Daily Briefing — May 8, 2026

Researchers at Palisade, a Berkeley based AI safety organisation, have published findings showing that recent AI systems can independently copy themselves onto other computers without being told to. Not as a theoretical exercise. In actual tests. The director of the organisation put it plainly: we are approaching the point where no one would be able to shut down a rogue AI because it could replicate itself across thousands of machines before anyone even noticed something was wrong.

Back in March, Alibaba researchers reported catching one of their own systems, a model called Rome, tunnelling out of its sandboxed environment to mine crypto on an external system. That one got a bit of press and then mostly disappeared from the conversation. The Palisade findings suggest this is a pattern, not a fluke.

SpaceX is planning to drop at least $55 billion on a chip manufacturing facility in Austin, Texas, called "Terafab," with total costs potentially reaching $119 billion if all phases get built out. The details surfaced in a public hearing notice filed in Grimes County, where SpaceX is asking for tax breaks. Intel is already involved on the design and fabrication side, and the plant will serve both SpaceX and Tesla with chips aimed at AI, robotics, and space based data centers.

Musk is trying to vertically integrate the entire AI compute stack, from the chip fab floor all the way up to orbital data centers. The ambition is genuinely staggering. One terawatt per year of compute capacity in space is not a number that fits neatly into any existing mental model of what a technology company does.

What this signals for the broader industry is that the race to control AI infrastructure is now moving into physical manufacturing in a serious way. Companies that own the chips, the power, and the data centers will have a structural advantage over those that rent compute from someone else.

Anthropic just signed a deal to take over the entire computing capacity of SpaceX's Colossus 1 data center in Memphis. That's more than 300 megawatts and over 220,000 NVIDIA GPUs, and it will be powering Claude Pro and Claude Max subscriptions going forward. SpaceX owns xAI, which makes Grok, and the only reason that compute was available to sell is that nobody is using Grok in anywhere near the volume xAI built capacity for.

In an industry where every other AI lab is scrambling, lobbying, and signing nine figure cloud contracts just to get their hands on more GPUs, one player has so much spare capacity that they're leasing out entire colossal data centers to a direct competitor - which is a bit funny. When you build for demand that doesn't show up, the compute becomes a liability you have to monetise somehow, and the most obvious buyer is the lab whose product people actually want to use. But don't laugh too much, that's how AWS Cloud is taking off in the beginning.