The Trillion-Dollar Illusion?

Summary The industrial system under construction around artificial intelligence is larger, more interconnected and more financially demanding than the familiar image of a software startup suggests. Anthropic filed its U.S. IPO registration statement in October, revealing explosive revenue growth and a layered program of future computing commitments valued in the hundreds of billions. These are genuine developments. The dispute is whether future customer spending will justify the scale, timing and rigidity of the obligations being created today. [1][2][7] An unusual financial pattern connects the major participants: cloud companies invest in laboratories that purchase their services; hardware manufacturers sell the processors and participate in financing demand; infrastructure operators borrow against long-duration customer contracts; alternative asset managers purchase data centers and lease them to technology tenants. The IMF now describes circular financing as a risk transmission mechanism rather than an anecdotal peculiarity. None of this proves fictitious demand. It does mean that investors should not mistake every signed contract, funded data center and equipment sale for independent confirmation of final-customer willingness to pay. [4][14][15][24] The immediate economic contradiction concerns time. AI capabilities are becoming rapidly cheaper: Stanford documented a more than 280-fold decline in the price of GPT-3.5-level inference in approximately eighteen months; the OECD reported nearly an 80% decline in its quality-adjusted model price index through April 2026. Meanwhile the industry is signing leases and computing commitments extending across much of the next decade. If the cost per useful task declines faster than revenues grow, utilization, margins and debt service become decisive. Increasing token usage is not itself proof of increasing profits. [22][23] The physical expansion is real. The IEA estimates data-center electricity demand at 485 TWh in 2025, with a central projection of 950 TWh by 2030. Equipment manufacturing, grids, permits, water and financing do not move at model-release speed. Project Jupiter in New Mexico demonstrates that power and construction delays can interfere with a sophisticated contract chain before the computing it promises has been delivered. A failed project can impose costs beyond its developer if grid upgrades or other public commitments were financed in anticipation of its demand. [17][19][21] The inquiry finds a legitimate commercial market, extraordinary growth and serious structural vulnerabilities. It finds no basis for claiming that laboratory sales are broadly fabricated or that a coordinated financial fraud has been proved. The strongest conclusion is less sensational and more important: AI may create enormous economic value even if some leading model developers, lenders or infrastructure operators fail to capture it. Technological success and the financial survival of particular firms are separate propositions.