America is planning more AI datacenters than its chip supply can fill

US datacenter deployments are projected to rise from about 11 GW in 2025 to 16-18 GW in 2026, but practical 2027 capacity appears capped in the low-20-GW range—far below the more than 80 GW implied by some announced 2028 pipelines. Jefferies and SynMax cite slowing land-clearing activity, permitting and power-access delays, and advanced AI-chip packaging capacity as key bottlenecks; existing packaging supports roughly 17.5 GW of total datacenter power, with planned 2027 projects adding about 6 GW. Concentration in Texas and Virginia heightens execution risk, particularly as unpredictable ERCOT curtailments can undermine datacenter uptime economics.

Analysis

The binding constraint is shifting from accelerator demand to installed, monetizable compute. That distinction matters: NVDA and AVGO can retain strong order books while quarterly revenue conversion becomes paced by packaging, server integration, power delivery and customer-site readiness. The near-term beneficiaries are suppliers whose revenue is recognized before a facility is energized—VRT, ETN, GEV, PWR and MYRG—although their backlog quality should be discounted where customer projects lack firm interconnection agreements. EQIX and DLR face a more mixed setup: constrained supply supports pricing on commissioned capacity, but delays defer development yields and can lift carrying costs.

A lower realized buildout path would be most damaging to the highly leveraged "AI power shortage" narrative embedded in merchant generators CEG and VST, particularly where contracted load assumptions underpin long-dated valuation multiples. Texas is the key nonlinear risk: interruptible load may improve speed-to-connection, but unreliable curtailment forces customers either to overbuild backup generation/storage or accept lower utilization. That creates an opening for distributed-power vendors such as GEV and Caterpillar (CAT), while potentially reducing the amount of grid-connected capacity required per announced campus.

Advanced packaging creates a second-order advantage for TSM and packaging-adjacent supply chains, but it also weakens the thesis that domestic wafer fabrication alone de-risks US AI infrastructure. The market may be underpricing the possibility that 2027-28 accelerator shipments are allocation-constrained rather than demand-constrained; that would preserve chip pricing but push out cloud capex depreciation and customer revenue realization. Jefferies (JEF) has limited direct earnings sensitivity, so this is principally a sector-positioning signal rather than a catalyst for its shares.

Over the next 1-3 months, construction starts, utility interconnection awards and accelerator-package lead times matter more than announced campus totals. Over 6-18 months, the thesis is falsified if visible construction activity and firm grid commitments accelerate enough to support materially higher energized capacity, or if packaging additions arrive ahead of schedule; either outcome would re-rate power producers and data-center developers higher.

Key Decisions for Investors

• Prefer a 6-12 month long VRT / short EQIX pair: VRT captures retrofit and power/cooling intensity even when campus completion slips, while EQIX is more exposed to delayed asset commissioning. Reassess if EQIX signs unusually large pre-leased capacity with firm utility energization dates or if VRT backlog conversion slows for two consecutive quarters.
• Trim or avoid adding to CEG and VST on AI-load optimism until contracts disclose interconnection status, curtailment terms and credit support. A practical risk control is a 3-6 month CEG/VST put spread around the next utility-load disclosure cycle; upside risk is faster-than-expected Texas grid connection or hyperscaler-backed long-term PPAs.
• Maintain core NVDA exposure but shift incremental AI infrastructure beta toward TSM and VRT rather than assuming every announced GPU order converts on the original schedule. The thesis fails if accelerator packaging lead times normalize while hyperscaler capex guidance remains elevated, which would restore a more direct volume upside for NVDA.
• Watch PWR and MYRG for confirmation rather than chase: initiate only after quarterly backlog growth is accompanied by margin expansion and disclosed data-center interconnection work. The favorable setup is 12-18 months of transmission and substation spending; the key downside is project deferral converting nominal backlog into lower-margin remediation work.