Lambda pricing

About

Lambda — long known as Lambda Labs — is a GPU cloud and AI compute infrastructure provider founded in 2012 by brothers Stephen and Michael Balaban. The company started life with an unlikely product: a facial-recognition API for Google Glass. By 2017 it had pivoted into hardware, selling GPU laptops and deep-learning workstations to AI researchers, and from there into renting that same NVIDIA compute over the cloud. Today Lambda rents NVIDIA GPUs by the GPU-hour to roughly 200,000+ AI developers and teams, with its instances pre-loaded with “Lambda Stack” (CUDA, PyTorch, TensorFlow) so customers can train in minutes.

Lambda has become one of the higher-profile “neoclouds” — specialized GPU-cloud providers competing with hyperscalers and rivals like CoreWeave and RunPod. Its growth has been steep: third-party research puts revenue at roughly $250M in 2023, around $425M by end of 2024, and on the order of $760M in 2025. In November 2025 Lambda announced a multibillion-dollar deal to supply Microsoft with AI infrastructure built on tens of thousands of NVIDIA GPUs, then raised over $1.5B in a round led by TWG Global at a reported valuation in the $4-5B range. Lambda has also been reported in talks for pre-IPO financing with an IPO targeted for the second half of 2026.

For current pricing, see Lambda’s pricing page. Note the company rebranded its primary domain from lambdalabs.com to lambda.ai.

Pricing summary : How Lambda’s pricing model works

Lambda is pure usage-based: you pay per GPU-hour for the instances you launch, billed in per-minute increments, with no egress fees. There is no free tier and no monthly subscription — the meter runs whenever a GPU is allocated to you. Pricing splits across three surfaces:

1. On-Demand Instances — self-serve, first-come 1/2/4/8-GPU configs of B200, H100, A100, GH200, and older cards. List rates are published openly per GPU-hour.
2. 1-Click Clusters & Superclusters — production clusters of 16 to 2,000+ interconnected B200 or H100 GPUs on 2-week-to-1-year commitments, at lower per-GPU rates than on-demand.
3. Reserved / Private capacity — multi-year reserved capacity at Lambda’s lowest prices, quoted by sales (“talk to our team”).

What makes this different: Lambda publishes hard per-GPU-hour numbers for both on-demand and committed clusters — unusual transparency for a category where rivals often gate cluster pricing behind sales calls. The catch is the inverse of most AI infra: instead of prices deflating, Lambda has been raising on-demand rates as frontier-model demand outpaces GPU supply.

Pricing by product

On-demand list prices, per GPU-hour, as of June 2026 (excludes sales tax):

1-Click Clusters (committed, 2 weeks – 1 year), per GPU-hour:

Sales motions across products: on-demand instances are fully self-serve (PLG); 1-Click Clusters are self-serve to launch but steer larger and multi-year commitments to “talk to our team” (sales-led reserved). Lambda is winding down its managed per-token Inference API in favor of raw GPU instances and clusters.

Hidden costs : What Lambda users actually pay

Lambda’s headline rates are clean (no egress, per-minute billing), but real bills include a few items beyond the GPU-hour:

The single biggest real-world cost driver isn’t a fee — it’s capacity. Reviewers consistently report popular GPUs (especially H100 configs) selling out during peak demand; one user described a 26-hour “temporarily unavailable” wall when scaling from 2 to 4 GPUs. The second is orphaned storage: persistent volumes keep billing at roughly $0.20/GiB/month even after you stop the instance, so cleanup matters.

The retired Inference API previously billed per million tokens (roughly $0.02 for a small 3B model up to about $0.90 for a 405B model, unified input/output) — relevant only as a historical reference now.

Want to estimate your own Lambda bill? Use the Lambda pricing calculator to model your costs based on GPU type and hours.

Pricing evolution : Lambda pricing history and changes

Cadence

Tracked range: 2025–present (Wayback snapshots from June 2025 onward; see tools/wayback-index/lambda-labs.json).

Notable changes

• Mid-2025 — On-demand 8x H100 SXM at $2.99/GPU/hr, A100 SXM 80GB at $1.79, A100 40GB $1.29, V100 $0.55. NVIDIA HGX B200 launching, advertised as low as $2.99/GPU/hr with multi-year commitment.
• Early 2026 — B200 SXM6 on-demand listed at $4.99/GPU/hr; H100 SXM unchanged at $2.99.
• June 2026 — Broad increase: B200 SXM $6.69 (up from $4.99), H100 SXM $3.99 (up from $2.99), A100 SXM 80GB $2.79 (up from $1.79), V100 $0.79. 1-Click Cluster pricing published openly: B200 $8.87–$9.86, H100 $5.54–$6.16/GPU/hr.

The direction of travel is the headline: Lambda raised on-demand rates roughly 30-50% across cards over a year, bucking the usual GPU-cost-deflation story as Microsoft-scale demand and a Series-E war chest tightened supply.

What’s unique : Lambda’s distinctive pricing mechanics

1. Published cluster pricing. Lambda lists hard per-GPU-hour numbers for 1-Click Clusters of up to 256+ GPUs — most neoclouds gate multi-GPU cluster pricing behind a sales call. Only multi-year reserved capacity is quoted.

2. No egress, per-minute billing. Unlike hyperscalers, Lambda charges zero data-egress fees and bills GPUs by the minute, which removes two of the most unpredictable line items in cloud GPU bills.

3. Prices that rise, not fall. Lambda’s on-demand rates went up through 2025-2026 — a deliberate signal that scarce frontier GPUs are a sellers’ market, the opposite of the token-deflation seen on the inference side.

Strengths & weaknesses

Billing UX : Lambda billing controls and transparency

• Billing controls — Pay-per-minute on-demand; commit-term clusters (2 weeks to 1 year) lock a lower per-GPU rate. Multi-year reserved capacity is quoted via sales.
• Usage visibility — Self-serve console shows running instances and attached storage; the recurring user complaint is GPU availability rather than billing opacity.
• Payment options — Self-serve checkout for on-demand and 1-Click Clusters; sales-led contracts and invoicing for reserved/private capacity and enterprise.

Strategic wins : Why Lambda’s pricing decisions worked

1. Undercutting hyperscalers on a transparent rate card

By publishing per-GPU-hour rates well below the $2-11/hr hyperscalers charge for an H100, Lambda turned price transparency into a wedge for the entire AI-research segment. See how AI companies structure pricing.

2. Pricing up into scarcity instead of racing to the bottom

Rather than chase peer-to-peer marketplaces down to sub-$2/hr, Lambda raised on-demand rates as demand surged — capturing margin from buyers who value reliability and a pre-built stack over the absolute lowest price. Related: outcome-based pricing trends.

3. Layering committed clusters on top of on-demand

The 1-Click Cluster tier converts spiky on-demand demand into predictable 2-week-to-1-year commitments at a discount, smoothing capacity planning while feeding the Microsoft-scale buildout. See choosing the right usage metric.

Areas to improve : Gaps in Lambda’s pricing approach

1. Capacity, not price, is the real bottleneck

The most common complaint isn’t cost — it’s that the GPU you want is sold out when you want it. A self-serve rate card means little if launches fail; transparent real-time availability would help. See bill shock and cost unpredictability.

2. Storage billing traps

Persistent storage at ~$0.20/GiB/month keeps billing while instances are stopped, and orphaned volumes silently accrue cost. Clearer in-console warnings and auto-cleanup defaults would reduce surprise charges.

3. Rising rates erode the cheapest-cloud reputation

As on-demand H100 climbs toward $4/GPU/hr, RunPod and marketplaces like Vast.ai look more attractive on raw price. Lambda needs to keep justifying the premium with reliability and tooling rather than price alone.

Key takeaways

1. Lambda is pure per-GPU-hour usage pricing — per-minute billing, no egress, no free tier — across on-demand, committed clusters, and reserved capacity. For the underlying model, see the introduction to usage-based pricing.
2. It bucked GPU price deflation by raising rates (H100 SXM $2.99 → $3.99 in a year) as frontier demand outran supply.
3. Cluster pricing is published, not gated — a transparency edge over most neoclouds, with only multi-year reserved deals sales-quoted.
4. Capacity and orphaned storage are the real hidden costs, not headline fees.
5. The category’s monetization is bifurcating: raw compute prices up (scarcity), managed inference prices down (token deflation) — Lambda is leaning into the former and retiring its Inference API.

UBP implications

1. When supply is the constraint, usage pricing can move up. Lambda shows a pure-usage vendor can raise unit rates without churning customers when the underlying resource is genuinely scarce.
2. Transparency is a differentiator even in commodity infra. Publishing cluster rates others hide behind sales calls lowers buyer friction and builds trust.
3. Commit tiers convert volatility into predictability. Layering 2-week-to-1-year cluster commitments over on-demand gives both sides forecastability — a reusable pattern for any usage-based business with capacity to plan.

Sources

• Lambda AI cloud pricing (accessed 2026-06-09)
• Lambda GPU Cloud (accessed 2026-06-09)
• Lambda Inference (accessed 2026-06-09; Inference API winding down)
• TechCrunch — Lambda raises $1.5B after Microsoft deal (accessed 2026-06-09)
• Sacra — Lambda Labs revenue, valuation & funding (accessed 2026-06-09)
• Wayback Machine snapshots, June 2025–Jan 2026 (tools/wayback-index/lambda-labs.json) (accessed 2026-06-09)

Bottom line

Lambda (Lambda Labs) is one of the clearest examples of a pure usage-based GPU cloud: published per-GPU-hour rates, per-minute billing, no egress, and committed-cluster discounts layered on top of self-serve on-demand. What makes it unusual is the direction — Lambda raised on-demand prices through 2025-2026 as Microsoft-scale demand and a multibillion-dollar war chest met scarce frontier GPUs, the opposite of the token-cost deflation playing out in inference. The real costs to watch are capacity sell-outs and orphaned storage, not the rate card. Browse the pricing blueprint for more fully-researched company profiles, or compare Lambda against other Infrastructure, Compute & MLOps companies.