Turning dinosaurs into unicorns: AI and the real economy

Every venture cycle carries a myth about where the money goes. This cycle’s myth is “AI is eating software.” The real story is that AI is eating labor, and it’s about to eat the most of it in the physical economy. AI absorbed roughly 80% of global venture funding in Q1 2026, $242 billion of $300 billion. OpenAI and Anthropic alone took 43% of everything venture invested globally in H1.

That kind of concentration cuts two ways. It’s the shape of a bubble right before it corrects, and it’s what happens when a frontier stops being a startup category and becomes infrastructure. Capital is moving one layer down, into the systems underneath intelligence (power, cooling, data centers, grid access) and one layer out, into the physical world intelligence is starting to touch.

Hardware didn’t suddenly get easy. Atoms still cost money, and factories, components and maintenance still matter. What changed is the distribution point. A machine used to leave the factory as the product it would remain. Now a deployed fleet keeps improving after the sale: new models sharpen perception, planning and manipulation across an installed base, while deployment itself generates the proprietary data that makes the next model better.

That collapses the old math. TAM stops being machines sold x hardware price, and becomes installed machines x intelligence distributed x labor value captured over time. It matters most where the physical world is already built around human capability: doors, shelves, tools, production lines. A general-purpose robot doesn’t just compete for a robotics budget; it competes for a slice of the labor market. A copilot sells a tool, an autopilot sells the finished job.

This is where “dinosaur” companies from the least digitized industries come in: manufacturing, logistics, construction, energy, healthcare. They were never priced as software markets, because they were never sold as software — they were priced as labor markets. For every dollar businesses spend on software, they spend roughly six on services. That six-to-one ratio is the real prize, and it’s why physical AI can rebuild industries software never touched.

One thing will decide who captures that value: how concentrated the intelligence layer stays. Frontier training keeps consolidating toward a handful of labs with the capital and compute to compete. Deployment doesn’t have to follow the same path. Open-weight models are improving quickly, and a factory that can’t wait on the cloud, or won’t risk its production data leaving the building, has real reasons to want intelligence it can run locally.

General-purpose AI is becoming a basic productive input, alongside capital, labor and energy. If access to it stays as concentrated as this year’s funding numbers suggest, so does the ability to capture what it produces. How that intelligence gets distributed - held behind a few platforms, or spread across the companies and machines that actually deploy it - will decide which old industries get rebuilt, and who ends up owning what they become. This is what I call turning dinosaurs into unicorns.