AI IPOs: Who Controls the "Harness"?

Summary

The upcoming IPOs of OpenAI and Anthropic hinge not just on their valuation, but on a deeper question: can they make AI cheap enough to use at massive scale while building proprietary "harnesses" to rent out the entire system? A 'token' is raw AI intelligence, while a 'harness' is everything that makes that intelligence useful – like tools, memory, and workflows. Companies like CodeX and Claude Code are examples of harnesses. The real bet is whether OpenAI and Anthropic can own this "work layer" above the models. While some analyses suggest users get far more API value than they pay for ($14,000 for a $200 plan from OpenAI, $8,000 from Anthropic), this likely reflects retail pricing with markups, not internal costs. If these companies can significantly lower their actual serving costs through inference efficiency and model routing, the low API prices could be a strategic subsidy to drive adoption while they race down the cost curve. The core idea is that as raw intelligence (tokens) becomes cheap and abundant, value shifts to what's built around it – the harness. OpenAI and Anthropic aim to sell this operating layer, not just raw intelligence, making their products sticky and difficult to switch from, even if models change. The challenge for them is overcoming a company's inherent advantage: private context. To bridge this, they employ "forward deployed engineering" to understand and adapt their harnesses to specific customer workflows. For businesses, the strategic question is whether they are renting a harness or owning it. Owning it means controlling the layer that decides which model to use, managing context, evaluations, and workflows, effectively turning AI labs into suppliers. The bull case for these IPOs relies on them managing token costs, competing with open-source models, and building superior harnesses that companies opt to rent. The bear case is that companies will build their own harnesses, leaving the labs with only token margins. When S1 filings are released, investors should look for indicators like improving gross margins with usage, customers buying scalable software over labor, and whether forward deployed engineering is a permanent need or a bridge to self-sufficient products. Ultimately, the key for individuals and companies is to understand that using AI tools is not the same as having an AI strategy; the valuable skill is harness building – defining recurring jobs, providing context, connecting tools, and checking outputs.