Two infrastructure stories landed four days apart this week, and they pull in opposite directions. On November 16, NVIDIA's Vera Rubin platform made its MLPerf debut, sweeping every MLPerf Training v6.1 benchmark and posting, in NVIDIA's words, more than 10x generational gains in throughput per watt for agentic AI S1. On November 20, President Trump promised a "tremendous tax" on companies that build data centers outside the US, even as states keep adding restrictions of their own S2. The chips keep getting more efficient. The map of acceptable places to plug them in keeps shrinking.

Read the debut with its conditions attached

The Vera Rubin numbers are vendor-submitted, so the fine print does work. "Up to 2x more performance per GPU" than a highly optimized GB300 NVL72 is a ceiling across workloads, not a uniform doubling. The figure with real planning value is scaling: on the DeepSeek-V3 benchmark, Vera Rubin NVL72 went from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to about four. That is a claim about networking and software staying coherent as racks multiply — what decides whether a purchased cluster delivers four times the work. NVIDIA also ran 10,240 GB300 GPUs, the round's largest scale.

What the per-watt gain buys depends on what you hold fixed. For a fixed workload, higher throughput per watt genuinely means less energy consumed. For an operator filling a fixed power envelope, the same gain shows up as headroom — more agentic work per watt — rather than a smaller draw; that is the reading NVIDIA emphasizes, arguing that because AI data centers are limited by power, tokens per watt decide how much agent work a system delivers [S1]. The supporting Endpoints figure is provisional: for Kimi K3, Vera Rubin NVL72 posted more than 10x the throughput per watt of GB300 at the same response speed.

The siting map tightens from both directions

Trump's threat came at a press conference following his summit with Xi Jinping and Vladimir Putin, floated publicly for the first time — "I haven't even said that to my people" — with few details, and a warning that otherwise "China will take them all." The Verge reports the promise could be difficult to fulfill with a new Democratic majority Congress being sworn in within a couple of months.

The domestic map is already closing, as reporting from September documents: Moody's counted at least 14 US states considering statewide moratoriums; New York had paused environmental permits for data centers of 50MW or more, Pennsylvania demanded local approval above 25MW, and Texas ordered an audit of an interconnection queue that had reached 474GW, more than five times its record peak demand S3. Read together, local rules push buildout across borders while the tax threat raises the price of going. Siting now carries political risk on top of power price and queue time.

What planners should do this quarter

Size clusters from scaling efficiency and same-GPU-count comparisons, not headline peaks, and track which results are still provisional. Treat per-watt gains as workload-dependent: fixed workloads consume less energy; fixed interconnection caps buy serving headroom. Neither moves the limit itself. Remember that queues set timelines — nearly 12GW of data center load requests sat in New York's interconnection queue as of May, and no benchmark doubling shortens that. Finally, hold siting optionality across jurisdictions, because one election or executive order can flip the rules a roadmap depends on.

Faster chips are the easy half of AI infrastructure. The hard half is still a socket on a grid that someone is willing to approve.

Sources and editorial method

This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.

Common questions

What did NVIDIA report in Vera Rubin's first MLPerf results?

On November 16, 2026, NVIDIA reported that Vera Rubin swept every MLPerf Training v6.1 benchmark, delivered up to 2x more performance per GPU than GB300 NVL72, scaled from 256 to 1,024 GPUs at about 97% efficiency on DeepSeek-V3, and posted a provisional 10x-plus throughput-per-watt gain for agentic AI in its first MLPerf Endpoints submission.

What did Trump propose about data centers built outside the US?

At a November 20 press conference, Trump said the US would not allow companies to build data centers outside the country and threatened a 'tremendous tax' on those that do, offering few details while states impose their own moratoriums and restrictions on new data center projects.