Nvidia Rubin Gets Up to 3.2x GB300’s Tokens per Dollar

Nvidia’s Vera Rubin NVL72 produced between 2.37 and 3.24 times as many Kimi K3 tokens per dollar of total cost of ownership as the GB300 NVL72, according to a SemiAnalysis InferenceX chart published on 27 October that labels the Rubin results an official preview.

SemiAnalysis wrote on X that Nvidia “just submitted its first Rubin NVL72 Kimi K3 submission, and it can reach over 850 tok/s/user, even at 2.8 Trillion parameters”, adding that it had “over 3x higher throughput” on vLLM, the open-source serving software.

Short explainer · Rubin on Kimi K3

A vertical news Short narrated from the published text of this piece and related FourWeekMBA pieces.

Business Pill · Tokens per watt explained

An evergreen Business Engineer explainer on tokens per watt, what an AI data centre actually produces. It explains the general idea and does not discuss Rubin or Kimi K3.

What the Chart Shows

The chart is titled “Kimi K3 2.8T Agentic Total Tokens per $1 TCO vs. P90 Interactivity” and is dated 2026-10-27. It compares two systems, both running Nvidia’s Dynamo with vLLM: Vera Rubin NVL72 and GB300 NVL72.

Tok/s/user, the unit SemiAnalysis quotes, is how many tokens per second a single user receives, in other words how fast an answer streams to one person. The chart measures it at P90, and puts total tokens per $1 of cost on the other axis.

At four matched speeds the chart prints the Rubin multiple over GB300: 2.37x at about 100 tok/s/user, 2.38x at about 150, 3.24x at about 280 and 3.07x at about 400, reading the speeds from the chart’s axis.

The Rubin curve runs to about 860 tok/s/user, against about 425 for the last GB300 point, also read from the axis. On both curves, faster service per user comes with fewer total tokens per dollar.

A Pricier Chip-Hour, More Tokens

SemiAnalysis prices a VR200 chip at $3.61 per hour and a GB300 at $2.31, under a cost tier labelled “Owning at Large Hyperscaler Volume” and its July 2026 pricing surveys and AI Cloud TCO Model. By our arithmetic, the Rubin chip-hour costs about 56% more.

The tokens-per-dollar multiples on the chart already include that higher cost. The same InferenceX page lists chip-hour costs of $2.26 for B300, $1.50 for MI355X and $1.03 for TPU v7.

Connecting the Coverage

On 21 October we reported that Huang put Vera Rubin at 30x Blackwell on tokens per dollar, and on 11 October that vLLM on Vera Rubin posted 7.8x GB200 throughput per chip. Those figures come from separate charts and are not combined with today’s multiples.

On 26 October we reported that Nvidia cut Vera Rubin memory bandwidth to 19.2 TB/s per GPU, and today that Nvidia will invest $3B in Meta’s stock as Meta adds Rubin. On 15 October we covered SemiAnalysis saying Google TPUv7 tops GB300 by 1.67x on cost.

The Structural Read

As we read it, the chart prints its largest Rubin multiples at the faster speeds: 3.24x at about 280 tokens per second per user and 3.07x at about 400, against 2.37x and 2.38x at about 100 and 150.

As we read it, the chart also shows the trade-off inside each system: on both curves, more speed per user comes with fewer total tokens per dollar.

As we read it, SemiAnalysis frames the result as Nvidia’s first Rubin NVL72 submission on Kimi K3, a 2.8 trillion parameter model, run on Dynamo with vLLM on both systems.

Three Implications

FOR AI CHIP BUYERS SemiAnalysis prices a Rubin chip-hour about 56% above GB300 by our arithmetic, and its chart prints 2.37x to 3.24x the tokens per dollar.

FOR INFERENCE PROVIDERS SemiAnalysis’s chart puts Rubin’s top speed at about 860 tokens per second per user, against about 425 for GB300, read from its axis.

FOR BENCHMARK READERS SemiAnalysis labels the Rubin results an official preview and says they may change as validation and publication continue.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of the Inference Economy.

The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, SemiAnalysis’s chart prices that inference work as tokens per dollar of total cost of ownership, which includes the cost of owning the hardware. The framework was written before this result and does not discuss Rubin or Kimi K3.

What Is Not Established

The Rubin results carry the label “InferenceX Official Preview: Results may change as validation and publication continue.” The tokens-per-dollar multiples depend on SemiAnalysis’s own cost model, and the speeds quoted here are read from the chart’s axis rather than printed values.

The post does not say what the “over 3x higher throughput” is measured against, and the InferenceX page notes that disaggregated setups count throughput per decode or prefill chip, so they are not an apples-to-apples comparison with other setups. We did not see a statement from Nvidia.

The Bottom Line

On SemiAnalysis’s preview chart, Rubin costs more per chip-hour than GB300 yet returns 2.37 to 3.24 times the Kimi K3 tokens per dollar at matched speeds, and its curve extends to about twice GB300’s top per-user speed, by our arithmetic.