NVIDIA moves the AI factory yardstick to tokens per megawatt

Malaysia has spent two years approving data centres. The harder question was always where the electricity comes from. On 15 September 2026, at the AI Infra Summit in Santa Clara, NVIDIA made that question its central sales argument.[1][4]

Ian Buck, the company's vice president of hyperscale and high-performance computing, told an audience NVIDIA put above 8,000 people that AI factories must now be codesigned from silicon to grid. The measure he offered was tokens per megawatt.[1]

The company's own post states that DSX MaxLPS, its factory-wide power management software, can deliver up to 1.4 times more tokens per megawatt. For the coming Vera Rubin NVL72 systems it claims up to 40 per cent more GPU capacity inside the same site power envelope, and up to 35 per cent higher token throughput with no new power lines.[1]

The one deployment figure comes from Lambda, which NVIDIA says ran 19 nodes within the power budget normally allocated to 16, lifting cluster-wide throughput by 24 per cent, from about four million to five million tokens per second, and performance per watt by 23 per cent. Lambda's president of cloud services calls it a proof of concept that reclaims stranded capacity.[1][2]

Two claims sit on third-party ground. MLCommons published MLPerf Inference v6.1 on 16 September 2026, describing the suite as architecture-neutral with every result peer-reviewed before publication. It added an end-to-end retrieval-augmented generation test and an edge agentic inference test, set a record for submitting organisations, and reported gains of up to 5.7 times against results from a year earlier.[3]

NVIDIA reads those results as Vera Rubin NVL72 delivering up to 3.7 times the throughput of GB300 NVL72. That entry was a preview submission, which is not a shipping product measured in production. NVIDIA separately cites the SemiAnalysis AgentX dashboard for up to 30 times higher throughput per megawatt and up to 45 times lower cost per million tokens.[1]

MIDA set out the local arithmetic in March 2025. Data centre energy consumption in Malaysia could pass 5,000 megawatts by 2035, about 40 per cent of Peninsular Malaysia's current power capacity, with total energy supply applications already above 11,000 megawatts. The same article records RM184.7 billion of data-centre-related investment between 2021 and December 2024, concentrated in Greater Kuala Lumpur and Johor.[4]

Efficiency cuts both ways. NVIDIA's own framing is that throughput per megawatt decides how much revenue a power-constrained AI factory can generate, so an operator who gains capacity inside the same envelope is likelier to sell it than bank it. The national plan anticipates that, proposing fallback clauses when a data centre misses efficiency or renewable targets, and GPU-hour credits for smaller firms.[1][5]