Lambda Cloud

Lambda Cloud rents NVIDIA GPU instances by the hour with a preinstalled AI software stack and no egress fees. On 31 October 2026 one H100 SXM was $4.29 an hour, and an 8-GPU box was $3.99 per GPU. It is a plain, predictable GPU cloud for training and steady serving. It costs more per hour than Runpod Community and has no serverless billing.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

For side-by-side reading, see Runpod vs Lambda. To filter by price and GPU across all hosts, use the hosting compare.

What Lambda sells

On-demand instances. You pick a GPU type and count, a region, and optionally a filesystem. The instance boots with Lambda Stack and you log in over SSH, with JupyterLab available. The docs warn that it can take a few minutes after launch for the instance to become reachable.

1-Click Clusters. The pricing page offers clusters of 16 to 2,000 or more interconnected HGX B200 or H100 GPUs for large-scale work.

Reserved capacity. The page lists reserved terms, from two weeks to a year, for B200 and H100 systems, and says to contact the team for reserved capacity at the lowest prices.

Filesystems. Regional network storage you attach at instance creation, which persists after the instance is gone. Per the docs, the filesystem must be in the same region as the instance, it cannot be moved between regions, and in most regions each account can have up to 24 filesystems.

Lambda is a GPU cloud and does not run your code for you. You get a server and you run vLLM, Ollama, PyTorch or anything else on it. If you want to manage no server at all, look at Modal or a per-token API.

The price list

On-demand prices from lambda.ai/pricing, checked on 31 October 2026, in dollars per GPU per hour. The page lists several instance sizes, and the price per GPU falls as the count rises.

Blank cells mean the page did not list that size when we read it. The 8x H100 box lists 208 vCPUs, 1,800 GiB of RAM and 22 TiB of SSD. The single H100 SXM lists 26 vCPUs, 225 GiB of RAM and 2.75 TiB of SSD.

Reserved-term lines on the same page are higher per hour than the on-demand rows because they describe different products. For example, 16 H100 GPUs for two weeks to a year are listed at $6.16 per GPU-hour, with the price falling to $5.85 at 64 GPUs and $5.54 at 256, and HGX B200 at $9.86, $9.36 and $8.87 for 16, 64 and 256 GPUs. These are cluster prices that, we assume, include interconnect and support.

What you pay in practice

The volumes are examples we chose.

Example one: a weekend fine-tune on one H100. 20 hours × $4.29 = $85.80. On Runpod Community the same hours would be $53.80, and on Modal about $79 (20 × $3.95).

Example two: a week-long training run on 8 H100s. 168 hours × 8 GPUs × $3.99 = $5,362.56. Add filesystem storage, billed per GB used per month in one-hour steps.

Example three: a dedicated inference box for a team. One A100 SXM 40 GB at $1.99 for 220 hours a month is $437.80. For a 120B-class model you need 80 GB, so a single A100 40 GB will not hold it. You would pay for an 8x A100 80 GB box at $2.79 per GPU, which is $22.32 an hour for all eight, or use one H100 at $4.29. The H100 at 220 hours is $943.80, and always-on for 720 hours is $3,088.80.

Example four: a 24 GB experiment. An A10 at $1.29 for 10 hours is $12.90.

The lesson is the same as with other GPU clouds: price per hour is only half of it. Idle hours cost full price. For bursty use, a serverless host is cheaper. For steady use, a GPU cloud is the better fit. At 720 hours a month, Lambda's single H100 costs $3,088.80, and Runpod's Community H100 SXM costs $1,936.80 in the same month, though Community capacity is a different product class.

Limits and gotchas

• First-come capacity. Lambda describes on-demand as self-serve with first-come access. At busy times the GPU you want may not be available in your region. Reserved capacity solves that for a price.
• Region lock for storage. A filesystem lives in one region and cannot be moved. Choose the region before you load a large dataset.
• No scale to zero. An instance bills until you terminate it. Forgetting one for a month on an H100 is a $3,088.80 mistake.
• A few minutes to boot. The docs say it can take a few minutes for an instance to become accessible after launch.
• Account limits. Per the docs, accounts can have up to 24 filesystems in most regions, and the Texas region differs.
• No serverless, no per-token billing. You manage the server and the model server on it.

Data and security

Lambda's pricing page and docs we read describe the product, not a data-retention policy. A rented GPU instance is yours: you control what runs, what is stored on the SSD and the filesystem, and what is sent over the network. That gives you strong control, since Lambda is not processing your prompts, you are. It does not tell you how Lambda handles data at rest on its hardware or what audits it holds.. If you handle regulated data, ask before you load it.

Practical steps: encrypt sensitive data at the application level, use a filesystem you delete when done, firewall the instance so only the ports you need are open, and keep model servers behind a proxy with authentication. Both vLLM and Ollama need that, as described on vLLM and Ollama.

Choosing an instance size

The price list rewards buying more GPUs per box, but only slightly: an H100 is $4.29 in a single-GPU instance and $3.99 per GPU in an 8-GPU box, a drop of about 7 percent. So do not buy eight GPUs to save on the rate. Buy the count your job needs.

For GH200, the software ecosystem is different from x86, since it is an Arm host. Check that your container images support it.

Multi-GPU boxes help in two ways. They fit models that do not fit on one card, and they allow tensor parallelism with vLLM, which the vLLM docs describe with a tensor_parallel_size setting equal to the number of GPUs in the node. They do not make a small model faster by themselves.

A training run, step by step

For a multi-day job, a few habits keep a rented GPU honest.

1. Stage data on a filesystem first. Attach it in the region you will train in. Loading a dataset on a billed H100 wastes money.
2. Test the script on a cheap GPU. An A10 at $1.29 is enough to find syntax and path errors.
3. Write checkpoints to the filesystem. If the instance dies, you resume from the last checkpoint, not from scratch.
4. Run inside tmux or a job manager so an SSH drop does not kill the run.
5. Log cost. Write the start time and rate to a note. A 168-hour run on eight H100s is $5,362.56, and a forgotten instance after the run adds $31.92 an hour for the box.
6. Terminate on completion. Use a script that shuts the instance down when the job exits.
7. Copy results out. There are no egress fees listed, so pulling checkpoints costs only time.

Lambda for inference

Serving works the same as on any GPU cloud. You install vLLM or Ollama, load a model, put authentication in front and monitor it. The questions are fit and cost.

For fit: take the listed size of your model, add a quarter for cache, and pick the card. gpt-oss 20B is listed at 14 GB, Qwen3-Coder 30B at 19 GB and gpt-oss 120B at 65 GB. For cost: compare with per-token pricing before you commit. A single H100 at $4.29 for 720 hours is $3,088.80 a month. At Groq's listed gpt-oss-120b rates of $0.15 and $0.60 per million tokens, $3,088.80 buys about 5 billion output tokens, which is far more than a ten-person team is likely to send in a month. That comparison ignores quality, latency and privacy, which are the real reasons to rent the GPU. Be honest about which of those is yours.

If privacy is the reason, a dedicated instance gives you full control of what is logged and where the data lives. Our private AI for teams page walks through the policy side, and on-premise AI covers when owning hardware makes sense.

Lambda against its nearest rivals

For the long reads, see Runpod vs Lambda and Modal vs Runpod.

Who should pick Lambda

Alternatives

• Runpod: lower hourly rates on Community Cloud, per-second billing and serverless.
• Modal: per-second billing in a function model.
• Replicate: models by the second without any server.
• Hugging Face Inference: dedicated endpoints at A100 80 GB $2.50 an hour on AWS.
• AWS Bedrock: managed models inside AWS.
• The AI hosting overview maps all of them, and Runpod vs Vast.ai covers the marketplace route.

Setting up a first instance

1. Create an account and add a payment method.
2. Create a filesystem in the region you plan to use, if you want data to persist.
3. Launch an instance with the GPU you need, attach the filesystem and add your SSH key.
4. Wait a few minutes, then SSH in. Lambda Stack gives you the NVIDIA driver, CUDA toolkit, cuDNN, NCCL, PyTorch, TensorFlow, JAX and Docker.
5. Start your server, for example vLLM or Ollama, behind a firewall rule that exposes only what you must.
6. Terminate the instance when done. The filesystem stays and bills per GB used.

For sizing, our GPU requirements for LLMs page works from model sizes to cards.

Our take

Lambda is the dull, dependable option, and that is a compliment for training runs. The no-egress-fee line and 8-GPU boxes at $3.99 per H100 are the strong points. The weak points are first-come capacity and the lack of scale-to-zero. We would pick it for multi-day training or a dedicated inference box that stays busy, and send bursty work elsewhere. Check Runpod's tiers before you commit, since Community Cloud can be cheaper for work that is not sensitive. These notes come from Lambda's pricing page and docs, and we ran no tests.

Related pages

• Hosting hub and AI hosting overview
• Runpod vs Lambda
• Private AI hub and how to self-host an LLM