Coursework for Big Data and AI in Business Strategy at IE. The question was whether OpenAI could defend a first-mover advantage in a market whose structure was still forming, and whose capital intensity had no real precedent. Three years on from the case being written, my answer was not flattering.
Written May 2026, before much of this was consensus.
A layered oligopoly, not a winner take all market
The value chain splits into chips, cloud, foundation models, applications and end users, and each layer has its own competitive structure while staying economically dependent on the others.
Concentration is highest at the chip layer, where NVIDIA's CUDA moat sits, and at the cloud layer between AWS, Azure and Google. Value migrates between layers but stays tight within the frontier model layer, which is where OpenAI lives and which is the single most contested part of the stack.
Every one of the five forces is tightening
Substitutes arrived fast, with Anthropic, Gemini and Llama all launching into the space OpenAI had to itself. Supplier power concentrated upstream, where there is no real alternative to NVIDIA at the current level of technology. Buyer power rose as multi-model platforms cut switching costs at the enterprise tier.
The barrier to entry moved too. The case put frontier training at 200 to 300 million dollars, which no longer holds. Chinese labs arrived with credible models, and open-weight releases under permissive licences put serious alternatives in front of end users at no cost.
The economics do not work like software
This is the part I think matters most. Software valuations rest on near zero marginal cost, where each additional customer is essentially pure margin once the product exists.
AI inverts that. Every query burns GPU cycles, electricity and cooling water. AI server racks pull 40 to 100 kilowatts against 5 to 15 for a traditional rack. Inference cost scales with usage, so revenue and cost rise together instead of pulling apart, and that is before the training bill.
The case assumed a 63% operating margin. That has not survived contact with reality: a reported burn of 7 to 13 billion dollars in the first half of 2025 against 4.3 billion of revenue.
Where the value is going
The market is not converging on one winner. It is moving away from compute and towards applications, agents, proprietary data and deployment into regulated industries.
Enterprise API spend doubled to 8.4 billion dollars between late 2024 and mid 2025, but OpenAI's share of it halved from 50% to 25% while Anthropic went to 32%. Domain specific models can beat larger generalists inside their niche, and government procurement has turned geopolitical alignment into a competitive vector aimed at a customer with a bigger budget than any enterprise.
The financing question
Underneath all of it, OpenAI's solvency is supported by circular arrangements that tie its capital base to the equity of its own suppliers, with a large part of the investment arriving as cloud credits rather than cash.
Set against an energy and macro backdrop that the market has not priced into the model layer, that is the structural issue, not the competitive one. A fading lead is survivable. A business model whose costs scale with its revenue, funded by the people it buys from, is a different problem.
Documents