SemiAnalysis: NVIDIA's Vera Rubin AI platform delivers 7x more tokens per megawatt than its predecessor Blackwell

NVIDIA's next-generation AI accelerator platform, Vera Rubin NVL72, runs agentic AI workloads up to 7x more efficiently per megawatt than its predecessor Blackwell, according to the first third-party benchmark SemiAnalysis published this week on pre-release software. The 7x tokens-per-megawatt number, not the 67x performance-per-dollar figure in the report's headline, is the one Google Cloud, Microsoft Azure, Oracle, and Meta will price their next inference capacity against.

Agentic AI is shorthand for AI assistants that take multi-step actions on a user's behalf. They browse the web, run code, query databases, and chain tool calls together, rather than answer a single question. That workload pattern is long-context, low-batch, and bursty, which means the binding constraint on how many agent sessions a cloud can serve is no longer silicon but data-center power and cooling. Tokens per megawatt is the procurement input that follows.

The benchmark itself is called AgentX. It replays real-world agentic traffic across SemiAnalysis's fleet of thousands of chips and has been validated by the four hyperscalers named above and by the open-source ML stack (vLLM, LMCache, SGLang, PyTorch, and Hugging Face). OpenAI, MiniMax, ZAI, Qwen, and Moonshot's Kimi have also signed on. SemiAnalysis says AgentX is currently the only inference benchmark covering TPU v7, NVIDIA, and AMD, with SambaNova and Trainium in the pipeline.

Vera Rubin is the first NVIDIA platform co-designed across six products for the agentic era: the Rubin GPU, the Vera CPU, the NVLink 6 Switch, ConnectX-9, BlueField-4, and Spectrum-6. At GTC 2026 in March, Jensen Huang presented that VR NVL72 achieved 3x performance per megawatt compared to Blackwell on trillion-parameter-scale models at roughly 200 tokens per second. SemiAnalysis now measures 7x on early software, and frames the gap as the chip vendor understating its own number. "Jensen is sandbagging," the analyst wrote, in the reporter's framing, not as an NVIDIA claim.

At GTC 2024, Huang claimed GB200 NVL72 would deliver 30x the performance of Hopper. SemiAnalysis later measured it at 98x. If Rubin follows the same software-maturation curve from announcement to shipping, the 3x-to-7x GTC-to-AgentX gap is the floor of the pattern, not the ceiling.

The 67x performance per dollar is a different number. It folds together the 7x per-megawatt efficiency, the rack-scale integration of the NVL72 form factor, and the dollars the workload spends on both silicon and electricity. That is the figure that will land in procurement spreadsheets, and it is also the most assumption-heavy: it presumes the pre-release kernel and library work SemiAnalysis ran is representative of what hyperscalers will deploy, and that the software stack keeps maturing at the Blackwell rate.

A third number, largely skipped by wire coverage, makes the strategic picture sharper. SemiAnalysis estimates that even on the early software builds, Rubin can earn over 2x more profit per gigawatt than the Blackwell platform. Profit per gigawatt is the metric that tells a cloud operator how much revenue each new watt of data-center capacity clears. Doubling it on pre-release software is the part of the report that lands hardest on AMD, whose MI400-class accelerators are still pitching on dollars-per-training-FLOP, the unit of value that defined the 2024 chip cycle.

Through 2024, the procurement question was how many H100s fit in a rack at what price per FLOP. Through 2025, it moved to total cost of ownership per token for batched inference. With Rubin on the AgentX workload, it is now tokens per megawatt of data-center capacity, because agentic traffic is bursty, long-context, and rarely batchable, and the binding constraint on how many agents a cloud can serve is no longer silicon but power.

The falsifier sits in the small print. Every number in the report comes from pre-release software, and SemiAnalysis is currently the only firm with visibility into it. If NVIDIA's software stack plateaus, the 7x-to-67x range compresses and AMD's MI400 finds its opening. If the stack matures at the Blackwell rate, the agentic inference market just got repriced, and the buyers who locked in 2025 capacity before Rubin ships will be the ones paying for it.