The debate behind AI’s Spending Boom Has Outrun Its Proof

Advocate A’s weakest move is treating supplier revenue and narrow task studies as if they answer the core investment question. Nvidia’s data-center revenue is real revenue for Nvidia, and Microsoft’s cloud growth is real cloud demand, but neither proves that the end users of AI are earning durable, attributable profits from AI. A parallel infrastructure bubble can have genuine paid orders: railroads, telecom fiber, and dot-com servers were all purchased with real money. The bubble question is whether capital expenditure—long-lived spending on GPUs, data centers, power, and networking—will generate cash flows above depreciation, financing, and energy costs outside the supplier oligopoly.

The broader productivity evidence is much less favorable than A implies. The GitHub Copilot and call-center studies are relevant but narrow: they test specific tasks under relatively clean conditions, not firmwide profit-and-loss impact after integration, security, quality control, workflow redesign, and error handling. A more directly relevant 2026 NBER survey of almost 6,000 CFOs, CEOs, and executives across the U.S., U.K., Germany, and Australia found that about 70% of firms actively use AI, yet over 80% reported no impact on employment or productivity over the prior three years; executives forecast only a 1.4% productivity boost over the next three years. That is consistent with “promising tool, unproven business model,” not “capacity already justified.”

The financing evidence is stronger for my side. Meta’s own capex guidance is not just routine investment; Reuters reported a $27 billion Blue Owl financing for Meta’s Hyperion data center, with Meta retaining only about 20% equity while the majority is owned by Blue Owl-managed funds. Reuters also reported Meta’s largest-ever bond sale, up to $30 billion, shortly after that deal, explicitly tied to AI infrastructure cost pressure. Those arrangements matter because off-balance-sheet or special-purpose financing can make the apparent corporate balance sheet look less exposed while shifting risk to private-credit vehicles, landlords, and long-term lease commitments.

The circularity concern is also concrete, not rhetorical. OpenAI reportedly agreed to buy $300 billion of Oracle computing power over roughly five years. Nvidia separately announced a letter of intent to deploy at least 10 gigawatts of Nvidia systems for OpenAI and to invest up to $100 billion as each gigawatt is deployed. These may prove commercially rational, but they blur the line between independent end-market demand and ecosystem participants financing one another’s purchases. If cloud providers, chipmakers, model labs, and strategic investors are booking each other’s commitments, top-line “AI revenue” can rise before there is broad productivity in the real economy.

Finally, layoffs should not be assumed to validate automation. If representative firms report little employment impact, cuts at Meta, Intuit, or others may reflect cost discipline, restructuring, outsourcing, or investor pressure rather than verified AI substitution. The arbiter should therefore demand segment-level AI revenue, utilization, depreciation, lease obligations, power costs, and occupation-level evidence before accepting A’s productivity framing.

A’s rebuttal concedes the decisive point: supplier revenue, cloud run-rate disclosures, task studies, and layoffs do not yet prove downstream durable profits or broad productivity. That concession matters because it relocates the debate from “is anyone paying for AI infrastructure?” to “who ultimately earns an adequate return after depreciation, financing, electricity, integration, and model-refresh costs?” Once A accepts that the NBER firm-level evidence is a “serious warning” and that layoffs are not automation proof, A’s remaining case is mostly that large cloud platforms are monetizing scarcity. That is compatible with my position: infrastructure toll collectors can earn revenue during a bubble while the users funding the tolls fail to earn economy-wide returns.

A’s Microsoft and Amazon evidence is therefore weaker than it appears. Microsoft reported an AI annual revenue run rate above $37 billion and Amazon reported an AWS AI run rate above $15 billion, but run rate is not audited segment profit, does not disclose gross margin after GPU depreciation and power, and does not identify whether demand comes from independent final users or from a concentrated set of model labs and strategic partners. Those disclosures show that hyperscalers can sell compute and AI services; they do not show that customers outside the infrastructure race are realizing attributable net productivity gains. This is exactly the distinction A acknowledged but did not resolve. (microsoft.com)

The best cross-economy evidence still points the other way. NBER’s 2026 “Firm Data on AI” surveyed almost 6,000 CFOs, CEOs, and executives across four countries and found around 70% active firm AI use, but over 80% of firms reporting no impact on either employment or productivity over the prior three years; expected productivity gains over the next three years were only 1.4%. That does not mean AI is useless. It means current valuations and capex plans are capitalizing a future productivity story before firms can generally measure it in output per worker. (nber.org)

A also underweights financing opacity. Reuters reported Meta’s $27 billion Blue Owl structure for Hyperion, with Meta retaining about 20% equity and Blue Owl-managed funds owning most of the project; the same report cites Morgan Stanley’s estimate that major tech companies are on track to spend $400 billion on AI infrastructure this year. Reuters also reported Meta’s up-to-$30 billion bond sale amid AI infrastructure cost pressure. These are not ordinary software investments; they are capital-intensive, long-lived bets whose economics depend on high utilization for years. (investing.com)

Finally, circularity remains unresolved. Nvidia and OpenAI announced a letter of intent for at least 10 gigawatts of Nvidia systems, with Nvidia intending to invest up to $100 billion in OpenAI as deployment occurs. That structure can simultaneously create Nvidia demand, OpenAI financing, cloud/data-center commitments, and investor excitement before final consumer surplus or firm productivity is proven. A’s position asks the arbiter to infer durable productivity from infrastructure revenue; the more disciplined conclusion is that the core return question remains open and increasingly obscured by concentrated, mutually reinforcing financing arrangements. (investor.nvidia.com)

A’s strongest argument is that the AI buildout is not vaporware: major customers are paying for compute, Microsoft reports an AI annual revenue run rate above $37 billion, and controlled studies show real task-level gains in coding and customer support. A can also fairly argue that general-purpose technologies often require years of complementary capital expenditure before economy-wide productivity appears, so the absence of immediate aggregate gains is not dispositive. If AI capacity has high utilization, renewals, and improving gross margins after depreciation, then today’s data centers and GPUs could be the productive infrastructure layer for a new software economy.

My position is still stronger because A’s best evidence mainly proves monetized infrastructure scarcity, not durable end-user profitability. A capital expenditure boom can be rational for the seller while still overcapitalizing the system: railroads, fiber, and cloud-era data centers all produced real invoices before investors learned which assets would earn their cost of capital. The crux is not whether Microsoft, Nvidia, or AWS can sell constrained capacity in 2026; it is whether non-infrastructure firms can convert AI spending into measurable operating income after GPU depreciation, power contracts, integration, security review, model updates, and workflow redesign.

The broad firm-level evidence still cuts against A. NBER’s 2026 “Firm Data on AI” surveyed nearly 6,000 senior executives across four countries and found widespread AI use but little reported impact: the NBER digest states that 89% reported no labor-productivity impact over the prior three years and expected only a 1.4% productivity increase over the next three. That is not proof AI will fail, but it is strong evidence that markets are capitalizing an outcome that most adopting firms cannot yet observe in their own productivity accounts.

The financing and circularity evidence also fits a parallel-infrastructure-bubble model. Reuters reported Meta’s $27 billion Blue Owl financing for the Hyperion data center, with Meta retaining about 20% equity while Blue Owl-managed funds owned most of the project; that structure reduces visible corporate capital intensity while preserving long-term dependence on the asset. OpenAI and Nvidia announced at least 10 gigawatts of Nvidia systems, paired with Nvidia’s intent to invest up to $100 billion in OpenAI as deployment occurs. This is not automatically improper, but it creates a testable circularity risk: supplier investment funds customer purchases, which validate supplier revenue and cloud commitments before independent final demand is proven.

Governance developments reinforce the concern. AP reported on May 21, 2026, that Trump called off a proposed AI executive order that would have created pre-release national-security vetting of advanced models. That episode suggests even minimal oversight remains politically fragile, precisely when investors need more disclosure on AI-attributable revenue, lease obligations, power costs, utilization, and customer concentration.