Ricardo Semler on the prisoner’s dilemma on data centers: trillions of dollars obsolete in just a few years

The article warns that the AI data-center buildout could leave municipalities and investors with stranded, highly specialized assets as computing hardware outpaces building infrastructure. North American data-center vacancy is just 1%, customers contracted for a record 25GW of generation capacity in H1 2026, and roughly $7 trillion of financing and investment is being lined up, but 77% of pipeline projects are shifting to frontier power markets. Hardware may become obsolete in 3-6 years versus 30-60 years for buildings, while AI racks have risen from 5-10kW five years ago to 100-250kW today, with proposals up to 600kW. The piece highlights risk transfer through structures such as Meta's $27B Louisiana Hyperion joint venture, where Blue Owl owns 80% and Meta uses a four-year lease with extensions, potentially leaving incentives, utilities, and pension investors exposed if the project fails.

Analysis

The investable distinction is between compute-refresh beneficiaries and long-duration asset owners. Higher rack-density and changing cooling/power architectures shorten the useful economic life of electrical distribution, liquid cooling and networking; VRT and ETN can monetize repeated retrofit cycles, while DLR/EQIX and private-credit vehicles are exposed if tenant demand shifts toward newer campuses before remaining lease economics support the original capital stack. This is less a near-term demand-collapse thesis than a residual-value and refinancing-risk thesis.

META's use of third-party capital can preserve reported leverage and near-term FCF optics, but it does not eliminate economic exposure if residual guarantees, renewal decisions, or power-delivery delays impair project returns. For OWL, the primary risk is not necessarily immediate balance-sheet credit loss; it is weaker fundraising, lower realizations and fee-related earnings multiple compression if investors reassess data-center infrastructure underwriting. The relevant 1-3 month catalysts are financing disclosures, lease tenor/residual-guarantee detail, utility interconnection slippage, and any widening in data-center ABS/private-credit spreads.

Consensus is likely too bearish on physical capacity demand but too complacent on who owns the stranded asset. Efficiency gains can expand aggregate compute consumption while still destroying the economics of a specific site or vintage. Over 6-18 months, scarce deliverable power should support premium, grid-connected campuses, whereas remote sites reliant on incentives, speculative power assumptions, or single-tenant leases deserve materially higher cap rates and lower terminal values.

Key Decisions for Investors

• Initiate a 6-12 month pair: long VRT / short DLR, sized market-neutral. VRT benefits from retrofit and higher-density cooling spend; DLR carries greater terminal-value and tenant-concentration sensitivity. Reassess if DLR discloses sustained cash rent spreads and long lease extensions that more than offset rising development yields.
• Maintain OWL as an underweight/short watch candidate rather than a full position until vehicle-level leverage, residual guarantees and investor concentration are disclosed. Trigger a short on evidence of fundraising slowdown or wider financing spreads; invalidate on resilient fee-related earnings guidance and continued inflows into infrastructure/private-credit strategies.
• Avoid treating META's off-balance-sheet structures as pure leverage relief. For the next two earnings cycles, monitor committed versus operating capacity, lease obligations, guarantees and capex guidance; reduce META exposure if those commitments rise faster than monetization evidence, but do not short solely on project-finance optics.
• Prefer ETN and VRT over broad data-center REIT exposure for a 6-18 month AI-infrastructure allocation. The risk/reward improves on any sector pullback driven by power-delay headlines, provided orders/backlog continue to convert and margins remain supported by liquid-cooling and electrical-content mix.