The true value behind OpenAI and Anthropic’s IPOs lies not in their AI models themselves but in the proprietary “harnesses”—the infrastructure and workflows that turn raw AI intelligence into practical, company-specific solutions, creating customer lock-in. Success depends on their ability to reduce token costs while embedding AI deeply into business processes through customized harnesses, as controlling this layer will determine dominance in the future AI economy.
The upcoming IPOs of OpenAI and Anthropic have sparked debates about their valuations, but the more pertinent question is what public investors are being asked to believe. Essentially, investors must trust that these companies can simultaneously make AI intelligence cheap enough to serve at massive scale and build proprietary layers—or “harnesses”—around that intelligence quickly enough that businesses prefer renting these systems rather than building their own. The “token” represents raw intelligence, while the “harness” is the infrastructure that turns this intelligence into actionable work, including tools, permissions, workflows, and evaluation mechanisms. The real value lies not just in the models themselves but in owning this work layer above the models.
API pricing often appears inflated compared to the value heavy users receive, but this reflects retail pricing with significant margins rather than the internal cost to serve tokens. OpenAI and Anthropic are likely subsidizing heavy usage while racing to reduce the cost of inference through efficiency improvements like model routing, caching, and hardware optimization. If they succeed in driving down token costs, raw intelligence becomes less defensible as a competitive moat, shifting the value to the harness layer that makes intelligence useful and accessible without requiring customers to understand the underlying technology.
A critical challenge for these AI labs is that they lack the private, company-specific context that businesses possess. They do not inherently know internal workflows, document locations, or approval processes. To overcome this, OpenAI and Anthropic are deploying forward engineers who work inside companies to map workflows, connect tools, and customize harnesses. This approach aims to transform generic AI systems into company-specific solutions, creating stickiness and lock-in by embedding the AI deeply into business processes. The question for companies is whether they will rent these harnesses or build and own them internally, as owning the harness means controlling the workflow and model orchestration layer.
The concept of recursive self-improvement is reframed here as a practical iterative advantage: smarter models help labs improve their own products faster, optimizing everything from code to routing and inference. This cycle could enable OpenAI and Anthropic to maintain cost leadership and build superior harnesses that most companies prefer to rent rather than develop themselves. However, if companies succeed in owning their harnesses, the labs become mere suppliers of intelligence tokens, reducing their strategic value and impacting their valuations.
Ultimately, the IPOs will reveal whether OpenAI and Anthropic can deliver on this dual promise of cheap tokens and powerful harnesses. Investors should look beyond revenue and user growth to metrics like cost-to-serve trends, margin improvements, enterprise adoption of scalable software versus custom labor, and the role of forward deployed engineering. For companies and individuals alike, the key takeaway is that AI strategy is not just about using models but about building and owning the harness that integrates AI into real work. Whoever controls this harness will dominate the future AI token economy, making it the trillion-dollar question at the heart of these IPOs.