Lambda
GPU cloud from a company that also builds AI workstations, offering on-demand NVIDIA instances and reserved multi-node clusters with InfiniBand.
Pricing Plans
| Plan | Price/mo | CPU | RAM | Storage | Notes |
|---|---|---|---|---|---|
| A10 24GB | — | 30 vCPU | 200GB | 1.4TB NVMe | Around $0.75/hr |
| A100 40GB | — | 30 vCPU | 200GB | 512GB NVMe | Around $1.29/hr |
| 8x H100 SXM | — | 208 vCPU | 1.8TB | 20TB NVMe | Around $2.49/hr per GPU |
· Updated
About Lambda
Lambda started by building GPU workstations and servers for deep learning researchers, then turned that hardware expertise into a cloud. That history shows in the product: the machines are configured the way someone who actually trains models would want them, and the software stack is opinionated in a useful way.
Lambda Stack is their Ubuntu package repository with CUDA, cuDNN, PyTorch, and TensorFlow versions that are known to work together. Anyone who has spent an afternoon reconciling a CUDA version against a PyTorch build will appreciate what that is worth. You can install it on your own Ubuntu machine too, independent of using their cloud.
Pricing
On-demand instances are billed per hour. An A10 sits around $0.75/hr, an A100 40GB around $1.29/hr, and 8x H100 nodes around $2.49/hr per GPU, which is aggressive against AWS and GCP for the same silicon. Reserved capacity for multi-week or multi-month training runs is negotiated and cheaper still.
The catch is availability rather than price. Popular configurations, H100 nodes especially, are frequently sold out, and there is no spot market to fall back on. Plan around capacity, not just cost.
Service Details
Instances run Ubuntu LTS with Lambda Stack preinstalled, accessed by SSH or a hosted Jupyter interface. Persistent filesystems attach across instances in a region, so a dataset survives after you tear down the GPU.
For multi-node training, the 1-Click Clusters give you H100 nodes wired together with NVIDIA Quantum-2 InfiniBand, which is the interconnect that makes distributed training scale rather than stall on gradient synchronisation. That is the part general-purpose clouds usually charge a lot more for.
There is no Windows option and no attempt at one. This is a Linux shop.
Available in: United States, Europe, Asia
Learn More
For more information about Lambda, visit their website at https://lambda.ai/