AI adoption may be inflated by subsidies, bundled distribution, and financing structures that obscure weak underlying economics.

Audited financials verified by the Financial Times confirm OpenAI lost $20.92 billion from operations in 2025 on $13.07 billion in revenue. Total costs hit $34 billion. Research and development alone, at $19.18 billion, exceeded the entire top line. The cost of serving users — inference — climbed from $2.65 billion to $7.5 billion in one year. More users means more compute. More compute means more loss. The loss grows with the revenue. That is not a growth phase. That is a structural ceiling.

OpenAI loses approximately $1.60 for every dollar it earns. Only 5.5 percent of its 900 million weekly users pay anything. The other 94.5 percent are being served for free at OpenAI's expense. ChatGPT Plus costs $20 per month. Sam Altman publicly admitted the $200 per month Pro tier loses money. If a $200 product is unprofitable, the $20 product is a charity. HSBC analysts estimate OpenAI requires over $207 billion in additional funding by 2030 just to keep the lights on.

Law of the Addict. The service is free or cheap enough to feel cost-free. So people use it constantly — far more than the price implies. OpenAI built a product so compelling it cannot afford the demand it created. Users feel like winners. They are not. Every free query is a micro-subsidy funded by SoftBank and Microsoft. The addiction is real. The business model is not. When the investor subsidy ends, the price goes up or the service degrades. Either way, the user pays eventually.

Anthropic projected its first-ever operating profit of $559 million for Q2 2026. Analysts at Prof G Media traced the mechanism: Anthropic signed a compute deal with SpaceX's Colossus data centres, with a heavily discounted ramp-up fee covering precisely May and June — the two months used to define the profitable quarter. Once the ramp-up ends, the contract resets to approximately $15 billion per year in compute costs alone. Analyst Ed Zitron stated the profit is the result of accountancy, not any improvement to the underlying business model.

Anthropic has raised approximately $132 billion across 18 funding rounds. It must sustain an estimated $80 billion in cloud infrastructure costs through 2029 via Amazon and Google partnerships — before the SpaceX contract is added. Its Q2 2026 operating profit of $559 million evaporates the moment the discounted rate expires. For UK and US enterprise buyers currently signing Anthropic contracts assuming a path to stable pricing, the cost structure of their AI vendor is more fragile than the headline suggests.

Law of Projection. Anthropic announced profitability. The press echoed it. The market priced it. The reality was a pre-negotiated discount window timed to a fundraising narrative. The announcement projects health onto a body still in surgery. People feel confident about a vendor that has never been sustainably profitable and may not be for years. That false confidence shapes procurement decisions, contract lengths, and strategic dependency. By the time the reality surfaces, the lock-in is already signed.

Microsoft spent $80 billion on AI infrastructure in 2025. Its Copilot product reached only 15 million paid seats from a base of 450 million Microsoft 365 users — a penetration rate of barely 3.3 percent. Copilot's share of the paid AI subscriber market contracted from 18.8 percent in July 2025 to 11.5 percent in January 2026. When Microsoft found its engineers preferring Anthropic Claude Code, it ordered them to switch back to Copilot — with sources confirming the decision was partly financial, closing quarter in June. Microsoft is spending its way into a product its own staff do not choose.

Copilot costs enterprises $30 per user per month on top of existing Microsoft 365 licensing. A Gartner survey found only 5 percent of IT leaders were expanding Copilot deployment. A July 2026 analysis confirmed that enterprises deploying 500 seats were seeing only 22 percent of employees reporting any productivity gain after six months. Morgan Stanley reported global tech companies have announced $740 billion in capital expenditures this year — a 69 percent jump. If the AI products cannot demonstrate ROI, those capex multiples compress. MSFT's AI narrative is the valuation story. The AI ROI data is the risk.

Law of the Landlord. Microsoft owns the infrastructure, the operating system, the office suite, and now the AI layer. Copilot is bundled into subscriptions millions of businesses already pay for. The landlord does not need the product to be the best. It needs it to be unavoidable. Enterprises adopt Copilot not because it outperforms but because it is already inside the building they rent. Preference data shows only 8 percent of users with access to all three platforms choose Copilot. That is not adoption. That is captivity dressed as convenience.