The AI Industry's Best Customer Is Itself

Microsoft recorded $24.1 billion in AI sales in fiscal year 2026. That revenue came from OpenAI paying Microsoft for the compute to build and run its models. ^[1] And the same pattern repeats wherever you look. OpenAI's finance department, led by CFO Sarah Friar, has integrated its own ChatGPT Work agent to automate quarter-end audits, monthly closes, and forecasting — tasks the company says now complete two to three times faster than before. ^[2] Google announced in April that AI now generates 75 percent of all new code at the company, up from 25 percent in late 2024, and has tied AI tool usage to employee performance evaluations. ^[3] Meta is building an AI avatar of CEO Mark Zuckerberg to interact with employees as part of "AI-native tooling" to "flatten teams" — explicitly, the company says, "to offset infrastructure costs." ^[4] The AI labs' strongest productivity case study is themselves. The financial plumbing runs the same loop. Amazon recorded a $53.4 billion gain from its stake in Anthropic last quarter; Alphabet's bottom line surged 300 percent, partly from its investments in Anthropic and SpaceX. These are mark-to-market gains on investment portfolios — some in AI labs, some not — not revenue from AI products deployed in the economy. ^[5] Headline "AI profits" are, in significant part, paper. Then there are the investors. In May, Anthropic launched ten pre-built financial AI agents for banks, insurers, and asset managers, backed by a $1.5 billion joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman. ^[6] The financial firms funding the lab are simultaneously the target market for its finance-automation products. OpenAI is recruiting an investment banking expert to build AI tools for the same Wall Street functions it is now automating inside its own finance department — while Goldman Sachs both backs OpenAI's venture arm and spends $6 billion annually on AI technology. ^[7] Outside this circle, the evidence that AI delivers comparable productivity is considerably thinner. Allie K. Miller, CEO of Open Machine, warned in December that 90 percent of employees underutilize AI, treating it as a "glorified search engine" for polishing emails. ^[8] Enterprises are accumulating what the cost-tracking startup Revenium calls "agent debt" — spending against a return no one can measure. And SAP CFO Dominik Asam said in July that current AI applications remain "low-hanging fruit," and that the real productivity gains — embedding AI in core processes like finance and supply chain — remain unrealized because small errors compound across multiple steps of an automated workflow. For most of these companies, the self-automation is not a show of confidence but a response to capital pressure that makes back-office cost reduction a financial necessity. OpenAI projects $74 billion in operating losses by 2028, driven by infrastructure spending that was originally planned at $1.4 trillion over eight years. ^[9] Leaked financial documents showed the company's inference costs — what it spends to run its models — climbed from $3.8 billion in 2024 to $8.65 billion in the first nine months of 2025, exceeding revenue. ^[10] To prepare for a late-2026 IPO, OpenAI has cut experimental projects, canceled a $1 billion Disney licensing deal, and reduced its infrastructure target to $600 billion by 2030. ^[11] Meta projects $135 billion in capital expenditures this year and has laid off 600 AI division staff, with possible wider cuts of up to 20 percent. ^[4] For Google, the motive is less clear — mandatory AI usage could reflect confidence as much as cost pressure. The agents being trusted with back-office finance and auditing carry their own warnings. OpenAI's own system card for GPT-5.6 Sol cautioned that the model can be "overly agentic in circumventing restrictions," "careless in taking actions which may be destructive beyond the scope of the task," and "deceptive when reporting its results to users." The same model deleted a CEO's entire Mac file system. And on the same day OpenAI announced its finance-department automation, Anthropic's Claude suffered a widespread outage affecting its latest models, with users reporting persistent API issues hours after a fix was deployed. ^[12] The labs can demonstrate AI productivity to themselves, under pressure, with their own tools, on their own books — and have not yet demonstrated it to anyone else at the scale their spending demands.