Lambda

What is Lambda?

Lambda is a GPU cloud and AI infrastructure company that rents NVIDIA GPUs, from single instances to clusters of thousands, for training and running AI models. It operates at lambda.ai, and the older lambdalabs.com address now redirects there.

You can launch on-demand instances by the hour, book 1-Click Clusters for weeks or months, or contract for private superclusters built on NVIDIA GB300 NVL72 systems. Default instances ship with Lambda Stack, a bundle of drivers and machine learning frameworks, so training can start soon after launch.

Lambda says it was founded in 2012 in San Francisco by machine learning engineers, with its first office in the Noisebridge hackerspace. It now describes itself as an AI-only company building AI factories at gigawatt scale, and it holds SOC 2 Type II certification.

This Lambda is not AWS Lambda. Amazon’s AWS Lambda is a serverless service for running short functions without managing servers, and it has nothing to do with GPU rental. Searches often mix the two, so check the domain before reading any review.

Key Features

Hourly access to current NVIDIA GPUs with a ready ML stack is Lambda’s core feature, backed by clusters that scale beyond 2,000 GPUs.

• On-demand GPU instances: You can rent NVIDIA B200, H100, A100 and older GPUs by the hour from the web console or the Cloud API. Instances come in 1, 2, 4 or 8 GPU configurations, so small experiments and full nodes use the same workflow.
• 1-Click Clusters: Lambda provisions HGX B200 and H100 clusters from 16 to more than 2,000 GPUs for terms of two weeks to a year. This gives teams multi-node training capacity without negotiating a long data-centre contract.
• Superclusters: Single-tenant NVIDIA GB300 NVL72 systems linked with Quantum-2 InfiniBand are offered for frontier-scale work. The homepage also lists VR200 NVL72 and HGX B300, which suits labs planning very large training runs.
• Lambda Stack: Every default instance runs Ubuntu 22.04 with the NVIDIA driver, CUDA, cuDNN and NCCL installed. PyTorch, TensorFlow and JAX are ready too, so you can start training within minutes of launching.
• Persistent filesystems: Networked filesystems keep datasets and checkpoints after an instance is terminated. They are billed per GB used each month in hourly increments, and each account can hold up to 24 filesystems.
• Built-in JupyterLab and SSH: Each instance opens in JupyterLab from the browser or over SSH on port 22. Researchers can move between notebooks and terminal work without configuring remote access tools first.

Pros

Zero egress fees give Lambda a practical edge for data-heavy training, alongside early access to new NVIDIA hardware and a preinstalled ML environment.

• No egress fees: Lambda’s pricing page states that it does not charge for data leaving its cloud. Teams moving large checkpoints or datasets avoid a cost that often surprises users of the big hyperscalers.
• Current NVIDIA hardware: The homepage lists GB300 NVL72, B300, B200, H200 and H100 systems. Teams that need the newest accelerators for training can often access them here before broader availability elsewhere.
• Ready-made ML environment: Lambda Stack ships drivers, CUDA and the main frameworks on every default instance. Engineers spend less time fixing driver mismatches and more time running experiments.
• Scales from one GPU upward: The same account covers a single on-demand GPU, a 1-Click Cluster and a private supercluster. A start-up can grow from prototyping to multi-node training without changing provider.
• Focused on AI only: Lambda describes itself as an AI-only company whose engineering and support serve GPU workloads. Support staff are therefore familiar with distributed training problems rather than general cloud tickets.

Cons

Cost for small users is Lambda’s weakness: there is no free tier, and the H100 on-demand rate is $3.99 per GPU-hour.

• No free tier: Lambda offers no free plan or trial credits on its pricing page, and every instance is billed. Students and hobbyists should look at notebook services with free GPU quotas before committing.
• Higher list prices than some rivals: The H100 on-demand rate of $3.99 per GPU-hour is above several GPU marketplaces. The lowest rates require reserved capacity, which means talking to sales and committing for longer.
• Few managed services: Lambda focuses on raw compute and does not offer the managed databases, queues and serverless tools of AWS or Google Cloud. You must run orchestration and serving yourself or add third-party tools.
• Capacity can be tight: Popular GPUs can sell out in busy regions, a common complaint across GPU clouds. Teams with deadlines should reserve capacity or build fallback options with another provider.
• Name confusion with AWS Lambda: Searches for Lambda often return Amazon’s serverless functions service, which is unrelated. New users may also find older material under the former lambdalabs.com address, which now redirects to lambda.ai.

Who is Using Lambda?

AI labs and machine learning teams that train large models are Lambda’s core customers, from start-ups renting one GPU to labs booking superclusters.

• AI research labs: Frontier and academic labs rent large clusters for pre-training and fine-tuning. Lambda’s about page says hyperscalers, enterprises and frontier labs rely on its compute.
• Machine learning start-ups: Small teams rent on-demand GPUs to prototype models. They move to 1-Click Clusters once training runs need many nodes.
• Enterprise AI teams: Large companies use single-tenant clusters for sensitive training data. SOC 2 Type II controls and hardware isolation support their security reviews.
• Individual ML engineers: Engineers rent a single GPU for experiments, fine-tuning or benchmarks. JupyterLab access means they can start work from a browser in minutes.

Pricing

Lambda charges per GPU-hour with no free tier, from $0.79 for a V100 to $6.69 for a B200 on demand (official pricing page, October 2026).

Some third-party blogs quote much lower Lambda rates, such as $2.50 for an H100, which do not match the official page; rely on lambda.ai for current figures. Prices exclude sales tax, VAT or GST.

Disclaimer: GPU cloud rates move often and depend on region, term and capacity, so confirm the live price in the console before launching a long job.

What Makes Lambda Unique?

One provider covers the full range from a single hourly GPU to single-tenant GB300 NVL72 superclusters, and every tier is built only for AI. Hyperscalers offer GPUs as one product among hundreds, while Lambda’s console, documentation and support are all shaped around training and inference work.

The pricing model is also simple. Per-GPU-hour rates are published, there are no egress fees, and Lambda Stack removes most driver and framework setup. The trade-off is fewer managed services, so you bring your own orchestration and serving tools.

Compatibilities and Integrations

Lambda works through a web console and a Cloud API, with Linux instances that run standard PyTorch, TensorFlow and JAX code without changes.

• Access methods: Instances launch from the Lambda Cloud console or the Cloud API. SSH and browser-based JupyterLab are both available on every default image.
• Software stack: Lambda Stack bundles the NVIDIA driver, CUDA, cuDNN, NCCL, PyTorch, TensorFlow and JAX on Ubuntu 22.04. Lambda Stack can also be installed on your own Linux machines.
• Regions: The documentation lists 14 regions across the US, Germany, Japan, India and Israel. Filesystems in the Texas region are capped at 10 TB.
• Storage and data: Persistent networked filesystems hold datasets and checkpoints between sessions. Data can leave the platform without egress charges, according to the pricing page.
• Third-party tools: We found no official Terraform or SkyPilot integration in the documentation we read. Teams usually script the Cloud API or bring their own orchestration such as Kubernetes or Slurm.

Lambda Tutorials

A first training job on Lambda needs an account, an SSH key and an instance, and takes about 15 minutes to set up.

1. Create an account at lambda.ai and add a payment method.
2. Add an SSH key in the Lambda Cloud console.
3. Optionally create a persistent filesystem in your chosen region for datasets and checkpoints.
4. Launch an instance, choosing the GPU type, GPU count and region.
5. Connect over SSH or open JupyterLab, then run your training script with the preinstalled frameworks.
6. Terminate the instance when the job ends so billing stops, keeping data on the filesystem.

How We Rated It

Lambda scores 4.1 out of 5, earned by fast current hardware and no egress fees, with marks lost for limited managed services.

Summary

Lambda suits AI teams that need serious NVIDIA compute, from one GPU for a fine-tune to thousands for pre-training. Published per-GPU-hour pricing, no egress fees and a preinstalled ML stack make it easy to start. The main limitation is cost for small users, with no free tier and few managed services beyond compute. Remember that it is unrelated to Amazon’s AWS Lambda serverless product. Store data on a persistent filesystem and terminate idle instances promptly, since billing runs until you stop them.