Lambda

What Lambda actually is

Lambda rents NVIDIA GPUs by the minute, and that focus is the product. Where the hyperscalers sell hundreds of services and hide GPU pricing behind instance families, reservation math, and regional rate cards, Lambda sells machine-learning compute and very little else. The company has existed since 2012 — founded by brothers Stephen and Michael Balaban in San Jose, California — and spent its first years building deep-learning workstations and servers before moving the business to cloud GPU capacity around 2018. That history shows in what you get: a small, legible cloud API, images with NVIDIA drivers and ML frameworks preinstalled (Lambda Stack), SSH access from the first minute, and no proprietary abstraction between you and the hardware.

The offering splits into three tiers. On-demand Instances are single nodes with 1, 2, 4, or 8 GPUs, deployed self-serve in minutes. 1-Click Clusters are 16 to 2,000+ interconnected NVIDIA B200 or H100 GPUs, booked for two weeks to a year through sales. Superclusters are single-tenant private-cloud deployments for very large buyers. Almost everyone reading this review will live in the first tier, so that is what we judge here.

Each instance is a Linux VM pinned to a fixed GPU configuration — GPU count, vCPUs, RAM, and local SSD all move together, and you cannot reshape one dimension without changing instance type. Persistent filesystems attach separately and are billed per GB used per month, and they keep billing after you delete an instance until you delete the filesystem too. That detail matters more than it sounds: teams that churn short-lived training nodes and forget their storage get a bill that looks nothing like the GPU rate card.

Pricing: what you really pay

Lambda publishes per-GPU-hour rates and bills by the minute, which makes cost math unusually easy. Current on-demand rates, from the live listings GPU Finder tracks against Lambda's own API: a single RTX 6000 at $0.69 per hour is the cheapest instance, an RTX A6000 (14 vCPU, 100 GB RAM) is $1.09 per hour, an A10 is $1.29, and a 1x A100 40GB is $1.99 per hour. Moving up, a 1x H100 PCIe runs $3.29 per hour, a 2x H100 SXM node costs $8.38 per hour, and an 8x H100 SXM node is $31.92 per hour — that is $3.99 per GPU-hour at full node. The newest silicon carries the usual premium: a 1x B200 SXM6 with 180 GB of VRAM is $6.99 per hour.

Three things make this pricing genuinely different from AWS or Azure list prices. First, billing granularity is per minute, so a 20-minute debug job costs a third of the hourly rate instead of a full hour. Second, egress is free — moving checkpoints and datasets out costs nothing extra, which for training workloads is not a rounding error; hyperscaler egress charges routinely surprise teams exporting large artifacts. Third, the rate you see is the rate you pay: there is no support-plan tier that changes compute pricing, no data-transfer line items, and no upfront commitment required for on-demand.

The other side of the ledger: filesystems, reserved capacity, and anything involving sales. Cluster pricing is not fully self-serve — 1-Click Cluster rates and Supercluster contracts are quoted by the team, and reserved capacity is where Lambda's lowest per-GPU prices live. For budgeting a burst research workload, the published on-demand numbers are what you will pay; for a steady 1,000-GPU footprint, you are in a negotiation. Side-by-side GPU-hour comparisons across providers live in our benchmarks.

Performance & infrastructure

Lambda's public cloud spans 14 regions by its own documentation: Virginia, Washington DC, Illinois, three Texas regions, California, Arizona, and Utah in the US; Tokyo and Osaka in Japan; one region in India; one in Germany (europe-central-1); and one in Israel. Ten of the fourteen are in the United States, and Europe is represented by a single German region — a real constraint for EU-latency workloads or anyone needing data residency outside Germany. Not every instance type is available in every region; the flagship B200 and H100 SXM configurations are largely US-default, while older parts like the A100 and A10 show up in Asia and Europe. If region matters to you, check per-SKU stock before you plan around it.

On the compute side, Lambda's pitch is direct hardware access without a virtualization tax on the GPU, preinstalled CUDA and framework stacks, and for multi-node work, high-bandwidth interconnects between GPUs. Within a node, GPU-to-GPU traffic rides NVLink-class interconnects; the 1-Click Cluster tier is explicitly sold on interconnected B200 and H100 nodes for distributed training. We would not over-read vendor interconnect claims without numbers of our own, but the configuration choices — SXM parts, full-node allocations, and clusters sized from 16 GPUs up — are the right shape for large-model training.

Storage is the weakest pillar. The platform offers network filesystems (up to 24 per account, with a 10 TB size cap in the Texas us-south-1 region) plus a Filesystem S3 Adapter that exposes those filesystems through an S3-style API. That adapter is a compatibility layer, not a real object store: no storage classes, no lifecycle rules, no multi-region replication. If you want S3 semantics at scale, you will bring your own object storage and treat Lambda as pure compute.

Operationally, there is a public status page (status.lambda.ai) covering frontend, infrastructure, and API components, minute-by-minute instance observability in the console, firewalls and SSO in the access layer, and full CRUD over instances through the Cloud API for CI/CD automation. The documentation is written for practitioners — region tables, filesystem limits, and known image issues (a December 2025 apt upgrade failure on the 24.04 base images) are all published rather than buried.

Where it falls short

The largest gap is contractual: Lambda publishes no uptime SLA for on-demand instances. A community member asked for one in 2024 and got no published answer, and we could not find an uptime guarantee in the docs or product pages. When a node or GPU fails, your remedy is a replacement instance, not service credits. For research and training that is usually acceptable; for anything with a production SLO behind it, it is a real difference from the 99.99% contracts available elsewhere.

Stock volatility is the second issue. GPU Finder's 30-day availability sampling from Lambda's own API shows the 1x H100 PCIe out of stock 18% of the time, the 4x H100 SXM in stock only 4% of the time, and the 8x A100 80GB out 11%. Cheaper and older SKUs sit available; the configs everyone wants sell out. If your training schedule cannot absorb a "no capacity" day, you need reserved capacity or a second provider.

Third, the platform surface is thin. There is no managed database, no CDN, no load balancer product line, and no bundled backup or snapshot service — backups are whatever you script onto the filesystems or ship off-platform yourself. Lambda's answer is that this keeps costs down and lets you bring your own stack, which is fair, but it means Lambda is a component in an architecture, not a place to run one.

Finally, geography. Nothing in Latin America, nothing in Australia, one EU region. Teams with data-locality requirements outside Germany and the US will find the map limiting even when the prices are attractive.

Who should pick this

We'd pick Lambda when GPU cost per training-hour is the dominant line item and the workload can tolerate occasionally missing capacity: fine-tuning runs, research experiments, batch inference, evaluation sweeps, and startups that want H100-class hardware without a hyperscaler contract. The per-minute billing and free egress make it particularly good for bursty, artifact-heavy jobs where you train for an hour and pull down a 50 GB checkpoint.

We would not pick it for latency-sensitive production serving with a contractual SLA, for EU workloads that must stay outside Germany, or for teams that want managed databases, object storage, and backups to come from the same vendor. In those cases the hyperscalers or a full-stack European provider earn their premium. For most ML teams, though, the honest comparison is not Lambda versus AWS at list price — it is Lambda versus the other GPU specialists, and on published pricing clarity and egress policy, Lambda sets a standard the others are measured against. More GPU-cloud options are in our gpu-cloud category.

Key takeaways

• Transparent per-minute GPU pricing from $0.69/hr (RTX 6000) to $6.99/hr (1x B200), with no egress fees and no commitment required for on-demand
• No published uptime SLA — availability is best-effort with a status page, not a contract
• 14 regions, ten of them US; Europe is Germany only, and flagship GPU SKUs are mostly US-anchored
• Popular SKUs sell out: 30-day stock tracking shows H100 configurations unavailable 4-18% of the time
• Compute-focused platform: filesystems plus an S3 adapter, but no managed databases, snapshots, or backup products
• Our pick for training and fine-tuning teams optimizing cost per GPU-hour; not our pick for SLA-backed production serving

Sources

• https://docs.lambda.ai/public-cloud/on-demand
• https://gpufinder.dev/providers/lambda
• https://lambda.ai/pricing
• https://lambda.ai/instances
• https://lambdalabs.com/llms.txt
• https://aiwiki.ai/wiki/lambda_labs
• https://status.lambda.ai/
• https://deeptalk.lambda.ai/t/gpu-cloud-service-level-agreement/4288