Three major stories are reshaping the AI landscape this week: Stanford's annual AI Index puts hard numbers on the global race, Meta goes all-in with a new superintelligence lab and a nine-figure capex pledge, and NVIDIA opens its robotics stack to the world.

Stanford AI Index 2026: Anthropic on Top, China Closing Fast

Stanford's AI Index 2026 is out, and the headline finding is stark: Anthropic's models now lead overall performance rankings, edging out OpenAI and Google on a composite of reasoning, coding, and safety benchmarks. But the more consequential number may be geopolitical — the report finds China has "nearly erased" the US lead in frontier model capability, closing a gap that was measured in years just two years ago.

Infrastructure tells the same story. Global data center power demand has hit 29.6 GW, more than doubling since 2022, as the compute arms race shows no sign of slowing. The Index also notes that AI is now a net job creator in most OECD economies — a finding that runs counter to the dominant public narrative.

Meta Bets $115–135B and Launches a Superintelligence Lab

Meta used this week to make two big swings. First, it introduced Muse Spark, the first model out of its newly formed Meta Superintelligence Labs — a reorganised unit designed to accelerate the path to AGI-class systems. Second, it confirmed a $115–135 billion capital expenditure commitment for 2026, the largest single-year AI infrastructure bet any company has publicly announced.

Meanwhile, OpenAI reported crossing $25 billion in annualised revenue and is reportedly taking early steps toward a public listing as soon as late 2026 — signalling that the race to monetise frontier AI is now as intense as the race to build it.

NVIDIA Opens Its Robotics Stack

NVIDIA's National Robotics Week announcements were quietly significant. The company released Isaac GR00T open models for humanoid robot training, made its Cosmos world models publicly available for simulation, and launched Newton 1.0, a physics engine built for robot learning environments.

Taken together, the releases amount to an open-source robotics platform that any team can now build on — lowering the barrier to physical AI in the same way that Hugging Face lowered the barrier to language models.

What to Watch

Stanford's finding that China has nearly closed the capability gap will land hard in Washington. Expect renewed pressure on export controls and compute access policy in the coming weeks. On the commercial side, Meta's capex number sets a new benchmark — watch whether Microsoft, Google, and Amazon feel compelled to match it in their next earnings calls. And with NVIDIA's robotics stack now open, the question shifts from "who can build a robot brain" to "who can build the best one fastest."