The bottleneck in AI data center expansion is shifting from compute to power, cooling, and scaled delivery

Overview

Bernstein invited Raj Parihar, who previously led data center infrastructure procurement at Meta and Microsoft and has more than 15 years of industry experience, to participate in an expert webinar. The discussion covered the procurement sequence for power and cooling in hyperscale AI data centers, equipment lead times, power architectures, liquid cooling, prefabrication, modularization, supplier selection, and dual-supplier strategies. The core conclusion is that whether AI data centers can go live on schedule increasingly depends on whether power and cooling infrastructure can be secured in advance and delivered reliably, rather than solely on the availability of computing hardware.

Core views

Power planning usually starts before chip and rack designs are fully finalized, while cooling solutions move forward after rack density, chip type, and the air-cooling or liquid-cooling path are clarified. Insufficient grid capacity is leading hyperscale cloud providers to use more gas turbines and gas reciprocating engines as bridge or behind-the-meter power sources, and to gradually replace part of diesel backup power with lithium-ion batteries. As campuses expand from 500MW to 1GW and even above 2GW, replicable modular architecture, dual sourcing of components, and cross-region delivery capabilities become key to procurement decisions. 800VDC has long-term penetration potential, but near-term adoption remains constrained by component lead times, standards maturity, and traditional facility-side architectures.

Analysis framework

The report uses expert interviews as the main evidence and analyzes the procurement process in chronological order: project approval, campus and capex planning, main power and bridge power procurement, rack design, primary and secondary cooling procurement, pilot commissioning, and formal go-live. It also breaks down the three-layer architecture of main power, bridge power, and backup power, and compares suppliers from the perspectives of delivery performance, capacity execution, global coverage, project management, and technology differentiation. Stock target prices use EV/EBITDA, EV/EBIT, sum-of-the-parts, and P/E methods, respectively.

Methodology notes

• Industry researchExpert interview method

  Identify real procurement processes and supplier preferences through industry experts with hyperscale cloud procurement experience.

  The expert previously worked on data center infrastructure procurement at Meta and Microsoft. The interview focused on equipment lead times, technology roadmaps, project delays, and supplier execution capabilities; the related judgments are expert opinions rather than comprehensive industry order statistics.

• Operational analysisProject timeline analysis

  Break data center construction into planning, procurement, construction, commissioning, and go-live stages.

  Power equipment is usually procured before the end of the second quarter after approval; after rack architecture is finalized, CDUs and cold plates are procured before the end of the third quarter; facility-level cooling is determined in the fourth quarter, followed by another 12 to 18 months for pilots, commissioning, and testing.

• Industry chain analysisThree-layer power supply architecture

  Divide power supply needs into main power, bridge or behind-the-meter power, and backup power.

  Main power comes preferably from the grid; when the grid is constrained, gas turbines or gas reciprocating engines are used; backup power is gradually shifting from diesel generators and lead-acid batteries to lithium-ion batteries, second-life EV batteries, and other solutions.

• Supplier assessmentScaled execution capability assessment

  Evaluate equipment suppliers based on delivery speed, capacity execution, global coverage, project management, and service capabilities.

  As projects reach GW scale, buyers prefer suppliers with successful delivery track records, the ability to expand capacity across phases, and the ability to support technology migration, rather than merely comparing individual product specifications.

• Valuation methodsMultiple valuation method

  Use EV/EBITDA, EV/EBIT, sum-of-the-parts, or P/E valuation based on different companies' business structures.

  VRT, TT, and JCI primarily reference NTM+1 EBITDA multiples; NVT uses sum-of-the-parts; CARR uses EV/EBIT; ETN uses a 2030 P/E multiple discounted back to the target price.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• VRT (Vertiv Holdings Co)

  Core beneficiary of AI data center power and cooling infrastructure; Outperform, target price $368.

  Strengths

  The expert lists Vertiv as a preferred supplier after Schneider, recognizing its delivery capability, capacity support, and hyperscale project management experience.

  Weaknesses

  Near-term supply chain and staffing pressures exist, and the company is sensitive to specific computing platforms and technology architecture migration.

  Comparison

  Ranks above Eaton and Delta in supplier preference, but below Schneider.

  Risks

  Slower data center expansion, significant improvements in cooling efficiency, faster replacement by custom chips, and failure to adapt in time to direct liquid cooling or 800VDC changes.

• NVT (nVent Electric PLC)

  Beneficiary of electrical connection and cooling infrastructure; Outperform, target price $220.

  Strengths

  Has a combination of electrical and cooling businesses; Bernstein uses sum-of-the-parts valuation, reflecting the standalone value of different business segments.

  Weaknesses

  The CDU business is still in the production ramp-up stage, and recruitment cycles for some key positions may be relatively long.

  Comparison

  Compared with large integrated suppliers, its opportunities depend more on differentiated niche products and expansion of the cooling business.

  Risks

  Accelerated commoditization of CDUs and open compute platform-related products, production line execution below expectations, and supply-chain talent gaps.

• TT (Trane Technologies PLC)

  Beneficiary of facility-level cooling, chillers, and liquid cooling demand; Outperform, target price $555.

  Strengths

  Has mature cooling equipment and service capabilities and can participate in data center secondary cooling and liquid cooling growth.

  Weaknesses

  Traditional residential and transport businesses may dilute the contribution from data center growth, and the chiller business may also face periodic pressure.

  Comparison

  Compared with pure-play data center equipment companies, its business is more diversified but less sensitive to AI infrastructure growth.

  Risks

  Escalation of price-fixing litigation, more liquid cooling competitors, deterioration in other end markets, and chiller headwinds exceeding liquid cooling incremental growth.

• JCI (Johnson Controls International PLC)

  Beneficiary of data center facility-level cooling and building systems; Outperform, target price $173.

  Strengths

  Has large-scale facility equipment and service capabilities and can participate in chiller and facility-side cooling system construction.

  Weaknesses

  The investment thesis depends on organizational lean transformation and improvement in operating leverage.

  Comparison

  Has broad coverage, but AI data centers are not the only earnings driver.

  Risks

  Lean transformation fails to be implemented effectively, chiller demand weakens, and operating leverage falls below management guidance.

• CARR (Carrier Global Corporation)

  Potential beneficiary of cooling and HVAC equipment; Market-Perform, target price $78.

  Strengths

  Continued data center growth and progress in the CDU business could create additional upside.

  Weaknesses

  U.S. residential and light commercial cycles remain more important earnings drivers, while the certainty of the data center contribution is relatively limited.

  Comparison

  Rated below the other five covered companies in the report, reflecting a more balanced current risk-reward profile.

  Risks

  Slowdown in hyperscale cloud capex, escalation of price-fixing litigation, and later-than-expected improvement in refrigerant supply and demand.

• ETN (Eaton Corp PLC)

  Core beneficiary of power distribution, electrical equipment, and data center power expansion; Outperform, target price $534.

  Strengths

  Power capacity has become the primary constraint, and the expert includes Eaton among the leading suppliers preferred by hyperscale cloud providers.

  Weaknesses

  The valuation is based on assumptions of long-term structural growth and relatively high earnings resilience, requiring strong growth delivery.

  Comparison

  Ranks behind Schneider and Vertiv in supplier preference, but has direct exposure to electrical equipment expansion.

  Risks

  Load growth, utility capex, or reshoring/manufacturing trends weaker than expected, as well as disruptions to global data center construction or cancellations of large projects.

• Schneider Electric

  The expert's highest-rated hyperscale data center power and infrastructure supplier, and an important competitive benchmark for the covered names.

  Strengths

  Outstanding delivery speed, capacity execution, global coverage, and customer understanding.

  Weaknesses

  The report does not provide its rating, target price, or detailed valuation.

  Comparison

  The expert rates it ahead of Vertiv, Eaton, and Delta.

  Risks

  Potential risks mainly come from a slowdown in industry capex, supply chain constraints, and changes in customer procurement architecture.

Key data

• Overall project cycle24 to 30 monthsFrom concept approval to formal go-live for production traffic.
• Pilot, commissioning, and testing cycle12 to 18 monthsOccurs after major equipment procurement and facility construction.
• Grid interconnection cycle for large projects24 to 36 monthsApplies to deployments of roughly 50 to 100MW or above.
• Period of tight utility capacityThrough 2028 to 2029The expert said available capacity at multiple utilities is largely fully booked.
• Common schedule delayApproximately 20%May lead to additional capex and commissioning overruns.
• Potential delay for GW-scale projects6 to 9 monthsDoes not yet include additional delays caused by regulatory or compliance issues.
• Gas turbine delivery cycle18 to 24 monthsHyperscale cloud providers may reserve supplier capacity in advance.
• Gas turbine costApproximately 2x grid powerDespite higher costs, it is usually a more efficient behind-the-meter power option.
• Potential 800VDC penetration rateApproximately 30% to 40% by 2030Actual adoption speed remains constrained by component supply, standards, and facility-side retrofits.
• Cycle from advanced custom chip design to rack18 to 20 monthsThe expert's estimate for 2 to 3 nanometer chips.
• Cooling procurement start timeApproximately 6 months after approvalUsually starts after rack architecture and facility design have a relatively high degree of certainty.
• Example campus expansion path500MW to 1GW, then to above 2GWIllustrates the importance of modular and replicable architecture for multi-phase expansion.

Impact & implications

Industry value may further shift toward power equipment, behind-the-meter generation, power distribution, liquid cooling, facility-level cooling, and engineering delivery. Suppliers with global manufacturing capabilities, shorter lead times, complex project management experience, and after-sales service networks are more likely to enter preferred lists and win orders in subsequent phases. Dual-supplier strategies do not necessarily lead to order cancellations: because infrastructure has a long service life and limited value decay, excess deliveries may translate into a larger actual construction scale. Conversely, suppliers that rely on highly customized solutions, a single technology path, or insufficient delivery capabilities face risks of share loss, price commoditization, and customer lock-in concerns.

Risks

• Hyperscale cloud providers' AI capex or data center expansion pace is lower than expected.
• Insufficient grid generation, transmission, and distribution capacity causes project delays and cost overruns.
• Lead times for transformers, solid-state transformers, gas turbines, and cooling components remain extended.
• Higher energy efficiency reduces cooling and electrical equipment demand per unit of compute.
• Migration from GPUs to custom chips changes rack, liquid cooling, and supplier competitive dynamics.
• CDUs, cold plates, and open compute platform-related products accelerate commoditization, compressing supplier margins.
• 800VDC standards mature and adoption proceed more slowly than expected, or suppliers fail to adapt in time.
• Highly customized solutions lead to supplier lock-in, premiums, and lower cross-generation compatibility.
• Supplier staffing, capacity ramp-up, or complex project management capabilities are insufficient.
• Litigation, regulatory approvals, and compliance issues create additional schedule and cost risks.

What to watch

• New generation, transmission, and distribution capacity and queue times at major U.S. utilities.
• The share of projects above 200MW adopting gas turbines, gas reciprocating engines, and other behind-the-meter power solutions.
• Orders and delivery cycles for gas turbines, transformers, solid-state transformers, CDUs, cold plates, and chillers.
• Maturity progress of 800VDC, sidecar power architectures, and related industry standards.
• The pace at which lithium-ion batteries and second-life EV batteries replace diesel backup power.
• The pace of data centers shifting from air cooling to liquid cooling and of increases in rack power density.
• Share and delivery performance of Schneider, Vertiv, Eaton, Delta, ACT, CoolIT, and Boyd.
• Whether dual-supplier orders from hyperscale cloud providers translate into above-plan capacity expansion.
• VRT's supply chain and staffing recovery, and NVT's CDU production line ramp-up.
• Whether data center cooling revenue at TT, JCI, and CARR can offset pressure in traditional end markets.