OpenAI and Anthropic IPOs: The Battle for the Work Layer

The Shift from Raw Intelligence to the Work Layer

The core value proposition for AI labs like OpenAI and Anthropic is shifting from selling raw intelligence (tokens) to owning the 'harness.' A harness is the operational layer that turns raw model output into actual work. It includes file access, tool integration, permissions, memory, evaluation frameworks, and routing logic between different model tiers. The success of these companies depends on their ability to build these systems faster than their customers can build proprietary internal versions.

The Economics of Token Subsidies

High-usage plans, such as the $200 tier, are often viewed as irrational spending, but they likely function as a strategic subsidy. API prices are retail figures that include significant markups. If labs can drive down internal inference costs through chip utilization, model distillation, and caching, they can afford to subsidize power users to gain market share. This strategy aims to make intelligence abundant and cheap, which ultimately commoditizes raw tokens and forces value to migrate toward the software layer that orchestrates them.

The Conflict of Context and Lock-in

Labs face a fundamental information asymmetry: they possess superior models and infrastructure, but they lack the private context of their enterprise customers. Forward-deployed engineering serves as a bridge to overcome this, allowing labs to map internal workflows and adapt their products to specific business needs. If a lab successfully embeds itself into a company's core workflow, the resulting lock-in is based on the system architecture rather than the underlying model. Companies that fail to build their own internal routing, evaluation, and workflow definitions risk becoming entirely dependent on the lab's proprietary harness.

Evaluating the IPOs

Investors should look beyond revenue and valuation numbers in upcoming S-1 filings. Key indicators of a sustainable business include:

• Whether heavy user costs are trending downward as inference efficiency improves.
• Whether gross margins expand as usage scales.
• Whether enterprise revenue comes from scalable software or high-touch custom deployment labor.
• Whether forward-deployed engineering is a temporary bridge or a permanent requirement for product functionality.