he central question surrounding the potential IPOs of companies like OpenAI and Anthropic is whether they can evolve beyond being providers of 'raw intelligence'—the metered consumption of API tokens—to become the dominant operating layer for global knowledge work. The thesis is that raw intelligence is destined to be a low-margin commodity, similar to electricity or bandwidth. To maintain a trillion-dollar valuation, these labs must convince public investors that they can build a 'harness' around that intelligence. A harness is the structural software layer that includes file visibility, tool access, permissions, memory, and routing logic, which ultimately transforms a model's output into a finished, actionable task.
The labs are essentially engaged in a race to build these harnesses faster than enterprises can build their own. This competition is driven by a fundamental information asymmetry: labs have speed, infrastructure, and model talent, but they do not know the 'private context'—the specific, messy internal processes, data sources, and truth-defining documents—that make a company run. Enterprises currently hold the advantage in this context-heavy domain, provided they take the initiative to build their own independent AI infrastructure rather than simply consuming lab-provided tools.
The labs combat this via 'forward-deployed engineering,' a tactic designed to map and adopt customer workflows directly into their own systems. While this may look like basic consulting, it is a strategic attempt to lock customers into their specific harness. If a customer's workflow becomes deeply entangled with a lab's proprietary system, the model underneath becomes irrelevant because the switching cost for the workflow itself is too high. This creates a long-term defensive moat that extends far beyond the intelligence of the model.
Ultimately, the IPO S-1 documents will reveal the true health of this strategy. Investors should look beyond revenue and burn rates to assess whether gross margins are improving as usage scales and whether enterprise customers are buying repeatable software or expensive manual labor. If companies learn to own their own harnesses, the labs will be relegated to being commodity suppliers of intelligence, significantly dampening their long-term valuation potential. The successful enterprise AI strategy involves owning the context, maintaining independent evaluation layers, and building routing logic that allows for model substitution, ensuring the company—not the lab—captures the value of their own workflows.