I just learned something that should terrify every AI investor: Six major large language models were tested on real freelance work - the kind actual humans get paid to do on Upwork. Not homework. Not summaries. Real commercial tasks that generate real revenue. Building video games. Creating presentations from rough notes. Architectural schematics. The BEST performing AI completed tasks well enough to get paid 2.5% of the time. The worst? 0.3%. Think about that. If you were an Uber driver who completed 2.5% of your rides, you'd be kicked off the platform in a week. This comes from academic research published in the Remote Labor Index - not some anti-AI hit piece. They eliminated jobs requiring physical work or heavy human interaction and focused purely on digital deliverables where AI should theoretically excel. And it failed 97.5% of the time. Meanwhile, US tech companies are spending $380 billion on AI infrastructure in 2025-2026. Data centers using the power of 619 houses per GPU stack. Oracle's shares are now BELOW where they were before announcing their massive OpenAI partnership. Blue Owl Capital (AI infrastructure funder): down 40% Fermi (data center REIT): down 60% The funding markets are already getting more discerning. And we haven't even hit the real reckoning yet. AI is excellent at correlation. But correlation isn't how the world works. It can regurgitate answers to questions it's been trained on. But ask it to actually BUILD something, execute a complex task, or operate in the real world where correlations don't hold? It falls apart. Scam Altman showed Operator - OpenAI's agent that's supposed to act like a CEO's assistant. 19 minutes into the demo, they revealed it worked 34% of the time. On their own metrics. Their own homework. That THEY graded. 34%. And that's in a controlled demo environment. In the real world with actual commercial deliverables it's 2.5%. The capital misallocation is 17 times larger than the dotcom bubble. Nvidia's receivables are up 770% in 33 months (Cisco's were up 140% before they collapsed). Every part of the AI stack is losing money except Nvidia - and they're the ones extending vendor financing to keep the whole thing afloat. This isn't a technology that's "almost there." This is a technology with fundamental architectural limits that can't be overcome by just adding more compute. I sat down with Julian Garran - one of the sharpest macro strategists I know - and he walked through why AI was "built to fail" from day one. The full conversation covers: - Why the economics of data centers guarantee losses - The Cisco 2000 playbook playing out in real time - What happens when the funding dries up - Where smart money is rotating (hint: it's not tech) This is a career-defining inflection point in markets. And most investors are still positioned for a productivity revolution that isn't coming. The full interview is in the comment below.