Lambda (Lambda Labs): GPU Cloud and Hardware for AI

Lambda is a company that sells access to graphics processing units, or GPUs, the chips that do most of the arithmetic in modern AI. It does this in three main ways: renting individual GPU machines by the hour through Lambda Cloud, renting or building large clusters of thousands of GPUs for big customers, and, historically, selling physical workstations and servers preloaded with AI software. For most readers the relevant product is the cloud: a developer can sign in, choose a machine with one to eight NVIDIA GPUs, connect over SSH, and start training or running a model within minutes, paying only for the time used.

The company's story starts with face recognition. Next Platform's 2020 profile reports that Stephen Balaban was the first hire at Perceptio, a neural-network face recognition startup that Apple acquired in August 2013, and that he founded Lambda a few months later with his twin brother Michael Balaban; TechCrunch and Verdict give 2012 as the founding year. In 2015 the founders built a selfie-to-art-style app on a single AWS GPU server. It went viral, reaching about 30,000 sign-ups a day at its peak, and the AWS bill climbed to roughly $40,000 a month. Building their own GPU machines for about $60,000, hosted at first in a friend's house, cut that bill by about 95%, and that experience became the business. At the end of 2016 Lambda began selling AI servers, followed in 2017 by its Quad workstations and Blade servers, and in 2018 it formally announced Lambda GPU Cloud and released Lambda Stack.

Renting a GPU on Lambda works much like any cloud virtual machine, with fewer layers. According to Lambda's documentation, a user picks an instance type and one of 13 regions, adds an SSH key, and optionally attaches a persistent filesystem in the same region. The default image is Ubuntu 22.04 with Lambda Stack, which bundles NVIDIA drivers, CUDA, cuDNN, NCCL, PyTorch, TensorFlow, JAX, Triton and JupyterLab, so a training script can usually run without installing a framework first. Only SSH port 22 is open by default. Filesystems are billed per gigabyte per month and keep billing after the machine is deleted, a common source of surprise charges.

Prices on Lambda's public pricing page in October 2026 illustrate the range. An NVIDIA B200 with 180 GB of memory costs $6.69 per GPU-hour on an 8-GPU instance and $6.99 on a single-GPU instance. An H100 SXM with 80 GB costs $3.99 to $4.29 per GPU-hour depending on instance size, an H100 PCIe $3.29, an A100 40 GB $1.99, and older cards such as the V100 as little as $0.79. For multi-node work, 1-Click Clusters reserved for two weeks to a year are listed at $5.54 to $6.16 per H100 GPU-hour and $8.87 to $9.86 per B200 GPU-hour, falling with cluster size. Longer and larger commitments are priced on request.

From about 2025 the company's scale changed sharply. Verdict reported a $480 million Series D in February 2025 and a Series E of more than $1.5 billion in November 2025, led by TWG Global, aimed at building what the company called gigawatt-scale AI factories. In November 2025 Microsoft signed a multibillion-dollar agreement for Lambda to deploy tens of thousands of NVIDIA GPUs, including GB300 NVL72 rack systems; CEO Stephen Balaban told TechCrunch that Lambda had worked with Microsoft for more than eight years. In May 2026, Bloomberg reported that former Sprint chief executive Michel Combes became CEO, with Stephen Balaban moving to chief technology officer and Michael Balaban to chief product officer. In October 2026 TechCrunch, citing The Wall Street Journal, reported that Lambda was raising up to $4 billion at a $14.5 billion pre-money valuation ahead of a planned 2027 public offering, and that its backlog of contracted orders had grown from $15 billion in June to $50 billion in September.

Compared with the large general-purpose clouds such as Amazon Web Services, Microsoft Azure and Google Cloud, Lambda offers a narrower menu focused on GPUs, simple hourly pricing and a preconfigured machine learning image, rather than a broad catalog of managed services. Compared with other GPU-focused providers such as CoreWeave and Nebius, which TechCrunch groups with Lambda as NVIDIA-backed neoclouds, the differences come down to price, available GPU models, cluster networking, regions and contract terms, all of which change often. Buying hardware outright remains an option for steady, round-the-clock workloads; renting suits bursty experiments and projects that need the newest chips without a capital purchase.

There are practical limits. Popular GPU types are not always available in every region, and on-demand capacity for the newest chips can run out at busy times. Each instance lives in one region, and filesystems must be in the same region. The GH200 instance uses an ARM processor, so software built only for x86 will not run on it. Lambda's documentation also prohibits cryptocurrency mining. At the company level, TechCrunch noted in October 2026 that a large share of Lambda's backlog depends on a single customer, Anthropic, which concentrates risk.

In practice, using Lambda requires an account with a payment method, an SSH key, and a plan for data: where the dataset lives, how it gets to the region, and how checkpoints are saved to a persistent filesystem before the instance is terminated. The main cost is GPU-hours, so the most effective savings come from testing code on a cheap single-GPU instance first, then scaling up only when the job is known to work.