The Buildings Are Full. The Finances Are Fraying.

In the same week that data center vacancy in Northern Virginia hit 0.3% — a record low, meaning effectively every rack is leased — S&P cut Oracle's credit rating to one notch above junk ^[1][2]. The physical layer is sold out. The financial layer is being downgraded. The gap between those two facts is what the AI sell-off is actually about. The demand is not a mirage. Micron's high-bandwidth memory — the specialized chips that make AI training possible — is sold out through 2027 ^[3]. Nvidia reports a trillion dollars in demand visibility for its Blackwell and Rubin systems through the same year ^[4]. Global AI spending is projected to approach $2 trillion in 2026 ^[4], and CBRE projects a supply-demand imbalance in data center space lasting at least three more years ^[1]. This is not a technology correction. The market is not questioning whether the data centers will be used. It is questioning whether the financial structures that built them can survive their own tenants. Start at the top of the chain, with the companies that rent the compute. OpenAI, the largest single AI infrastructure buyer, cut its projected 2030 infrastructure spending from $1.4 trillion to $600 billion in February, aiming to align capital expenditure with a $280 billion revenue forecast ^[5]. Its adjusted gross margin fell from 40% in 2024 to 33% in 2025 as inference expenses quadrupled ^[5]. In April, the company missed internal revenue and user targets, and CFO Sarah Friar warned it may struggle to fund its $600 billion in compute contracts without accelerated growth ^[6]. At the same time, DeepSeek permanently cut its V4-Pro API prices by 75% in May, making its models 12 to 19 times cheaper than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 for equivalent tasks ^[7]. These are parallel pressures — one from costs, one from competition — and both compress the revenue per token that model providers need to service long-term cloud contracts. That tenant fragility propagates downward. On July 9, S&P downgraded Oracle's credit rating from BBB to BBB- — one notch above junk — citing negative free cash flow and extreme customer concentration: roughly half of Oracle's $638 billion in contracted future revenue depends on a single tenant, OpenAI, via a $300 billion cloud deal ^[2]. The shares have fallen 65% from their 52-week high ^[2], driven, as one report put it, by "investor concerns over the financial stability of its partner, OpenAI" ^[8]. Oracle added $43 billion in debt during fiscal 2026 and plans to raise another $40 billion in the current year to fund AI data-center expansion ^[2]. The market is now pricing a question that was not on the spreadsheet when those contracts were signed: what happens to the infrastructure provider when the tenant's margins keep shrinking? Even the wealthiest hyperscaler is not immune. Microsoft, OpenAI's primary backer, saw its stock fall 20% in June despite 40% Azure growth ^[8][9]. The company froze hiring in Azure cloud and North American sales to fund AI infrastructure spending, with Azure Core chief of staff Hilary Macfadden warning that "until we have credible, executable plans locked to address that gap, pressure will continue to cascade" ^[10]. And Microsoft is now replacing OpenAI and Anthropic models with in-house MAI models across Excel, Word, Outlook, and Teams to cut token costs ^[11]. Microsoft AI CEO Mustafa Suleyman made the logic explicit.

We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost — Mustafa Suleyman

The largest AI customer is actively working to reduce its spending on external model providers — the very tenants whose long-term contracts underwrite the infrastructure buildout. At the bottom of the chain sits Nvidia, the hardware vendor that supplies the chips everyone else borrows to buy. In July, Nvidia launched a revenue-sharing financing model that allows AI startups to access GPU infrastructure without upfront capital, in exchange for a share of product revenue ^[12]. The traditional capital-expenditure model is breaking down: the hardware vendor is now financing its own customers' access to compute because they cannot raise the capital independently. The cost pressure that makes such financing necessary is not abstract. Nvidia's own vice president of applied deep learning, Bryan Catanzaro, made it concrete.

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

The market has already begun to differentiate. Goldman Sachs notes a strategic shift in investor preference from semiconductor stocks toward hyperscaler stocks, as enterprises are expected to demonstrate tangible returns from AI investments ^[13]. The Magnificent Seven lost $2.3 trillion in market value in June, concentrated in the capex spenders — Microsoft down 20%, Nvidia down 13% — while the broader S&P 500 dropped only 2% ^[9]. This is not a broad market panic. It is a specific repricing of the financial intermediation layer: the companies sitting between real physical demand and uncertain end-user revenue. The entire chain — from tenant margins to vendor credit — is leveraged against a single assumption: that AI revenue will grow fast enough to cover the cost of the infrastructure built to serve it. The market is no longer taking that assumption on faith.