Monday, 25 May 2026

The spring of 2026 is revealing something uncomfortable about enterprise AI adoption: the gap between what we hoped these tools would do and what they actually cost to run.

Start with OpenAI's Codex. The company has quietly transformed it from a code completion engine into a genuine desktop agent that can watch your screen, control applications on Mac, and run on mobile devices. This is meaningful progress—not because agents are new, but because putting them in the hands of everyday developers and knowledge workers at scale changes the economics of the problem. A tool that can actually click through your apps and execute sequences is fundamentally different from one that generates code snippets. By the way, this matters for the broader industry conversation about whether agents will be tools or workers. If Codex can handle real workflows on commodity hardware, then suddenly the question isn't "will agents work?" but "how much will we actually use them?"

Which brings us to the harder story. Microsoft has started cancelling most of its direct Claude Code licenses for employees and redirecting developers to GitHub Copilot instead. The reason, reported matter-of-factly across multiple sources this week, is that Claude Code costs more to run than it saves. Uber is facing similar arithmetic. This is the moment where the hype collides with actual spreadsheets, and it's worth sitting with: companies spent months deploying cutting-edge AI tools only to discover they were paying more for the compute than they were gaining in productivity. That's not a minor implementation detail—it's a fundamental validation failure. I find this honestly more important than any new model release, because it forces us to ask whether the baseline unit economics of these systems actually work outside the lab.

The other thread running through this week is control and compliance. Nvidia's Jensen Huang arrived in Taiwan urging Supermicro to fix its export compliance after a $2.5 billion smuggling bust earlier this year. Taiwan is also cracking down on AI chip smuggling to China. Zscaler is acquiring Symmetry Systems to handle AI agent governance—specifically, to map identity and access for AI systems. These moves reflect a quieter recognition: as AI systems become agents that control access and execute workflows, the security and compliance surface explodes. You can't just deploy an agent and hope it stays in its lane.

What strikes me about this particular moment is that the real problems aren't technical anymore. OpenAI can build agents. Figure AI's humanoids sorted packages for 24 hours without intervention. These are solved problems in isolation. What's breaking is the economics, the governance, the actual integration into work. That's where the real difficulty lives.