Daily Briefing — March 23, 2026

Europe wants a piece of the AI compute gold rush, and the bottleneck is not chips, not talent, not regulation. It is literally the wires in the ground. Grid operators across the continent cannot move electricity fast enough to feed the data centers that AI labs are desperate to build. National Grid in England and Wales has over 30 gigawatts of proposed data center demand sitting in a connection queue. That is roughly two thirds of Great Britain's entire peak electricity demand, just waiting in line.

The problem is not that Europe lacks the power generation. Renewables are coming online at a solid clip. The problem is transmission infrastructure, which takes years and serious capital to upgrade. So while AI spending globally runs into the hundreds of billions, some European data center projects are simply collapsing because they cannot get a grid connection approved in time to matter.

Grid operators are now scrambling to squeeze more capacity out of existing networks, experimenting with things like higher performance cables and smarter load balancing. That is interesting engineering, but it is also a sign of how far behind infrastructure planning has fallen relative to the pace of AI demand. The gap between where the industry wants to go and what the physical world can support right now is genuinely large.

Someone compromised Trivy, one of the most widely used vulnerability scanners in software development, and the attack was surgical. Hackers used stolen credentials to force push malicious code across virtually all versions of the trivy action tags on GitHub, essentially hijacking the tool that thousands of teams rely on to find security problems in their own code. The irony is almost too much to take: the thing you use to check if your pipeline is safe just became the attack vector.

The malware does not sit quietly. Once triggered, it actively hunts through your development environment for GitHub tokens, cloud credentials, SSH keys, and Kubernetes tokens. It encrypts what it finds and presumably ships it somewhere useful to the people who planted it. If your CI/CD pipeline ran a compromised version between Thursday and when this was caught, you should assume anything that pipeline touched is exposed.

This is a textbook supply chain attack, and it is a reminder that the tools you trust to protect you are themselves part of your attack surface. Aqua Security's maintainer confirmed the compromise and the advice is blunt: rotate everything. Not tomorrow. Now.

Jensen Huang went on the Lex Fridman podcast and dropped what is probably the most loaded three words in tech right now: "I think we've achieved AGI." To be fair, the definition Fridman was working with was pretty specific. He framed AGI as an AI system that can essentially do your job, meaning start, grow, and run a successful tech company worth over a billion dollars. Huang heard that and said, basically, yeah, we're there now.

Here is the thing though. AGI means whatever the person saying it needs it to mean at that moment. Tech leaders have spent months running away from the term, coining new phrases that sound more precise but are functionally identical. The reason that matters beyond semantics is that AGI is literally written into major contracts, including the one between OpenAI and Microsoft, with real money tied to when or whether it gets declared.

Huang also pointed to OpenClaw, an open source AI agent platform that went viral, as evidence that agents are already doing things people did not expect. Whether or not you buy the AGI framing, the underlying point lands: agentic AI is moving faster than most job descriptions have accounted for.

Everyone has seen the demo. The AI agent books the meeting, summarizes the document, routes the ticket, handles the edge case. It looks great. It always looks great. Then someone tries to actually deploy it inside a real company and the whole thing turns into a slow-motion disaster.

That is what enterprises are learning right now. The gap between a controlled demo and a live production environment is not a technical gap so much as an organizational one. Fragmented data sitting in seventeen different systems, workflows nobody has properly documented, and escalation rates that spiral because the agent keeps hitting situations it was never trained to handle. The technology works. The organization is not ready for it.

Greyhound Research analyst Sanchit Vir Gogia put it plainly: the challenge starts when agents are asked to operate inside the complexity of a real organization. That complexity is exactly what demos are designed to hide. If your company is in the middle of an AI agent rollout, or planning one, you are likely about to meet that complexity face to face.

The enterprises that are actually succeeding with agentic deployments are doing three things differently: cleaning up their data pipelines before launch, mapping workflows with embarrassing granularity, and building clear escalation rules from day one. None of that is glamorous work. All of it is the actual work.

AI is moving into HR, and this time it is not just sending automated rejection emails. We are talking about systems that screen your resume before a human ever sees it, flag you as a flight risk before you have even updated your LinkedIn, and recommend your next internal role based on pattern matching across thousands of other employees. A new wave of startups and enterprise platforms are pitching this as the solution to slow, inconsistent, and administratively bloated HR functions. And honestly, the pitch is not wrong on its face.

But here is the part that should make you pause. HR is not accounts payable. It is the function that decides whether you get hired, promoted, supported, or quietly managed out. The CEO of HR tech unicorn Phenom put it plainly: the bar for trust here is higher than in other enterprise functions because the decisions directly affect people's lives and identities. Algorithms optimizing for consistency can just as easily lock in existing biases at scale, and do it faster than anyone can catch.

The deeper shift is that the definition of what HR professionals actually do is being rewritten in real time. If AI is screening candidates and predicting attrition, the humans left in that function need to be doing something the machine genuinely cannot do. That is a very different job description than the one most people in HR signed up for.