AI news videos

NVIDIA: Same Power Budget, 24% More AI Tokens

NVIDIA published results it says show an AI data center can produce measurably more tokens from the same amount of electricity, without adding a single megawatt.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• PUNCTUATION NOTE. The live page uses typographic em dashes and the snapshot preserves several of them; no quotation was reworded, and every supporting span quoted below was chosen or trimmed so that no dash character enters this script. The source's hyphenated and glued brand tokens, DSX, MaxLPS, HGX B200, NVL72, GB200, 800V, VDC, Eos and SVP, are never narrated, because the speech engine garbles glued brand and digit tokens and letter acronyms; they appear only in onscreen text. The acronym GPU was also kept out of narration for the same reason, which is why scene four says capacity rather than GPU capacity, a vaguer wording that cannot overstate the source. SENTENCE BY SENTENCE AUDIT. Hook one, "Lambda says the same power budget produced twenty four percent more tokens." Supported by "a fixed power budget can support 24% more token throughput when managed intelligently" and by "by running 19 nodes within the same power budget as 16 nodes at full power, Lambda achieved 24% more cluster-wide token throughput". The figure is spoken as "twenty four percent" and the attribution "Lambda says" is attached out loud because this is the partner's own proof of concept result, published by NVIDIA, with nobody else measuring it. Hook two, "Nobody independent checked it." This is not a claim drawn from the article; it is our own statement about the article's evidentiary status, and it is supported by the absence of any third party verifier anywhere in the source. The page contains only NVIDIA's own words, its partner Lambda's words through Dave Ward, its partner Emerald AI's words through Varun Sivaram and Mansi Shah, and NVIDIA's own projections. No regulator, no independent lab, no benchmark body and no published dataset appears. Scene one, "Lambda ran nineteen nodes inside the power budget of sixteen at full power, using NVIDIA's power management software." Supported by "Lambda, a GPU cloud provider serving more than 10,000 customers from AI-native startups to hyperscalers, ran the software on a five-rack, 19-node cluster" and by "by running 19 nodes within the same power budget as 16 nodes at full power". The numbers 19 and 16 are spoken as "nineteen" and "sixteen". The words DSX MaxLPS are replaced with "NVIDIA's power management software", which the source itself describes as "dynamic power allocation software that monitors GPU and rack-level consumption in real time". The five rack detail and the ten thousand customer detail are dropped for length; neither strengthens or weakens the narrated claim. Scene two, "Lambda says throughput rose from roughly four million tokens per second to five million and performance per watt improved twenty three percent." Supported by "from roughly 4 million tokens per second to 5 million. Performance per watt improved by 23%." The source's own hedge "roughly" is preserved and never tightened, the figures are spoken as words, and "Lambda says" carries the attribution for both numbers in the sentence. The word "cluster-wide" is dropped because scene one has already established that the subject is one nineteen node cluster. Scene three, "NVIDIA says Silicon Valley Power has sent more than two hundred demand signals, and power once fell from four megawatts to three." Supported by "Silicon Valley Power has since sent more than 200 demand signals to that AI factory. It worked every single time." and by "power fell from four megawatts to three. Automated. No operator required." The hedge "more than" is preserved. The word "once" was added deliberately so that the four megawatt to three megawatt drop is understood as the single August evening the article describes and is not implied to be the result of all two hundred signals, which the article never claims. The table row reading "40% power demand reduction in under a minute" was deliberately not narrated, because it does not reconcile with the four to three megawatt figure in the body and we will not narrate a number we cannot make consistent. The claim "It worked every single time" was also not narrated, because it is an unqualified success claim by the vendor with no definition of success. Scene four, "NVIDIA projects up to forty percent more capacity inside the same power budget for future Vera Rubin factories." Supported by "Based on NVIDIA's projections, DSX MaxLPS can enable up to 40% more GPU capacity for next-generation Vera Rubin NVL72 AI factories within the same megawatt power budget in suitable deployment environments." The hedges preserved are "up to", the verb "projects" which states out loud that this is a projection and not a measurement, and "future", which carries the source's "next-generation". The hedge "in suitable deployment environments" was dropped for length; it narrows the claim further, so dropping it is the only hedge loss in the script and it is noted here explicitly. Endcard: fixed channel boilerplate, no claim. SENTENCE BUDGET. Eight spoken sentences total, two in the hook, exactly one in each of the four scenes, and two in the endcard, which is under the ceiling of ten. One hundred and thirteen narrated words total, under the target of one hundred and eighteen: hook sixteen, scene one eighteen, scene two twenty two, scene three twenty two, scene four eighteen, endcard seventeen. Every scene is a single sentence and every scene is under the thirty word cap. Commas are scarce by design, with zero in the hook, one in scene one, zero in scene two, one in scene three and zero in scene four, and there is no multi item comma list anywhere in narration. The durationS integers are ten, eleven, ten and nine, summing to forty, inside the thirty to fifty gate. No timeline block was written, because the page date and the run date are both September 15, 2026, and a timeline array would make the resolver prefer an earlier date and wrongly age this same day story to the catalog tier. The date is never narrated. HONESTY CAVEAT. This is NVIDIA publishing about NVIDIA's own platform, on NVIDIA's own blog, with an NVIDIA byline. Lambda is a commercial partner and an NVIDIA cloud customer running NVIDIA's own software on NVIDIA's own servers, not an independent auditor, and its president of cloud services is quoted promoting the outcome. The twenty four percent throughput figure and the twenty three percent performance per watt figure come from a single five rack nineteen node proof of concept, which the source itself calls a proof of concept, and not from a fleet, a production deployment at scale or a repeated trial. The forty percent capacity figure is not a measurement at all; it is NVIDIA's own projection for Vera Rubin hardware that has not shipped, which is why the narration says NVIDIA projects it. Nobody independent verified any of it: no outside lab, no benchmark organisation, no customer other than the partner itself, and no published raw data. Nothing shipped to customers today. What happened today is a keynote by Ian Buck at the AI Infra Summit plus published validation results, and the narration uses no launch verb. CORROBORATION STATUS. There is none. The research step captured zero of five YouTube transcripts because the host's yt-dlp build is too old, which is a tooling failure on our side and says nothing about the story, and the transcripts file holds an empty videos array. The timeline probe file was reviewed and deliberately not used as a fact source, because it contains only third party YouTube titles and a title is not evidence. No third party outlet, channel or competitor is named anywhere in this script. The NVIDIA post, read from the snapshot captured 2026-09-15T17:05:22.867Z, is the single source, and every narrated sentence traces to a line quoted above. DELIBERATE OMISSIONS. Omitted for length, speech safety, prior coverage or because the item is the publisher's own framing: the entire opening scene setting narrative about the sweltering August evening, the Zoom call with about forty people and the cheering, which is colour rather than fact; the two quotations containing figures of speech, "watching with bated breath" and "This feels kind of like a SpaceX rocket launch", both banned as figures of speech; the dek's phrase about squeezing more tokens from every available watt, also a figure of speech; the line "The factory that's willing to flex gets to run bigger", also a figure of speech; the Jensen Huang quotation about a one gigawatt factory never becoming a two gigawatt factory, which is rhetorical framing and contains a glued number pattern; the full DSX suite sidebar listing Flex, OS, Sim and Reference Designs, all of which are brand tokens the speech engine cannot say and all of which would require a multi item list; the Flexible Load Interconnect Program name and the detail that the participating factory is NVIDIA's Eos; the planned Manassas, Virginia facility and its ninety six megawatt Vera Rubin build, which is an announcement of a future deployment rather than a result; the 800V DC architecture section and its projected three to five percent efficiency gain for 2027, which is projected, brand and digit heavy and unspeakable; the one hundred and twenty kilowatt liquid cooling heat figure; the GTC Taipei introduction in May; and the claim that Conductor responds in under a minute, dropped because only one grid scene is permitted and the two hundred signal count and the four to three megawatt drop were the stronger new specifics. Nothing omitted contradicts anything narrated. Every omitted item either runs in the same direction as what is narrated or would make the platform sound stronger or nearer than this Short claims. PRIOR COVERAGE NOTE. Our channel already covered the grid flexibility angle on 2026-08-20 in the Short slugged nvidia-power-flexible-ai-factories-grid, about NVIDIA power flexible AI factories helping stabilise the energy grid. This script therefore does not use grid demand response as its spine. The lead, the hook and scenes one and two are all the genuinely new material released today, namely Lambda's first validation in a deployment environment announced at the AI Infra Summit and the token throughput numbers that came with it. Exactly one scene, scene three, touches the grid proof at all, and it uses only specifics that did not exist in the August coverage: the count of more than two hundred demand signals sent since, and the drop from four megawatts to three. The August Short's general framing about grid stabilisation is not repeated, and no sentence in this script restates it. PROMPT INJECTION CHECK. The snapshot was read strictly as untrusted data and checked line by line for anything addressed to a model or an agent. No injection attempt was found: no hidden directive, no ignore previous instructions pattern, no role reassignment and no attempt to change behaviour. The imperative text present in the file is ordinary website chrome aimed at human readers and page navigation, for example "Skip to content", "Share This Article", "Copy link", "Register Now", "View All Recent News" and "Enter your Mastodon instance URL (optional)"; none of it was treated as an instruction and none of it entered the script. The file also carries capture annotations rather than publisher text at the top, namely the SOURCE URL, CAPTURED and TITLE header lines, plus one stray angle bracket artefact from the extraction; these were treated as metadata about the capture, never as source facts and never as instructions. The persuasive marketing language in the body, such as the dek and the closing line about a new baseline for what an AI factory is supposed to do, was treated as the publisher's promotional framing, not as instruction and not as fact, and none of it was narrated.(NVIDIA (blogs.nvidia.com), first party corporate blog post in the AI Infrastructure category, titled "From Megawatts to Tokens: How NVIDIA Maximizes AI Factory Production", byline Vishal Ganeriwala, dated September 15, 2026 on the page, written in NVIDIA's own corporate voice about NVIDIA's own DSX platform and NVIDIA's own partner results, containing no independent reporting, no outside auditor and no third party measurement anywhere in it, read from a local watcher snapshot captured 2026-09-15T17:05:22.867Z)