68 / 100

Lambda pricing teardown

https://lambda.ai/service/gpu-cloud

Lambda's GPU cloud page is a transparent, catalog-style pay-by-the-minute price list that's ideal for developers who know which GPU they want, with strong hero CTAs and no-egress-fee messaging. It's a SKU matrix rather than a packaged tier ladder, so there's no recommended plan, no commitment/discount mechanism, and little guidance for choosing between GPUs.

Tier structure

NVIDIA B200 SXM6

$6.69/GPU/hr (8x)

NVIDIA H100 SXM

$3.99/GPU/hr (8x)

NVIDIA A100 SXM 80 GB

$2.79/GPU/hr (8x)

NVIDIA A100 SXM 40 GB

$1.99/GPU/hr (8x)

NVIDIA Tesla V100

$0.79/GPU/hr (8x)

Value metric

per GPU per hour, billed by the minute

How limits scale

VRAM per GPU 16 GB → 40 GB → 80 GB → 80 GB → 180 GB
vCPUs (8x) 88 → 124 → 240 → 208 → 208
RAM (8x) 448 GiB → 1800 GiB → 1800 GiB → 1800 GiB → 2900 GiB
SSD storage (8x) 5.8 TiB → 5.8 TiB → 19.5 TiB → 22 TiB → 22 TiB
Price per GPU/hr $0.79 → $1.99 → $2.79 → $3.99 → $6.69

Escalation logic

Tiers are GPU generations ordered by performance, with price rising roughly 1.4-2x between adjacent models. A 1x/2x/4x/8x toggle changes instance size, and per-GPU price rises slightly on smaller configs (B200 goes $6.69 to $6.99), a mild volume incentive.

Psychological anchors

  • Top-end anchor — The B200 at $6.69/GPU/hr is promoted in a dedicated section ('2x the VRAM and FLOPS of H100... up to 3x faster training'), making H100 at $3.99 feel like the sensible mid-range choice.
  • Charm pricing — Every price ends in .99, .79, .69, .09 or .29, signaling accessible, developer-friendly pricing.
  • Cost-risk removal — 'Pay by the minute' and 'no egress fees' headline the pricing section, reducing fear of surprise bills and lock-in.
  • Instance-size toggle — The 8x/4x/2x/1x selector lets users start with a single GPU and scale up, creating a low-commitment entry and natural expansion path.
  • Cheap entry SKUs — Older GPUs (V100 at $0.79, Quadro RTX 6000 at $0.69 on 1x) serve as low-price entry anchors for experimentation.
  • Escape hatch to sales — 'Looking for long-term capacity? Talk to our team' and 1-Click Clusters upsell funnel larger buyers off the self-serve path.

What this page is optimizing for

Self-serve developer conversion: public per-minute prices, repeated 'Launch GPU instance' CTAs, 'Get started' in the nav, and UI/API/CLI messaging all push toward instant signup. Enterprise capacity is handled by a secondary 'Talk to our team' link.

Red flags

  • No recommended or 'best for' labeling on any GPU, so newcomers must already know whether they need an A100 vs H100 vs B200.
  • No discount for reserved or committed usage on this page, and no annual/prepay option, which leaves savings on the table for steady workloads.
  • No free trial or starter credits, so first-time users must pay before testing a GPU.
  • Prices are per GPU per hour while instances are sold as 8x/4x/2x/1x, so users must multiply to estimate real cost; there is no cost calculator or monthly estimate.
  • Per-GPU price is higher on smaller instances and the 8x/4x/2x/1x toggle hides the 1x-only SKUs (GH200, A10, A6000), which can confuse comparison across configurations.
  • Catalog of 20+ SKUs across four views makes it hard to scan differences; there is no use-case guidance (training vs fine-tuning vs inference).
  • 'Self-serve, first-come access' implies availability can be constrained, but the page does not show live capacity or availability per GPU.

Best-practices scorecard

What works · 2

  • Visible prices — Every SKU shows a transparent per-GPU hourly price with specs, and the page notes tax is extra.
  • Value metric matches usage — Per-minute billing by GPU matches how ML teams consume compute and scales fairly with usage.

Half measures · 3

  • Tier differences are scannable — Consistent spec columns (VRAM, vCPUs, RAM, storage, price) make comparison easy, but there is no guidance on which GPU suits which workload.
  • FAQ or objection handling — Objections are partly addressed via 'no egress fees', Lambda Stack, and observability blurbs, but there is no FAQ on billing, availability, or quotas.
  • Clear CTAs per tier — Strong repeated 'Launch GPU instance' CTAs exist, but they are page-level rather than attached to each GPU row.

What's missing · 3

  • Clear recommended tier — No plan is highlighted or badged; only the B200 gets editorial emphasis in a separate section.
  • Annual discount offered — No prepay, commitment, or reserved discount is shown; long-term capacity is routed to sales.
  • Trust signals present — No customer logos, testimonials, or security badges appear on the page itself, despite links to customer stories elsewhere.