Detailed Analysis
OpenAI's IPO filing has reignited debate about the trillion-dollar valuations being assigned to frontier AI companies, with Anthropic reportedly moving toward a similar public offering. The article argues that the central question for public investors is not whether these valuations are numerically justified, but rather whether OpenAI and Anthropic can simultaneously drive the cost of AI inference toward commodity-level pricing while building proprietary "harnesses" — the structured layers of tooling, memory, permissions, routing, and workflow management that transform raw model output into productive work — fast enough to prevent enterprise customers from building those systems themselves. The distinction between a token and a harness is the analytical core of the piece: tokens are raw intelligence sold by unit, while harnesses are complete work environments that abstract away the complexity of deploying that intelligence. Products like Codex and Claude Code are cited as the clearest expressions of this harness strategy in practice.
The article directly addresses a widely circulated analysis — attributed to SemiAnalysis — estimating that heavy users of the $200/month plans from both OpenAI and Anthropic are extracting API-equivalent value of $14,000 and $8,000 respectively. Rather than accepting the surface-level interpretation that these companies are destroying capital, the article reframes the math by noting that public API prices are retail prices that already embed significant gross margins, potentially 70–80%, and that internal inference costs are substantially lower than those sticker prices. The $200 plans, viewed through this lens, may represent a deliberate strategy: allowing high-volume power users to consume intelligence at near-cost while the labs race down the inference cost curve through improvements in caching, batching, model distillation, chip utilization, and routing between cheap and expensive models. The subsidy today funds the scale data and infrastructure position needed to make the economics work at mass penetration tomorrow.
This cost curve argument is central to the entire IPO thesis. If inference costs plateau and token prices remain high, the unit economics of serving intelligence at scale become structurally difficult, and the business model becomes vulnerable. But if the labs can continue compressing the cost of serving intelligence — a trend that has held historically across compute cycles — then cheap tokens become a feature rather than a problem. Abundant, low-cost intelligence would commoditize the raw model layer, mirroring how electricity, bandwidth, and compute became infrastructure inputs rather than differentiators. The value, as the article argues, would then migrate upward into the harness layer: the proprietary work surface that makes intelligence useful without requiring customers to understand the underlying architecture.
The deepest strategic tension the article identifies is the information asymmetry between the labs and the enterprises they are trying to serve. OpenAI and Anthropic possess models, infrastructure, engineering velocity, and aggregated usage data. But they lack the one input that makes harnesses genuinely powerful at the organizational level: private context. The labs do not know which internal documents are authoritative, which approval workflows are real versus ceremonial, which data fields in enterprise systems actually drive decisions, or which institutional histories explain why processes are structured the way they are. Enterprise customers hold that context entirely. The entire competitive fight over the work layer, then, is a race between the labs building general-purpose harnesses capable enough to be adopted before enterprises develop sufficient internal tooling capability, and enterprises accumulating the AI engineering talent and platform knowledge to build their own context-aware harnesses before the labs' products become indispensable. The IPO story is ultimately a bet on which side wins that race — and at what speed.
Read original article →