LAMBDA LABS · TECH
Lambda Labs: dedicated GPU cloud in the USA, A100-40GB from USD 1.10/h, H100 from USD 2.49/h
Lambda Labs is a US GPU cloud with ML engineer focus. A100-40GB from USD 1.10/h, H100 from USD 2.49/h, reserved contracts with 1-year/3-year rebates.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is Lambda Labs?
Lambda Labs (officially Lambda, Inc.) is a US GPU cloud platform headquartered in San Francisco, founded in 2012 originally as a GPU workstation manufacturer. As of May 2026 Lambda Labs operates dedicated GPU data centres in the USA (several sites on the West Coast and in the Midwest) and is one of the established providers in the ML engineer market. Unlike Vast.ai (auction marketplace) or RunPod (mix of Secure and Community Cloud), Lambda Labs operates exclusively its own hardware in its own data centres.
The product portfolio is tailored to ML workloads. On-demand GPU cloud with hourly billing. Reserved Cloud with 1-year or 3-year contracts and significant rebates (typically 30-50% reduction over on-demand). Lambda Stack as a pre-installed ML environment (CUDA, cuDNN, PyTorch, TensorFlow, JAX, all in matching versions). Lambda Workstations (physical hardware for the office) as a hardware sale. And since 2023 Lambda Inference as a serverless API for open-weight models.
May 2026 inventory: NVIDIA A10 (24 GB) from USD 0.60/h, A100-40GB from USD 1.10/h, A100-80GB SXM from USD 1.79/h, H100-80GB PCIe from USD 2.49/h, H100-80GB SXM from USD 2.99/h, H200-141GB from USD 3.99/h, B200 (Blackwell) from USD 4.49-6.99/h depending on configuration. Multi-GPU instances up to 8x H100 SXM in one node with full InfiniBand topology are available. Reserved contracts for 8x H100 SXM over 1 year from around USD 18,000-23,000/month, over 3 years even cheaper.
The platform is built for ML engineering workflows. Pre-built clusters with pretrained models, NVIDIA stack pre-installed, Slurm scheduler options for multi-node training, Jupyter notebooks startable via web UI, SSH access standard. CLI with `lambda-cli`, Python SDK, REST API. Custom image support for own Docker or container configurations.
Why it matters
Three points make Lambda Labs the standard choice in the ML engineer market: ML-focused platform maturity, clear reserved contracts for sustained workloads, and larger multi-GPU clusters with full InfiniBand topology.
ML engineer maturity: the platform has been established in the ML market for over 10 years. Lambda Stack as a pre-installed environment solves the most common pain point in ML setups: compatible version combinations of CUDA, cuDNN and framework. Anyone starting a pod has a working environment immediately – no 4-hour ordeal with version conflicts. That is a significant productivity advantage in ML daily life.
Reserved contracts for sustained workloads: anyone running a productive inference setup with 24/7 load benefits massively from Lambda Reserved Cloud. 1-year reserved A100-80GB from USD 1.39/h (versus USD 1.79/h on-demand), 3-year reserved from USD 1.05/h. Calculation example: 8x A100-80GB over 12 months on-demand: USD 1.79 × 8 × 720 × 12 = USD 123,500. With 1-year reserved: USD 95,800 – saving 22%. With 3-year reserved: USD 72,500 – saving 41% versus on-demand 12 months.
Multi-GPU clusters: Lambda Labs offers in May 2026 genuine 8x H100 SXM nodes with full InfiniBand topology, which is central for multi-node training (32+ GPUs). CoreWeave is the direct competitor here, Lambda Labs often slightly cheaper per hour, with a clear reserved path. For a training set with a 70B parameter model Lambda Labs is one of the economically most attractive options.
Lambda Inference: since 2023 Lambda Labs has operated a serverless inference API for open-weight models (Llama 3.3 70B, Llama 3.3 405B, Qwen, DeepSeek). Pay-per-token, OpenAI schema compatible, with high throughput guarantees. This variant competes with Together AI and Fireworks AI, with advantages on cost-efficient tariffs for some models.
Regulatory position: Lambda Labs is a US Delaware corporation, subject to US law and the CLOUD Act. A US subpoena can directly compel Lambda Labs to surrender data. As of May 2026 Lambda Labs has no EU or CH region. For workloads with CH or EU data residency requirements, Lambda Labs is not directly usable. For training data without personal reference (open-weight model finetune on public data), the constellation is pragmatic, but the TIA remains heavy due to US location.
How it works
Ordering: through the portal lambdalabs.com. Account creation by email verification with company details, payment by credit card or bank transfer (USD), prepaid model or reserved contract. Instance provisioning typically in 1-3 minutes for on-demand pods, immediate for reserved clusters.
Sample prices May 2026 (on-demand): NVIDIA A10 24GB: USD 0.60/h. A100-40GB: USD 1.10/h. A100-80GB SXM: USD 1.79/h. 8x A100-80GB SXM cluster: USD 14.32/h. H100-80GB PCIe: USD 2.49/h. H100-80GB SXM: USD 2.99/h. 8x H100-80GB SXM cluster: USD 23.92/h. H200-141GB: USD 3.99/h. B200: USD 4.49-6.99/h.
Reserved prices (1-year): A100-80GB SXM from USD 1.39/h, H100-80GB SXM from USD 2.49/h. Reserved (3-year): A100-80GB SXM from USD 1.05/h, H100-80GB SXM from USD 1.89/h. Multi-GPU clusters with discount on larger orders.
Lambda Inference: Llama 3.3 70B Instruct from USD 0.20/million input tokens, USD 0.30/million output tokens. Llama 3.3 405B Instruct from USD 0.80/million input, USD 0.80/million output. Qwen 2.5 72B similar. DeepSeek-V3 from USD 0.25/million input, USD 0.40/million output. OpenAI schema compatible, simple endpoint swap possible.
Network: each pod has a public IPv4 and IPv6. SSH access by public key. Private networks between pods in the same account available. InfiniBand topology on multi-GPU nodes with full bandwidth between GPUs.
Storage: local NVMe SSD on the pod for fast I/O, plus persistent volumes (file storage) that survive pod restarts. For large datasets S3-compatible object storage is typically connected via AWS S3 or Cloudflare R2, since Lambda Labs has no own S3 service in May 2026.
Contract details: on-demand pay-as-you-go. Reserved contracts with 1-year or 3-year minimum term. DPA per GDPR Art. 28 available on request for EU/CH customers but, given US location and CLOUD Act applicability, not fully GDPR-compliant without additional measures.
Migration: when switching from Vast.ai or RunPod to Lambda Labs, transfer the Docker image and training scripts, upload data via S3 sync, start the pod. Typically completed in half a day.
Lambda Labs setup for multi-node training in 5 steps
- 01Create an account at lambdalabs.com, register company details, configure credit card or bank transfer as payment method, request a DPA for EU/CH customers.
- 02Decide cluster sizing: 8x H100 SXM for 70B model pretraining, or 32x H100 SXM across 4 nodes with InfiniBand for faster throughput.
- 03Weigh reserved contract vs. on-demand: 1-year reserved for predictable sustained workloads (saving 20-30%), 3-year for production (saving 40-50%).
- 04Choose a Lambda Stack image with matching CUDA/PyTorch versions, start the pod, upload data via S3 sync from AWS S3 or Cloudflare R2.
- 05Start training with the Slurm or Ray scheduler in parallel across all nodes, configure checkpoint logic every 30 minutes to object storage, export data at the end and document the TIA entry.
When to use Lambda Labs
Lambda Labs is the right choice when (a) multi-GPU clusters with InfiniBand for serious training are needed, (b) reserved contracts for 24/7 sustained inference become economical, or (c) open-weight model inference via Lambda Inference is simpler than self-hosting. Concrete cases: AI lab needing a 32x H100 cluster for 70B model pretraining. SaaS company with sustained inference workload on a 1-year reserved contract. ML team using Lambda Inference as a cost-efficient open-weight API.
For 8x H100 SXM nodes with full InfiniBand topology, Lambda Labs is one of the economically most attractive options. CoreWeave is competitive but often priced higher. For multi-node training (32+ GPUs) both providers should be seriously evaluated.
Lambda Inference for open-weight models is in May 2026 an attractive price-performance point. Llama 3.3 70B from USD 0.20/M input tokens is cheaper than Together AI or Fireworks AI for the same model. For a Swiss SME with volume, Scaleway Generative API with EU residency remains the clean choice.
For reproducible ML experiments Lambda Stack is one of the best pre-installed stacks on the market. CUDA, cuDNN, PyTorch, TensorFlow, JAX in matching versions – no version-mismatch chase. That is a marked advantage in research and in productive ML pipelines.
When not to use
Lambda Labs is not appropriate for workloads with Swiss or EU data residency requirements. As of May 2026 Lambda Labs has no EU or CH region – all locations are in the USA. A TIA with a US location is hard to defend for data under nFADP. For EU/CH workloads Scaleway, OVHcloud or Exoscale is the clean choice.
Anyone seeking extremely low hourly prices for experimental workloads is cheaper at Vast.ai or RunPod Community Cloud. Lambda Labs is around factor 2-3 more expensive than Vast.ai interruptible.
Anyone needing broad regional choice with global distribution is better served by OVHcloud (33 sites) or AWS/Azure. Lambda Labs focuses clearly on the USA, without a global multi-region architecture.
For mandates under professional secrecy (Art. 321 SCC) or banking secrecy (Banking Act Art. 47), Lambda Labs is not acceptable – the CLOUD Act risk cannot be eliminated by standard contractual clauses. Self-hosting on Hetzner GPU or Exoscale A100 in CH is the correct alternative.
General caveat: Lambda Labs availability of H100 SXM and H200 is occasionally on a waitlist in May 2026, especially for reserved contracts. Anyone with a critical deadline should inquire early and possibly keep CoreWeave in parallel as a backup option.
Trade-offs
STRENGTHS
- ML-engineer focused platform with pre-installed Lambda Stack
- Genuine 8x H100 SXM multi-GPU nodes with full InfiniBand topology
- Clear reserved contracts with 30-50% discount versus on-demand
- Lambda Inference as cheap pay-per-token API for open-weight models
WEAKNESSES
- No EU or CH region as of May 2026, all sites in the USA
- CLOUD Act applicability for all workloads, TIA with US location hard to defend
- H100/H200/B200 occasionally on waitlist, reserved contracts with lead time
- No own S3-compatible object storage, connection to AWS S3 or Cloudflare R2 needed
FAQ
Does Lambda Labs have an EU region?
As of May 2026 no. All Lambda Labs data centres are in the USA. For EU/CH data residency requirements Lambda Labs is not directly usable – Scaleway Paris, OVHcloud Frankfurt, Exoscale Zurich or Hetzner GEX are the EU/CH alternatives. Lambda Inference (pay-per-token API) is also US-hosted; for EU residency on inference, Scaleway Generative API or Infomaniak Apertus fits better.
What does an 8x H100 cluster cost over a year?
On-demand: USD 23.92/h × 8,760h = USD 209,500 for 12 months. With 1-year reserved contract: about USD 145,000-165,000 for 12 months, depending on configuration and cluster size. With 3-year reserved: about USD 105,000-125,000 per year over 3 years. With these reserved tariffs Lambda Labs is among the most attractive multi-GPU cluster providers price-wise. CoreWeave is at a comparable level, with somewhat different cluster topology optimisation.
How does Lambda Inference work?
Lambda hosts open-weight models (Llama 3.3 70B, 405B, Qwen 2.5 72B, DeepSeek-V3 and more) on its own H100 hardware in the USA. The API is OpenAI chat completions schema compatible – an existing OpenAI integration can be redirected to Lambda with an endpoint swap. Billing per million input and output tokens, with cheaper tariffs than Together AI or Fireworks AI for some models. Pro: cost-efficient, easy start. Con: US-hosted, CLOUD Act applicability, not usable for EU/CH mandates without TIA.
How does Lambda Labs stand against CoreWeave?
Both are US GPU clouds focused on multi-GPU training workloads. In May 2026 Lambda Labs is often slightly cheaper per H100 hour, with a clear reserved path. CoreWeave has larger clusters (16+ H100 with top InfiniBand topology), a larger multi-region presence (US East/West, UK, Spain), and has moved deeper into enterprise contracts competing with hyperscalers. For standard ML engineering workflows Lambda Labs is often the more natural choice. For top-end enterprise training with the largest clusters CoreWeave is more established.