Sunday, 31 May 2026

The valuation math has gotten absurd. Anthropic just crossed $965 billion on a $65 billion funding round, and the thing that strikes me isn't the number itself—it's that we're now pricing AI companies on faith in a future business model that barely exists yet. OpenAI is worth roughly the same. Neither has demonstrated sustainable unit economics at scale. By the way, this matters because it signals where capital thinks the winners will sit, and right now capital is betting that whoever controls the most capable frontier models wins everything downstream. The question is whether that's actually true.

That downstream layer is where the real fight is happening now, though. Both OpenAI and Anthropic just announced multi-agent automation platforms aimed at enterprise workflows, and I find this more interesting than either company's valuation. Long-running, complex tasks—the kind that require agents to reason across multiple steps, handle failures, and adjust strategy—are genuinely hard. They're also where real organizational value lives. If you can automate a process that currently costs a company millions in labor and error, the ROI calculus changes completely. The constraint isn't capability anymore; it's adoption and integration.

Here's where the technical progress becomes relevant again. MeMo, a memory architecture that lets teams upgrade their language models without full retraining and see a 26 percent performance jump, addresses a real pain point: the sunk cost of training and fine-tuning. If you've already optimized a model for your specific use case, retraining on a new base model is expensive and risky. Separating memory from reasoning could actually make the frontier-model treadmill bearable for enterprises.

The consumer side tells a different story. GitHub Copilot's switch to token-based billing has developers openly calling it a joke—the flat-rate option is disappearing, which means the calculus of "I'll use this for everything" evaporates. Meanwhile, Microsoft and Nvidia are reportedly building AI PCs that run actual agents instead of just Copilot, and Google is releasing Veo 3.1 Lite to democratize video generation through cost reduction rather than capability maximization. Each of these is a rational move, but together they sketch a fragmentation: premium models for enterprises willing to pay per token, cheaper alternatives for everyone else, and a race to own the hardware layer where agents actually run.

The pessimists—Yudkowsky and Soares, judging from their new book—are warning that continued capability improvement risks extinction. I'm skeptical of that frame, but I understand the anxiety. We're moving from models that respond to prompts to systems that autonomously manage workflows, make decisions, and interact with the world. The alignment problem doesn't get easier when the system acts without waiting for human approval.

What matters next isn't valuation or even individual breakthroughs. It's whether the enterprise automation layer actually works at scale, and whether the economic returns justify the infrastructure costs.