Lambda

Who owns it

Lambda sells GPU capacity for AI training and inference from the United States, and a UK or EU buyer contracts with the US company under US law. The public record on the company itself is thinner than its size would suggest. Lambda's own about page gives a founding date of 2012 and an origin at Noisebridge in San Francisco, but it names no founders, no legal entity and no registered address. The funding is better documented: a $480m Series D announced on 19 February 2025, co-led by Andra Capital and SGW with NVIDIA among the investors, at a $2.5bn post-money valuation, then more than $1.5bn in a Series E led by TWG Global in November 2025, taking the total raised to roughly $2.3bn. NVIDIA holding a stake in a company that rents out NVIDIA hardware is worth knowing when you are weighing up who gets allocation when supply is short. A listing has been reported for the second half of 2026 with Morgan Stanley, JPMorgan and Citi engaged, but no S-1 was public on 22 September 2026. Lambda buys and runs its own GPUs and puts them in other companies' buildings, including Aligned's DFW-04 in Dallas-Fort Worth, EdgeConneX in Chicago, LAX01 in Vernon, California and a new site in Kansas City. It is not a marketplace: no third-party host is renting you their machine.

## What it costs

Prices are in US dollars, VAT not stated, read 22 September 2026. Lambda charges per GPU per hour, and the rate goes up as you take fewer GPUs, which is the opposite of what most buyers expect. One H100 SXM 80 GB costs $4.29 an hour, or $3,132 a month at 730 hours. Take the whole eight-GPU node and the same GPU costs $3.99 an hour, so the node is $31.92 an hour and each GPU works out at $2,913 a month. Two and four GPU configurations sit in between at $4.19 and $4.09. An H100 PCIe is $3.29 an hour. The cheapest GPU on the list is a Quadro RTX 6000 24 GB at $0.69 an hour, about $504 a month, and a GH200 is $2.29. No H200 appears on the public price list at all, despite the pricing page's own marketing copy mentioning it. There is no spot, preemptible or interruptible tier, so there is no cheap way to buy compute you are willing to lose. Reserved pricing for a single on-demand instance is sales only: no rate, no term and no discount figure is published.

The one published committed option is a 1-Click Cluster, and the arithmetic is counterintuitive. Sixteen H100s cost $6.16 per GPU-hour on a term of two weeks to a year, falling to $5.85 at 64 GPUs and $5.54 at 256. All three are higher than the $3.99 on-demand rate for an eight-GPU node, so a cluster buys InfiniBand and guaranteed capacity rather than a discount. The smallest cluster you can book is 16 GPUs for two weeks, roughly $33,100 paid up front, which puts it out of reach for most small teams. Storage is billed separately and keeps running after the instance has gone: the documentation gives $0.20 per GiB a month as an example and shows the real rate when you create the filesystem, so a 2 TB filesystem left behind costs around $400 a month. Ingress and egress are free and stated as such. On-demand instances bill in one-minute increments and are invoiced weekly in arrears, while clusters bill weekly and are paid in advance. Instances are billed for as long as they are running whether or not they are being used, and no stop-and-keep-the-disk state is documented, so the practical choice is to terminate or keep paying the full rate. The smallest realistic monthly bill for a small team is a Quadro RTX 6000 at $503.70 plus a 500 GB filesystem at about $100, so roughly $604 a month with no transfer charges on top.

## What you get and what to watch

Signup is self-serve, but the card rules stop some buyers at the door. Lambda accepts major credit cards only, and prepaid and debit cards are not supported, which rules out a lot of UK sole traders and small agencies working from a business debit card or a virtual card. A $10 pre-authorisation is taken to check the card and refunded within a few days. Purchases are supported from 18 countries including the UK and the EU, with Lithuania excluded, and transactions attempted over a VPN are flagged as suspicious and may be declined. No new-account GPU quota is published, and the only documented limits are on storage: 24 filesystems per account, capped at 10 TB each in the Texas us-south-1 region. There is no free tier and no signup credit for general users. One secondary account describes manual review of new accounts taking around three days, which Lambda's own documentation does not confirm.

Fourteen regions are listed and nine of them are in the United States. There is one EU region, europe-central-1 in Germany, and the city is not published. There is no UK region and no UK or EU contracting entity, so the most a UK buyer gets is EU data residency under a US contract. Lambda does not publish which GPUs run in which region, which matters more than usual when your only European option is a single site. The platform sells one, two, four and eight GPU instances, 1-Click Clusters from 16 GPUs upwards with NVIDIA Quantum-2 InfiniBand and SHARP, and managed Kubernetes or Slurm on those clusters. Managed Kubernetes is not available on a single on-demand instance. Persistent filesystems can only be attached when an instance is launched, not to one already running. There is a public Cloud API, but the documentation page would not render on 22 September 2026, so endpoints, authentication and rate limits could not be checked. No official Terraform provider, command line tool or MCP server was found.

Compliance is the strongest part of the published record: SOC 2 Type II, ISO 27001, ISO 27017, ISO 27701, ISO 22301 and CCPA, with a trust portal at trust.lambda.ai for evidence requests. GDPR, a public data processing agreement and the CLOUD Act are not addressed on the trust page. Support runs through a portal at support.lambda.ai, with 24/7 monitoring and dedicated escalation channels described for enterprise customers. No uptime percentage and no service credit schedule is published for the self-serve product, and no status page was found, so a card-paying customer has no published uptime commitment to hold Lambda to. Trustpilot shows 2.6 from four unprompted reviews, every one of them a single star. The themes are narrow and recent: an account banned over a payment issue with data and models removed (30 January 2026), an account auto-suspended over a payment problem with persistent storage destroyed alongside the instances (18 January 2026), a card charged as soon as the details were entered, which may be the documented $10 pre-authorisation (9 December 2025), and support unable to close an account (30 November 2025). Four reviews is far too thin to draw a conclusion from, but nothing in Lambda's documentation says what happens to your data when an account is suspended, which is exactly what two of them describe. No outage, breach, regulator action or lawsuit was found on the public record.

Not yet tested by Saaskly. Assessed on published pricing, vendor documentation, company filings, the vendor's own status and trust pages, and third-party reviews.