Susquehanna sees 8% of AI data center projects at risk

Susquehanna estimates that 60 GW—about 40% of 148 GW in training-focused AI data center projects planned for 2025–2030—could face delays pending regulatory approval; in a worst-case scenario, 12 GW planned for 2028–2030 could be canceled. Near-term disruption appears limited because most 2027 projects are permitted, while enterprise inferencing and recovering non-AI semiconductor demand could partly offset a shortfall. The firm estimates a worst-case 10% year-over-year semiconductor-industry revenue decline; the PHLX Semiconductor Sector Index is down about 38% from its June peak and trades near a historical forward P/E trough of approximately 15 times.

Analysis

The key distinction is between a delayed buildout and permanently lost demand. Permitting uncertainty is concentrated far enough out that the first-order impact is more likely to be a longer revenue-conversion cycle than an immediate collapse in chip orders; earnings risk rises if delayed projects cause hyperscalers to reduce—not merely defer—capital budgets. Exposure is also uneven: semiconductor equipment and data-center power/cooling suppliers would be vulnerable to deferred campus construction, while enterprise inference could redirect spending toward installed compute, networking, and utilization rather than replace every planned training facility one-for-one.

The broad semiconductor valuation and drawdown argue against treating this as a fresh sector-wide short signal. The more credible risk is estimate dispersion: consensus may underweight the gap between announced capacity and permitted, financed, and ultimately energized capacity. Over the next 1–3 months, watch permit outcomes and hyperscaler capex guidance; over 6–18 months, monitor whether project deferrals show up in equipment orders and chip-revenue revisions. The thesis weakens if approvals clear without meaningful schedule slippage, or inference demand and non-AI recovery sustain orders. It strengthens if cancellations spread beyond the stated worst-case pool or major customers cut budgets.

Key Decisions for Investors

• Avoid adding an outright semiconductor short solely on this report: the potential earnings effect is delayed and could be partly offset by inference and non-AI demand. Treat broad sector exposure as a monitor, not a high-conviction directional trade.
• For existing AI-capex-heavy exposure, consider a small, defined-risk 3–6 month SOXX put spread, preferably entered on a rally rather than after further weakness. Close or reduce the hedge if permitting clears and major customers maintain capex plans; reassess if semiconductor revenue estimates begin moving down.
• Favor relative-value scrutiny over blanket selling: compare semiconductor equipment and data-center power/cooling exposures with businesses more levered to utilization or enterprise inference. Do not assume announced projects translate into orders; verify order backlogs, cancellation clauses, and customer capex guidance before sizing company-specific positions.
• Set a catalyst alert for the next 1–3 months: track permitting decisions and hyperscaler commentary. A string of approvals with no capex cuts would falsify the near-term bearish case; repeated delays accompanied by order cancellations or guidance reductions would warrant a stronger underweight.