AI data-center semiconductor, powered-shell and electricity infrastructure supply chain: UBS sees a path to power AI data-center demand through 2030, but execution of grid projects and behind-the-meter supply is critical.

Overview

This thematic UBS study tests whether the semiconductor-led AI infrastructure buildout can be physically supported by data-center construction and power availability. UBS concludes that there is theoretical capacity to support its above-Street accelerator assumptions through 2030, but the conclusion relies on substantial execution across grid, construction and on-site power supply chains.

Core views

UBS begins with accelerator roadmaps and converts expected GPU and ASIC shipments into rack-level power demand. Rising die counts, accelerator density and thermal design power mean that gigawatts deployed can increase even when rack volumes decline. The firm estimates about 340GW of cumulative incremental compute capacity between 2026 and 2030, which translates into roughly 380GW of data-center facility, or powered-shell, capacity using a PUE of 1.12. Annual compute capacity rises from about 10GW in 2024 to about 105GW in 2030. Merchant GPUs account for about 200GW, or 59%, of the incremental compute capacity by 2030, while ASICs account for about 140GW, or 41%. The report expects the deployment mix to shift toward ASICs, from 65% GPU and 35% ASIC capacity in 2026E to 57% and 43%, respectively, by 2030. However, UBS argues that merchant GPUs retain a disproportionate revenue share because of wider workload applicability, greater system-level content and software monetization. Its dollar-per-gigawatt analysis therefore implies roughly 70% of data-center revenue for merchant GPUs and 30% for ASICs. UBS forecasts NVIDIA compute capacity to rise at a 32% CAGR from about 17GW in 2026 to about 51GW in 2030; AMD rises at a 90% CAGR from about 1GW to about 9GW, albeit from a far smaller base. Google and AWS are the largest ASIC capacity contributors in UBS's assumptions. Physical shell construction is the first constraint. UBS Evidence Lab identifies about 276GW of global data-center capacity under construction or planned, enough for only about 70-75% of the 380GW implied by its accelerator estimates. The pipeline rose by about 75GW, or roughly 45% quarter-on-quarter, between 1Q26 and 2Q26, with Pennsylvania, Ohio and West Virginia showing the largest revisions; the United States accounts for 80% of the recent increase and 70% of visible global planned or under-construction capacity. UBS believes another 100-105GW of unannounced shells would be needed, but notes that projects announced over the next one to two years could still be operating by 2029-30. A greenfield data center takes at least 15 months after land and power procurement to go live, while 18-24 months is a more conservative assumption when power procurement, local opposition and labor availability create delays. Power is the more important bottleneck in UBS's analysis. The firm estimates approximately 307GW of global grid-supported data-center capacity during 2026-30: about 271GW in the United States, 16GW in Europe and 20GW elsewhere. The US base case supplies only about 76GW, while the larger 271GW figure requires about 24GW from higher plant-completion rates, 151GW from early-stage unpermitted projects and about 20GW from fewer plant retirements. UBS estimates that alternative power additions could provide about 156GW in a full-absorption case, sufficient to support the 151GW early-stage-project upside case. China could add another 30-40GW of on-grid capacity but is excluded because US accelerator exports to China are constrained. Behind-the-meter generation is the balancing mechanism because grid capacity alone falls short of implied shell demand. UBS estimates about 108GW of additional behind-the-meter support from Caterpillar and Innio, with potential upside from Cummins. Caterpillar contributes an estimated 90GW based on equipment-capacity expansion and allocation toward data-center prime power, while Innio contributes up to 18GW if it raises capacity by 1GW annually. Together, the report's 307GW grid estimate and 108GW behind-the-meter estimate produce about 415GW of potential power availability, modestly above the 380GW required by the compute forecast. The near-term balance is tighter. For 2027, UBS estimates about 57GW of implied data-center capacity versus only about 38GW of powered-shell visibility excluding early-stage plants. For 2028, the comparison is about 76GW versus 50GW. Each year has roughly 50GW of early-stage grid projects in the pipeline, so approximately 40-50% must convert into connected supply. UBS considers this achievable but demanding, requiring accommodating legislation and limited supply-chain delays; behind-the-meter equipment could bridge part of the gap by being reallocated from other end markets. UBS flags regulatory and equipment constraints as the principal downside to this path. State and local restrictions can delay or cancel data-center projects: the report cites New York's one-year moratorium on new data centers above 20MW while environmental studies are conducted, and Texas's review of grid-interconnection projects. Grid interconnection-to-operation timelines have extended to more than four years. Large power-transformer lead times can reach 210 weeks, backup-power systems and switchgear about 18 months, and on-site gas turbines more than three years. These bottlenecks can create a mismatch between available power and implied compute deployment. Finally, UBS applies its gigawatt and dollar-per-gigawatt framework to NVIDIA. The framework implies roughly $408B of NVIDIA data-center revenue in CY26E, rising to $1.1T in CY28E. CY26E is broadly aligned with the existing model, but UBS sees a widening gap in CY27-28, partly because the model's HGX-versus-NVL72 mix affects NVIDIA revenue per gigawatt. UBS argues that increasingly integrated AI clusters should enable NVIDIA to capture more infrastructure spend per deployed gigawatt while improving the amount of AI work delivered within a fixed power envelope.

Analysis framework

UBS works backward from expected GPU and ASIC shipments, rack configurations and thermal-design-power assumptions to derive compute gigawatts and then converts them to facility capacity using a PUE of 1.12. It compares that demand with visible data-center shells, grid-supplied capacity and behind-the-meter equipment capacity, then applies the resulting gigawatt assumptions and revenue-per-gigawatt content to its NVIDIA model.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI-compute capacity demand versus data-center shell and power supply

  UBS converts semiconductor deployment assumptions into facility-power requirements and tests whether construction, grid capacity and on-site generation can meet them.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Semiconductor shipments, rack power density, data-center construction and electricity infrastructure

  The report traces how accelerator roadmaps create demand for racks, data-center shells, grid connections and behind-the-meter power equipment.

• Other

  Power usage effectiveness conversion and revenue-per-gigawatt framework

  UBS uses a PUE of 1.12 to translate compute load into required facility capacity, then combines gigawatt estimates with NVIDIA content per gigawatt to derive implied revenue.

Asset mapping & comparison

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

• NVIDIA (NVDA)

  Primary semiconductor beneficiary in UBS's power-based AI infrastructure framework.

  Strengths

  UBS expects rising integrated-system content and revenue per deployed gigawatt; implied revenue reaches ~$408B in CY26E and $1.1T in CY28E.

  Weaknesses

  Revenue per gigawatt depends on the assumed mix between HGX and NVL72 architectures.

  Comparison

  Merchant GPUs retain roughly 70% of implied data-center revenue despite ASICs gaining capacity share.

  Risks

  Power-project completion, data-center construction, permitting and supply-chain bottlenecks could curb compute deployment.

• AMD (AMD)

  Merchant GPU participant in UBS's accelerator-capacity forecast.

  Strengths

  UBS estimates compute capacity rising at a 90% CAGR from ~1GW in 2026 to ~9GW in 2030.

  Weaknesses

  Growth begins from a much smaller base than NVIDIA.

  Comparison

  UBS estimates merchant GPUs represent 59% of incremental compute capacity through 2030.

  Risks

  Deployment depends on the same powered-shell and supply-chain conditions affecting the broader accelerator market.

• Caterpillar

  Supplier of behind-the-meter prime-power equipment for data centers.

  Strengths

  UBS estimates ~90GW of incremental data-center shell capacity could be supported by Caterpillar equipment between 2026 and 2030.

  Weaknesses

  Capacity is shared with mining, oil and gas and other end markets.

  Comparison

  Caterpillar is UBS's largest estimated behind-the-meter contributor, ahead of Innio.

  Risks

  Data-center allocation and the mix of prime-power applications determine realized contribution.

• Innio

  Supplier of behind-the-meter data-center prime-power equipment.

  Strengths

  About 80% of current data-center-equipment backlog is tied to prime power; UBS estimates upside to ~18GW of cumulative installed base.

  Weaknesses

  Estimated contribution depends on execution of 1GW annual capacity additions.

  Comparison

  UBS estimates Innio can complement Caterpillar's larger behind-the-meter contribution.

  Risks

  Equipment-capacity expansion and allocation to data centers are required.

Key data

• Cumulative incremental compute capacity~340GWUBS estimate for 2026-2030.
• Implied data-center shell capacity~380GW2026-2030 requirement using a PUE of 1.12.
• Potential total power availability~415GW~307GW grid support plus ~108GW behind-the-meter capacity.
• Visible global data-center pipeline~276GWUnder construction or planned; supports about 70-75% of UBS's implied accelerator capacity.
• 2027 implied capacity versus powered-shell visibility~57GW versus ~38GWExcludes early-stage power plants from visible powered-shell capacity.
• 2028 implied capacity versus powered-shell visibility~76GW versus ~50GWRequires early-stage projects or behind-the-meter supply to bridge the gap.
• NVIDIA implied data-center revenue~$408B in CY26E and $1.1T in CY28EDerived from UBS's power-based gigawatt framework.

Impact & implications

UBS argues that power availability does not necessarily invalidate semiconductor AI-demand estimates through 2030, but turns completion of early-stage power projects, data-center construction and behind-the-meter equipment allocation into critical determinants. Its framework indicates potential upside to NVIDIA's modeled data-center revenue trajectory, particularly in CY27-28.

Risks

• State and local data-center legislation, moratoriums and stricter cost-allocation rules could delay or cancel projects.
• A low conversion rate for early-stage grid projects would leave powered-shell capacity below implied accelerator demand in 2027 and 2028.
• Grid interconnection delays, now extending beyond four years, could postpone usable power supply.
• Long lead times for transformers, switchgear, backup systems and gas turbines could constrain buildouts.
• Power procurement, local community opposition and labor shortages could extend greenfield data-center construction beyond UBS's base timing assumptions.

What to watch

• Conversion of roughly 50GW of early-stage grid projects in each of 2027 and 2028.
• The pace of new data-center project announcements needed to close the visible 100-105GW shell-capacity gap.
• US grid capacity additions, retirements and reserve-margin trends.
• Allocation of Caterpillar, Innio and potentially Cummins equipment toward data-center prime-power uses.
• State and local permitting, interconnection and data-center legislation.
• The mix of NVIDIA HGX and NVL72 systems, which affects revenue per gigawatt.