AI's Bottleneck Just Moved, and the Checks Followed It
Last week's biggest rounds weren't about what AI can do. They were about whether it works when it matters.
Look at where the money went last Thursday and you’ll see a quiet shift. Runpod raised $100 million at a $1 billion valuation, with more than a million developers on its AI cloud. Sail Research pulled $80 million for infrastructure that keeps AI agents running over long horizons. Scaled Cognition took $100 million for AI reliability. Trase raised a $107 million seed for healthcare automation. Different companies, one theme: nobody’s funding “can the model talk” anymore. They’re funding “does it hold up in production.”
That’s not a vibe. It’s a constraint moving.
The mental model: Theory of Constraints
Eli Goldratt’s idea is simple and unforgiving. Every system has exactly one bottleneck that caps its output at any given moment. Work on anything except that bottleneck and you get nothing. Throughput doesn’t improve until you fix the actual constraint. And here’s the part people miss: once you fix it, the constraint moves somewhere else, and the job starts over.
Watch how it played out in AI. In 2023 the constraint was raw capability. Could the model produce something coherent? Solved. In 2024 it moved to access. Could you wrap an API into an app fast enough? That era minted the wrapper startups. By mid-2026 the constraint moved again, and it’s the hardest one yet: reliability. Can the system run in production, inside a regulated workflow, where a wrong answer costs real money?
So the capital followed. It’s flowing to reliability, observability, agent infrastructure, and regulated automation, because that’s the part breaking right now. Investors aren’t being clever here. They’re funding the bottleneck.
Spence’s take
The mistake builders make is pouring effort into the part that isn’t the constraint. A slicker interface on a system that hallucinates in production is polishing a non-bottleneck. It feels like progress and changes nothing. Find what breaks first, fix that, and leave the rest alone until it becomes the new limit.
Here’s the contrarian piece. The market keeps asking “what can AI do?” That’s last year’s constraint. The real question now is “what makes AI safe enough to actually ship?” And the answer is deeply unglamorous: reliability, compliance, the boring plumbing nobody posts about.
I see this every day. In regulated markets the bottleneck was never the idea. It’s whether you can run the thing inside the rules. That’s why /mkt builds on Reg A+ with tZERO’s infrastructure. When trust and compliance are the constraint, that’s where the work has to go, not the front end.
Find your bottleneck. Then have the discipline to work only on that.
If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.
This post is for informational and educational purposes only. It is not investment advice, an offer to sell, or a solicitation of an offer to buy any security. References to specific companies, funding rounds, or platforms are illustrative and not recommendations. Securities offerings under Regulation A+ involve risk, including possible loss of principal, and are qualified by their official offering documents. Nothing here should be relied on for any investment decision. Consult a licensed professional before acting. Funding figures are drawn from publicly reported sources including Crunchbase and TechCrunch and reflect reporting at the time of writing; they may be revised.