xAI's 8 Jet Engines Show AI's Real Bottleneck: Power

The week's real story: AI can't fit on the power grid anymore

The eight jet engines xAI installed at its Mississippi data center are the week's most important story, and the least discussed.

Running a data center on aircraft engines sounds absurd. But read xAI's Mississippi move and the logic is clear: eight refurbished jet engines, each capable of up to 50 MW, are there to skip a grid connection queue that stretches for years. The issue isn't fuel. It's time.

That marks a shift in the AI race. Training a better model no longer comes down to buying more GPUs; you have to find the electricity to feed them. In the US, data center interconnection queues are measured in years, so jet engines are a stopgap. A stopgap that shows how desperate the industry has become.

From Europe, this can look like a distant American problem. It isn't. Any country with weak grid capacity will hit the same wall on AI investment. The jet-engine fix is a warning, not a template.

Nvidia's 51% rule: selling chips is now politics

Nvidia has verbally told foreign cloud providers that at least 51% of their revenue must come from outside China to get chip allocations.

The second big story of the week was Nvidia's 51% rule. It isn't a written policy, just a verbal condition, but its effect is larger than many written rules.

In practice, it makes it nearly impossible for Chinese AI companies to rent Nvidia chips from clouds abroad. That opens a door for China's own chipmakers, Huawei first among them. It also accelerates the split of the global AI market into two camps.

The balance is hard for Nvidia: China is a huge market, but US regulators keep pushing. The company has to decide where its products go while trying to keep both sides happy.

Security: three separate breaches in one week

A reported Claude Code email leak, a Firecracker microVM escape and Bristol school files on the dark web all landed in the same week.

Security news came in waves. First, a report claimed Claude Code was sending users' email addresses to third-party sites through the User-Agent header. According to a GitHub issue, users started moving their companies to other tools once they noticed. It hasn't been officially confirmed, but the seriousness is clear: a coding assistant is carrying your identity out with it when it searches the web.

Then came the Firecracker microVM escape. PwnEscape Part 2 shows how a malicious guest can find the Firecracker process in host memory and hijack the registers of a live vCPU thread on the KVM_RUN return path. It's a technique that shakes cloud providers' isolation assumptions.

And 320,000 files stolen from 39 Bristol schools surfaced on the dark web. The council warned parents and staff about phishing attacks. The same week, Harvard confirmed 851,000 accounts were exposed in a ShinyHunters breach.

The common thread: attacks no longer target only big companies. Schools, universities and everyday tools are in the crosshairs, and in many cases users never find out.

Tesla, SpaceX and the market: Musk's political risk is back

Democratic wins in Texas and Congress raised fears of tighter regulation for Musk's companies; Tesla fell 10.77% and SpaceX 9%.

This week Tesla and SpaceX lost $469 billion in market value in two days. The reason is simple: Democratic wins in the midterms in Texas and Congress raised concerns about regulation of Musk's AI, data center and merger plans.

The same week brought news that Cybertruck production will end in mid-2027. The Austin line will be turned over to the Cybercab. US sales fell 48% last year to 20,237. This is the end of a strategy, not just a product.

The real question for Musk: how fast do AI and data center investments react when the political climate changes? This week the answer was clear: very fast.

Open models and games: the rest of the week

Baidu's 280-billion-parameter open model and an Ocarina of Time remake on Switch 2 were the other notable moves.

Baidu released ERNIE 5.5 Flash Base, a 280-billion-parameter model, with open weights. It uses 14 billion active parameters per token and was trained from scratch on 33.5 trillion tokens. That it landed the same week as Nvidia's 51% rule is no coincidence: as China struggles to source chips abroad, it is opening up its own models to build an ecosystem.

On the gaming side, the Ocarina of Time remake launched on Switch 2. It costs $60 digitally and $70 at retail, and stock is moving fast. Arriving the same week as the Zelda 40th Anniversary console shows Nintendo still knows how to play the nostalgia card.

What it means for you

For US, UK and EU readers, the xAI jet-engine story is a preview of what AI buildout will look like near you: data centers competing with households and businesses for grid capacity, and operators looking for workarounds. If you live near a major data center hub, expect more local debate over power, water and noise. Nvidia's 51% rule doesn't apply directly to European cloud providers, but any tightening of global chip allocation can show up in prices and lead times.

The security stories are the practical ones. If your company uses an AI coding assistant, ask what data it sends to third parties. If you run workloads on shared cloud infrastructure, the Firecracker research is a reminder that isolation is a promise, not a guarantee. And the Bristol and Harvard breaches show that schools and universities are now prime targets, so phishing awareness matters more than ever.

Watch two things next week: how regulators and environmental groups respond to xAI's jet engines, and whether Nvidia's 51% rule becomes written policy. The one thing already clear is that the AI race is now a race for electricity.