Open-source language models are now running at 207 tokens per second on consumer hardware. This isn't progress—it's the end of the AI capex story, and corporate balance sheets don't know it yet.
For eighteen months, we've watched tech companies justify $50M infrastructure spends on the premise that proprietary AI systems would generate moat-like competitive advantages. Claude, GPT-4, specialized enterprise tools—the thinking goes, you pay for the edge. But Qwen3.6-Max just proved that 90% of the way to state-of-the-art lives in open-source now. Hand-optimized kernels on a three-year-old RTX 3090. No cloud bill. No vendor lock-in.
What happens when enterprise customers discover they've funded the infrastructure for someone else's free alternative?
The precedent is brutal and immediate. This is Docker moment—when you realize the expensive thing you're paying for is now commodity. Except Docker eventually became valuable *because* it commoditized the base layer. AI's different. There's no Kubernetes here. There's just... cheaper inference.
The signal is already embedded in the data. HDFC, ICICI, Yes Bank—major financial institutions in India just reported Q4 earnings with *stagnant pricing power*. They can't raise margins anymore. In a rising-rate environment, that's the canary. Parallel signal: Rheinmetall just began series production of autonomous surface vessels. Sight Machine is shipping autonomous manufacturing agents. GitHub's fake-star economy reveals that developer enthusiasm metrics—the thing we've been using to validate AI infrastructure demand—are measurement games.
All of this points to the same structural problem: capital efficiency is breaking down because *the automation ROI hasn't materialized yet*, but companies have already committed the capex. They're chasing a productivity story that open-source LLMs just made thirty-percent cheaper. Margins stay compressed. Capex discipline either breaks (and rates spike), or it holds (and growth disappoints).
Here's what makes this dangerous: it's invisible to Wall Street consensus right now. The 10Y is at 4.26%, the curve is steep at 54bp positive. Markets are pricing in soft landing + stable yields + innovation-driven growth. That thesis *requires* companies to generate ROI on AI investments. If they don't—if Qwen at $0 marginal cost becomes the endgame—then the productivity story collapses and multiples re-rate lower.
And middle-market escalation in the Middle East (buried in PSEi calm, by the way) just adds energy-price tail risk on top. Yields spike to 5.2%, curve inverts, and companies that committed 18% of capex to unvalidated AI infrastructure suddenly look very bad.
The closest mirror: energy companies post-2014. Capex committed to high-cost extraction. Commodity crashed. Years of margin degradation followed.
Tim Cook's handoff to Ternus in September isn't a succession—it's liability transfer before the realization hits. Apple's China revenue is deteriorating. Services growth is slowing. And now the entire AI arms race they signed up for has a cheap, reproducible alternative.
If the board knew that in April, they moved Cook to the sidelines now.