AI's Infrastructure Is Built for Decades. Its Models Last Months.

In late June, Mark Zuckerberg told Meta shareholders something unusual for a CEO presiding over $145 billion in annual capital spending.

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

It was the clearest tell yet from inside the build. The man writing the largest check in the industry is already hedging — and the same month, Meta launched a new business called Meta Compute to sell its excess GPU capacity to other companies ^[1]. The announcement sent shares of CoreWeave and Nebius down as much as 17%, because it punctured the premise that AI compute would remain scarce indefinitely. The AI market has split into two layers, and they run on incompatible clocks. The infrastructure clock moves at the tempo of concrete. U.S. data center vacancy has fallen to 0.3% in Northern Virginia, the world's largest market. CBRE's Pat Lynch described the supply-demand gap in plain terms.

My thoughts are we’re at least three years out before we catch up with the supply that equal the demand that's out there. — Pat Lynch

Construction is running at twelve times 2020 levels, and hyperscalers are projected to invest $3.7 trillion over five years ^[2]. The financing structures are built for decades. Nexus Data Centers is seeking roughly $15 billion — a $14 billion bridge loan plus a revolver led by Morgan Stanley — for a single Texas campus serving Anthropic, with Google providing default guarantees in exchange for equity and Broadcom funding the chips through vendor financing ^[3]. Google is backstopping another $35 billion in loans for Anthropic's chip purchases across five facilities, with Apollo and Blackstone providing the debt and Google repaying bondholders if the operator defaults ^[4]. These are thirty-year assets, financed against the genuine scarcity of kilowatts — and that scarcity is real. The model clock moves at the tempo of a software release cycle. On May 23, DeepSeek made its 75% price cut on the V4-Pro model permanent, pricing it at 12 to 19 times less than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 for equivalent tasks ^[5]. In early August, Arcee AI trained a competitive open-weight model for $20 million in 33 days ^[6]. Chinese models now process four times the global token volume of U.S. models. All six of the most-used models globally are Chinese ^[7]. The gap between these two clocks is already producing a bifurcation in how the market's biggest participants behave. PNC Financial CEO Bill Demchak is investing in his own GPU compute to escape token-based pricing entirely.

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

Enterprises are deploying routing tools that assign routine queries to cheap open-source models like DeepSeek, reserving premium models only for complex workloads ^[8]. Pinterest uses DeepSeek R-1 for its recommendation engine and reports that in-house open-source models are 30% more accurate than the leading proprietary alternatives ^[9]. Airbnb runs Alibaba's Qwen for customer-service agents. Microsoft is canceling thousands of internal Claude Code licenses and redirecting engineers to GitHub Copilot CLI after a surge in experimentation drove up costs; Uber exhausted its entire 2026 AI coding-tool budget in the first four months of the year ^[10]. A new sector has emerged in the gap. In early August, a wave of AI cost-saving startups — coaches, measurers, and query-routers — began explicitly targeting what they call irrational enterprise spending on frontier models ^[11]. Their entire business model rests on the premise that companies are using expensive, versatile models for menial tasks that do not justify the cost. Runware's assessment is the bluntest thing anyone in the industry has said on the record.

The incentive right now is for people to say it is paying off and to find a way to justify it, and because of this, we don't have an accurate view of the market, as nobody wants to be the canary in the coal mine. — Runware

The counterargument deserves its weight. BlackRock's Jay Jacobs argues the build-out is less speculative than past cycles because compute is "almost instantaneously being monetized," citing 17x growth in token consumption last year and a supply-demand gap that suggests under-investment, not over-investment ^[12]. Google Cloud's backlog exceeds $460 billion, and CEO Sundar Pichai says revenue from generative AI products grew nearly 800% year over year ^[13]. Gartner and Goldman Sachs analysts argue that even as per-token prices fall, autonomous agentic AI systems will drive consumption higher — the Jevons paradox, in which cheaper unit prices expand total demand ^[10]. Token consumption can grow 17x while per-unit prices collapse, and the buildings still depreciate on a schedule the models do not respect. The infrastructure is financed against thirty-year leases; the product those buildings house is being commoditized before the concrete cures. Even Microsoft, the most successful AI monetizer, tells the story in both columns: its AI business hit a $37 billion annual revenue run rate, growing 123% year over year — and its quarterly capex reached $30.88 billion, its OpenAI investment losses widened to $3.1 billion, and its stock fell 17% year to date ^[14]. The data centers being financed today will still be standing — and still being paid for — long after the models they were built to house have been replaced four times over.