimal Python workload from the Databricks AI Runtime quickstart:
experiment_name: python-quickstart environment: version: databricks_ai_v6 compute: num_accelerators: 1 accelerator_type: GPU_1xA10 code_source: root_path: ./src command: python "$CODE_SOURCE_PATH/main.py" On Nodus you install the CLI with pip install nodus-compute, sign in with nodus login, and run the same script from its directory. This command is copied from the Nodus quickstart:
nodus run --name cli-quickstart --gpu L4 --image nodus/pytorch \ --max-cost 1 --timeout 10m --no-warm \ --output metrics=/nodus/outputs/metrics.json -- python train.py --gpu replaces accelerator_type, --image nodus/pytorch replaces the Databricks AI environment, and --max-cost puts a dollar cap on the Job. Add --dry-run first to see the estimate without starting paid compute. For a run that should survive interruption, nodus run --checkpoint /nodus/state makes that directory the checkpoint path.
Change two things in the script itself. Write checkpoints to /nodus/state instead of /Volumes/... paths, and write results to /nodus/outputs. If you are leaving SageMaker or Vertex AI too, SageMaker alternatives and Vertex AI alternatives cover those migrations, and Modal alternatives for training compares the serverless options in more depth.
FAQ
Is Databricks Mosaic AI Training still available for fine-tuning?
The Foundation Model Fine-tuning API, which grew out of the Mosaic AI training product, has reached end of life. Databricks says the databricks_genai package and its fine-tuning UI are no longer available, and points users to AI Runtime instead.
Can I keep my experiment tracking if I train on Nodus?
Yes, if you track with Weights & Biases. A Nodus WandB Connection sets the run variables for the Job, and a run that Nodus restarts after lost capacity resumes the same W&B run rather than starting a new one (connections guide).
Can I spend my AWS, GCP or Azure credits instead of buying new compute?
On Nodus, yes, for single-node work. You can connect your own AWS account, GCP project or Azure subscription and let Nodus place Jobs there first, within a spending limit you set, so your cloud credits are used before Nodus capacity. This is Beta and is described in Connect your cloud accounts.
Does Databricks charge for data transfer on AI Runtime?
It can. The pricing page says data transfer costs may apply when data moves between regions, zones or out of a cloud. The overview adds that AI Runtime sometimes uses GPUs in another region during high demand, which may incur egress costs.
What is the cheapest way to try an alternative without a contract?
Run one short job with a hard cap. On Nodus, the quickstart's --max-cost 1 --timeout 10m limits a first run to one dollar and ten minutes, and creating an organization requires email and card verification. Prices for every GPU are on the Nodus pricing page.
Originally published on the Nodus blog.

