On August 24, NVIDIA released an early set of benchmark results for its Vera Rubin NVL72 system: on agentic-style workloads, per-megawatt throughput reaches 30 times that of the previous-generation GB300 NVL72, while per-token cost drops to roughly 1/35th. A line beneath the numbers notes that the results are still awaiting review from SemiAnalysis.
The benchmark ran on SemiAnalysis's AgentX workload, which is built from real agentic coding traces rather than synthetic requests generated in a lab. Five models were used for scoring: Kimi K3, MiniMax M3, GLM5.3, Qwen3.5, and DeepSeek V4 Pro — all open-weight, four of them from Chinese teams. The lineup itself says something: to judge whether a new generation of inference racks is actually good, NVIDIA chose to benchmark against the open models developers are actually running.
Why the metric is "per megawatt"
Traditional generation-over-generation comparisons tend to use peak compute or per-card performance. This time, NVIDIA swapped the denominator for electricity.
The reason lies in how the workload is structured. According to OpenRouter's usage data, a single agentic request consumes 15 times the tokens of an ordinary chat request — an agent has to query databases, retrieve documents, invoke sub-agents, and drag an ever-longer context through multiple rounds of reasoning. Once per-task token consumption jumps by an order of magnitude, what actually caps scaling is usually the megawatts a substation can deliver, not the GPU count.
NVIDIA's own point of vertical comparison: GB300 NVL72's per-megawatt throughput on DeepSeek V4 Pro was already 15 times that of the Hopper architecture. Vera Rubin multiplies that figure again.
Broken down, the 30x doesn't come from any single source. DSX MaxLPS, NVIDIA's power-delivery and cooling technology, packs 40% more GPUs into the same megawatt budget, which effectively thins out the denominator first; NVLink 6 pushes packet rate to 10 times that of off-the-shelf Ethernet solutions while cutting latency to a third; layered on top of that is a full hardware-software stack — disaggregated serving, distributed KV caching, and expert parallelism. Each of these contributes its own multiplier, and stacked together they add up to the final number.
A rough back-of-the-envelope calculation
Using NVIDIA's own figures, it's worth working out roughly what this multiplier means in real terms.
Assume a 100-megawatt inference cluster, with electricity priced at about ¥0.5 per kWh; running at full load for a year would cost roughly ¥440 million in power. If per-megawatt throughput really does rise 30x, that same power bill would buy 30 times the token output; looked at the other way, producing the same number of tokens would cost about 1/30th as much in electricity. As for the 35x drop in per-token cost — slightly higher than the 30x throughput gain — that gap comes from the rack's own efficiency improvements, which is a separate matter from the increase in total output.
This kind of rough math has obvious distortions: real data centers never run at full load for an entire year, and factors like PUE, idle capacity, and mixed model deployment all pull the multiplier down. But it does explain why NVIDIA puts this particular metric front and center. For data center operators who already have a power allocation but are still waiting in line for new substation capacity, output per megawatt is a lot closer to the concept of "production capacity" than total compute is.
What to keep in mind before the review lands
These numbers are currently self-reported by the vendor. SemiAnalysis's review hasn't concluded, and NVIDIA itself has labeled this an early measurement.
The qualifiers matter too. The 30x figure applies to agentic coding traces — a long-context, high-token-density scenario that happens to sit right in Vera Rubin's architectural sweet spot. Swap in short chat, image generation, or batch embeddings, and the multiplier won't look the same. A gap between cross-generation claims and what actually ships is an old pattern in this industry — the move from Hopper to Blackwell played out the same way.
"Agentic AI workloads consume 15x more tokens than a simple chat request."
Stretching the timeline out a bit makes the picture clearer: Dell is already quoting inference pricing on Vera Rubin servers, and SpaceX plans to send NVL72 racks into orbit next year to test AI computing in space. The hardware hasn't shipped at scale yet, and the math around it has already been run several times over. What will actually validate the 30x figure are the numbers each operator gets running in their own data centers, once independent review is complete.
Sources: NVIDIA's official technical blog, SemiAnalysis's AgentX benchmark documentation, CocoLoop, OpenRouter usage statistics; the per-megawatt throughput multiplier, per-token cost reduction, and 40% increase in GPU deployment density have all been checked against official figures, while the power-cost estimate is the editorial team's own rough calculation.