Summary

  • On 12 November 2026 Lambda's pricing page stated that the figures it publishes take effect on 16 November 2026 at 00:00 UTC, and told readers to log in to see the price currently charged.
  • All four published 1-Click Clusters rows carry the same action label, "Talk to our team", and the row covering one-year-plus commitments from 16 GPUs shows an em dash where a price belongs.

The footnote that resets the page

Lambda's pricing page opens with a promise: "Clear, straightforward pricing for Instances, 1-Click Clusters™, and Superclusters." The footnote beneath the tables takes back the most useful part of it. "These prices take effect November 16, 2026, at 00:00 UTC," it reads. "Until then, log in to your Lambda Cloud account to see current pricing."

A price list is supposed to answer the question it raises. This one answers a different question — what the operator intends to charge four days after the reader arrives — and refers the live question to an account not every reader has. Announcing a repricing before it bites is more candid than swapping the numbers overnight under a live page. The narrower fact still stands: on the day it was read, the page did not state what a new customer pays.

A ladder with a gentle slope

Lambda's 1-Click Clusters page publishes four NVIDIA HGX B200 rows under the column headings Plan, DURATION, GPU COUNT, PRICE/GPU/HR* and Action. At 16 GPUs across a commitment of two weeks to one year, the figure is $9.86 per GPU-hour. At 64 GPUs it is $9.36. At 256 GPUs and above, $8.87. On the row for one year and beyond, starting at 16 GPUs, the price cell holds an em dash.

The slope is real and it is modest. Moving from 16 to 64 GPUs takes 5.1% off the hourly rate; moving from 64 to 256-plus takes off another 5.2%; end to end the discount is about 10%. Multiplied out at each row's own GPU count, the cluster costs $157.76, $599.04 and $2,270.72 per hour respectively. A year of 16 GPUs at the published $9.86, if that rate held flat for all 8,760 hours, would come to roughly $1.38 million. These are arithmetic on Lambda's own figures, in the publisher's own unit, not prices anyone was quoted.

The pricing page puts the same accelerator in a second context: NVIDIA H100 systems, 16 GPUs, two weeks to one year, at $6.16 per GPU-hour. On those two published numbers the B200 costs just over 60% more per GPU-hour than the H100.

The column that ends every row

Run your eye across the cluster table instead of down it and the ladder stops mattering. All four rows carry one action: "Talk to our team". No checkout, no confirmation control, no second price for the buyer who wants to click and go. The page then adds its own gloss: "Short-term and long-term contracts with POC environments available upon request."

"Available upon request" is the same instrument as "talk to our team", written in a softer register. The pattern matters because it is consistent: the table prices three tiers precisely enough to compare, then declines to price the commitment it most wants — a year or more — and routes the transaction through a conversation that leaves no public trace of the agreed rate.

The asterisk, and the one that does not resolve

The header PRICE/GPU/HR* carries an asterisk. What condition that asterisk attaches is not resolved anywhere in the text captured from the page, and it will not be guessed here. The pricing page's separate asterisk does resolve cleanly, and to something mundane: "Plus applicable sales tax/VAT/GST." Tax treatment sits outside the quoted rate; the cluster header's condition remains unstated in what Lambda's page delivered.

The evidence boundary

This briefing rests on two readable captures of Lambda's own pages, both taken on 12 November 2026. Lambda's terms-of-service page was requested twice that day and returned no readable snapshot on either attempt, so nothing here says whether Lambda's contract contains or omits availability commitments, service levels, credit remedies or termination rights — and the absence of that document from this briefing is not evidence about its contents.

Both retained pages are fragments of rendered text rather than complete captures. Rows, footnotes or clauses outside the captured text are not described, and their absence here says nothing about their absence from the page. Lambda's BTW directory entry sits at btw.media/en/directory/lambda-ai, where the company's record is kept alongside the earlier BTW report from July 2026, which examined whether GPU work could be started, checkpointed, restarted and billed predictably. That question is not re-argued here.

Why the distinction pays

GPU capacity is bought on schedules of weeks to years against hardware whose value on the second-hand market falls faster than the contract clock. The published rate card is the one document a buyer can compare across suppliers without opening a negotiation. Date that card to a future instant, and the comparable number becomes a forecast.

Leave the longest commitment unpriced, and the largest purchase on the page has no reference value at all.

The buyer's move follows from that, and it is not complicated: treat $9.86 as the benchmark Lambda has published it to be, ask what the operative rate is on the day of signing, and require the one-year-plus figure in writing before committing capital to a table that declines to quote it.