Huang Puts Vera Rubin at 30x Blackwell on Tokens per Dollar

On stage at Nvidia’s GTC Berlin keynote on 21 October 2026, Jensen Huang said Vera Rubin is another 30 times beyond Blackwell, while a slide behind him showed “30X” in tokens per dollar on a SemiAnalysis chart. Nvidia’s own release the same day puts the figure differently: Vera Rubin NVL72 systems “deliver 30x higher throughput per megawatt” for agentic workloads.

In plain terms, tokens are the units of text an AI model reads and writes, so tokens per dollar is how much AI output a buyer gets for its money, and throughput per megawatt is how much it gets for its electricity. Huang also showed a Vera Rubin node on stage, which TradeTheNews reported takes less than five minutes to assemble, against two hours for a Grace Blackwell node.

Business Pill · TOKENS PER WATT

A Business Engineer explainer on tokens per watt, what an AI data centre actually produces. It covers the concept in general and says nothing about Nvidia.

Watch · The Keynote

NVIDIA GTC Berlin 2026 keynote with Jensen Huang, 21 October (NVIDIA’s YouTube channel). Quotations in this piece come from Nvidia’s releases, the keynote slide and X posts, not from the video.

What Huang Said on Stage

• Another 30x. Huang said Blackwell was about 30 times faster than Hopper at large language model processing, and that Vera Rubin is another 30 times beyond Blackwell in two years, according to an automated transcript of the stream compiled by our monitoring. He did not give benchmark conditions on stage.
• In production. Huang reiterated that Vera Rubin is already “in full production”, TradeTheNews reported from the keynote.
• Faster to build. TradeTheNews reported that Vera Rubin is “100% liquid-cooled” and takes less than five minutes to assemble versus two hours for a Grace Blackwell node. On stage, Huang said Grace Blackwell was 80% liquid cooled, according to the transcript.
• Ten years. “With NVIDIA extreme co-design, we’ve accelerated AI performance a millionfold over 10 years, and we’re not stopping there,” Nvidia posted on X during the keynote. On stage, Huang compared that with about a hundred times under Moore’s Law, according to the transcript.

What Was Posted on X

TradeTheNews reported from the GTC Berlin keynote that Nvidia reiterated Vera Rubin is already in full production, is 100% liquid-cooled and takes less than five minutes to assemble versus two hours for a Grace Blackwell node. This piece uses its post as the source for the build-time figures.

Nvidia posted that with extreme co-design it has accelerated AI performance a millionfold over 10 years, and that it is not stopping there. This piece cites it as the company’s own framing of the past decade.

Three Versions of the 30x

The same number appeared three ways on 21 October, and each measures something different.

• On stage: Vera Rubin as another 30 times beyond Blackwell, with no workload or metric stated, according to the transcript.
• On the slide: “30X” above “Tokens/$”, on a SemiAnalysis InferenceX chart of agentic tokens per dollar of total cost of ownership that compares Vera Rubin NVL72 with GB300 NVL72. The slide’s label calls the Vera Rubin result an official preview and says results “may change as validation and publication continue.”
• In the release: “For agentic workloads, NVIDIA Vera Rubin NVL72 systems deliver 30x higher throughput per megawatt,” Nvidia wrote in its OpenShell Reference Architecture release. That sentence does not name the system it is compared with.

The slide carries a quotation credited to SemiAnalysis: “Vera Rubin will generate significantly cheaper tokens than the next-leading accelerator and thus make you significantly more money.”

Our SemiAnalysis piece showed why one multiple rarely tells the whole story. SemiAnalysis’s own Vera Rubin post gave about 67x the throughput per TCO of GB300 at 170 TPS, and 1.4x to 3x at 60 to 100 TPS, where it said most providers would serve. The multiple depends on where on the speed curve it is read.

Why the Build Time Matters

A node is one tray of chips in a rack. Huang’s point was that a Vera Rubin node has no cables to connect and is cooled entirely by liquid, so it can be put together in minutes instead of hours. Nvidia’s DSX release on the same day says “NVIDIA Vera Rubin infrastructure supports 100% liquid cooling with coolant temperatures of up to 45 degrees Celsius.”

On our arithmetic, two hours against less than five minutes is a cut of more than 24 times. TradeTheNews reported that the change improves manufacturing cycle times and reliability.

Connecting the Coverage

Earlier today, our DSX POD piece covered Nvidia’s claim that factory-built Vera Rubin containers can take deployment at a prepared site from six months to as little as 14 days, and our cooling-water piece covered its claim that DSX designs can bring facility cooling water to near zero. Both were announced at GTC Berlin the same day.

Our $164.5B piece covered The Information’s tally of Nvidia guarantees and commitments behind new data centres. Our OCP clip piece covered a separate claim, from the OCP Global Summit, that a third-generation MGX design cut build time from 90 minutes to one minute. Our vLLM piece reported a vLLM figure of up to 7.84x the throughput per GPU of GB200 NVL72 at matched interactivity, also labelled an InferenceX preview.

What is new today is Huang putting Vera Rubin’s gain, its build time and its production status in one keynote, next to an Nvidia release that states the 30x per megawatt.

The Structural Read

As we read it, the tokens-per-dollar framing on the slide is the one aimed at buyers, because it folds the chip price, the power and the rest of the cost of ownership into what one unit of AI output costs.

As we read it, the build-time claim matters as much as the speed claim: TradeTheNews reported that a Vera Rubin node takes less than five minutes to assemble against two hours for Grace Blackwell, which bears on how fast new capacity can be put to work.

As we read it, the 30x is a point on a curve, not a constant: SemiAnalysis reported about 67x at 170 TPS and 1.4x to 3x at 60 to 100 TPS in its own Vera Rubin results, as our SemiAnalysis piece covered.

Three Implications

FOR AI CLOUD BUYERS The slide measured Vera Rubin in agentic tokens per dollar of total cost of ownership against GB300 NVL72, and marked the Vera Rubin result as an official preview whose results may change.

FOR DATA CENTRE OPERATORS Nvidia’s release measures the gain per megawatt, the unit that matters where power is the limit, and its DSX release says Vera Rubin supports 100% liquid cooling.

FOR INVESTORS Huang put Blackwell at about 30 times Hopper and Vera Rubin at another 30 times beyond Blackwell, per the automated transcript, without stating benchmark conditions for either figure.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework starts from one unit: the token. “It is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”

As we read it, that is why the slide’s tokens-per-dollar framing is the commercial version of Huang’s 30x: it states the gain in the unit the buyer pays for. The framework was published in May 2026 and does not discuss Vera Rubin.

What Is Not Established

Huang did not give benchmark conditions for the stage 30x. The release’s 30x per megawatt names agentic workloads but no baseline system. The slide’s Vera Rubin result is marked as a preview. Stage remarks here come from an automated transcript and from TradeTheNews, not from an official transcript, so only phrases a second source shares are in quotation marks. Nvidia published no Vera Rubin production release on 21 October that we found, and gave no shipment volumes. The 24x figure is our arithmetic.

The Bottom Line

Huang told GTC Berlin that Vera Rubin is another 30 times beyond Blackwell, on a slide that measured it in tokens per dollar, while Nvidia’s release measured it as throughput per megawatt for agentic workloads. As we read it, the claim to watch is the cost of a token, because that is the figure buyers pay.