AI semiconductor to data-center power infrastructure supply chain: UBS sees enough potential power to support AI infrastructure through 2030, but execution is crucial

Overview

This UBS thematic AI-infrastructure study connects semiconductor shipment assumptions to the physical requirements for data-center construction and electricity supply. UBS concludes that power availability can support its accelerator estimates through 2030, but only if early-stage grid projects and behind-the-meter generation are delivered at scale.

Core views

UBS begins with semiconductor and accelerator-roadmap assumptions and translates them into electricity demand. Rising accelerator density, more dies per package, and higher thermal design power per rack mean that gigawatt requirements can continue increasing even when rack volumes decline. UBS estimates approximately 340GW of cumulative incremental compute capacity between 2026 and 2030; applying a 1.12 power-usage-effectiveness assumption raises the implied data-center facility requirement to approximately 380GW. Annual deployed compute capacity is projected to rise from approximately 10GW in 2024 to approximately 105GW in 2030. The report expects merchant GPUs to provide approximately 200GW, or 59%, of the incremental 340GW of compute capacity through 2030, while ASICs contribute approximately 140GW, or 41%. The compute mix shifts toward ASICs, from 65% GPU/35% ASIC capacity in 2026E to 57%/43% by 2030. However, UBS estimates that merchant GPUs retain about 70% of data-center revenue opportunity versus 30% for ASICs because GPUs have broader workloads, more system-level content and more software monetization. NVIDIA's deployed compute capacity is projected to grow at a 32% CAGR from approximately 17GW in 2026 to approximately 51GW in 2030; AMD grows at a 90% CAGR from approximately 1GW to approximately 9GW, albeit from a much smaller base. UBS ties these projections to supply-chain capacity. It expects industry CoWoS advanced-packaging capacity to rise from 160,000 wafers per month at the end of 2026 to 270,000 at the end of 2027, led by TSMC, ASE and Amkor. UBS assumes NVIDIA has supply for approximately 14.3 million chips in 2027 and AMD approximately 7.0 million, alongside sizable ASIC supply for TPU, Trainium, OpenAI, Meta and Microsoft programs. The report notes that higher-density architectures change the relationship between units, racks and power: NVIDIA's anticipated Feynman configuration reaches approximately 1.2MW per rack, while AMD's MI500 MegaPod is expected to reach approximately 1MW per rack. Data-center shell visibility is currently below the capacity implied by those semiconductor assumptions. UBS Evidence Lab identifies approximately 276GW of global data-center capacity under construction or planned, enough for roughly 70-75% of the implied 380GW requirement. UBS argues that the pipeline is moving higher rapidly: it increased by approximately 75GW, or approximately 45% quarter-on-quarter, between 1Q26 and 2Q26. The United States accounted for 80% of that revision and 70% of global planned or under-construction capacity. UBS believes projects announced during the next one to two years could become operational by 2029-30, although approximately 100-105GW of additional shells would be needed to close the visible gap. 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 timeframe because of permitting, power procurement, local opposition and labor constraints. Power is the more binding condition. UBS estimates approximately 307GW of global incremental data-center capacity can be grid-supported between 2026 and 2030, including approximately 271GW in the US, approximately 16GW in Europe and approximately 20GW elsewhere. The US utility base case provides only approximately 76GW; the balance relies on approximately 24GW from higher plant completion rates, approximately 151GW from early-stage projects and approximately 20GW from fewer retirements. UBS believes approximately 156GW of usable capacity from alternative sources could support the early-stage-project upside, but notes that the US reserve margin declines to 15% by 2030. China could add approximately 30-40GW of on-grid data-center capacity, but is excluded because US accelerator shipments into China are constrained by regulation. Behind-the-meter generation is UBS's key relief valve because it can bypass grid-connection delays. UBS estimates approximately 108GW of incremental capacity from such solutions, mainly approximately 90GW from Caterpillar and up to approximately 18GW from Innio; Cummins represents potential additional upside as its prime-power platform ramps later. Combining approximately 307GW of grid supply and approximately 108GW of behind-the-meter supply yields approximately 415GW of potential powered-shell capacity, modestly above UBS's approximately 380GW requirement. This conclusion requires about 150GW of early-stage on-grid projects to be completed and connected, as well as broad deployment of behind-the-meter resources. The near-term balance is tighter than the 2030 aggregate view. UBS projects approximately 57GW of data-center capacity demand in 2027 but sees only approximately 38GW of powered shell excluding early-stage plants; for 2028, it estimates approximately 76GW of demand against approximately 50GW of visible powered shell. With approximately 50GW of early-stage grid projects planned in each year, a 40-50% conversion rate would be required to bridge the gaps. UBS considers that achievable but dependent on accommodating legislation and limited supply-chain delays. UBS identifies legislation and interconnection as material constraints. New York has enacted a one-year moratorium on new data centers larger than 20MW while it conducts environmental studies, while Texas is reviewing data-center interconnection proposals. Grid interconnection-to-commercial-operation timelines now exceed four years. Critical equipment lead times include up to 210 weeks for large power transformers, approximately 18 months for backup-power systems and switchgear, and more than three years for on-site gas turbines. These constraints could delay or cancel projects and create mismatches between semiconductor-driven compute demand and available power. Finally, UBS applies its gigawatt and revenue-per-gigawatt assumptions to NVIDIA. Its framework implies approximately $408B of NVIDIA data-center revenue in CY26E, rising to approximately $1.1T in CY28E. While 2026E is broadly consistent with UBS's existing model, the widening 2027-28 gap suggests that its model may not fully capture the potential scale of deployed compute infrastructure, partly because the HGX-versus-NVL72 mix affects NVIDIA revenue per gigawatt. UBS argues that integrated AI systems allow NVIDIA to capture a greater portion of infrastructure spending per deployed gigawatt.

Analysis framework

UBS works backward from projected GPU and ASIC shipments, rack configurations and rack power requirements to calculate compute gigawatts, then converts those requirements into data-center shell demand using a PUE of 1.12. It compares the resulting capacity need with planned data-center shells, grid-supported power scenarios and behind-the-meter equipment capacity, before applying revenue-per-gigawatt assumptions to NVIDIA.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Semiconductor-to-data-center-to-power supply-chain mapping

  The report traces how accelerator volumes and rack designs translate into data-center construction needs and, ultimately, grid and on-site generation requirements.

• Industry AnalysisSupply-demand framework

  Comparison of implied data-center demand with shell and power availability

  UBS compares required facility capacity with visible project pipelines, grid capacity scenarios and behind-the-meter supply to assess whether infrastructure can support projected compute deployment.

• Industry AnalysisVolume-price decomposition

  Gigawatt deployment and revenue-per-gigawatt analysis

  The report separates physical compute capacity from revenue opportunity, showing why GPUs can retain a larger revenue share despite ASICs gaining capacity share.

Asset mapping & comparison

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

• NVIDIA (NVDA)

  Primary merchant-GPU beneficiary of growing AI compute capacity and integrated-system spending.

  Strengths

  UBS projects deployed compute capacity rising from ~17GW in 2026 to ~51GW in 2030 and sees higher revenue content per deployed GW.

  Weaknesses

  Revenue per GW is affected by the mix of HGX racks and NVL72 systems.

  Comparison

  Merchant GPUs are projected to account for 59% of incremental compute capacity but about 70% of data-center revenue versus ASICs.

  Risks

  Dependent on power availability, data-center construction, customer monetization and supply-chain execution.

• Advanced Micro Devices (AMD)

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

  Strengths

  UBS projects compute capacity to rise from ~1GW in 2026 to ~9GW in 2030, with a 90% CAGR from a small base.

  Weaknesses

  Smaller initial capacity base than NVIDIA.

  Comparison

  UBS estimates ~1.3 million GPU units in CY27 and ~1.8 million in CY28 as MI400 and MI500 ramp.

  Risks

  Execution of accelerator ramps and the same data-center-power constraints affecting the broader market.

• Caterpillar

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

  Strengths

  UBS estimates approximately 90GW of incremental data-center shell capacity could be supported by its equipment through 2026-30.

  Weaknesses

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

  Comparison

  UBS identifies Caterpillar as the largest modeled behind-the-meter contributor, ahead of Innio.

  Risks

  Upside depends on a greater allocation of equipment capacity toward data centers.

• Innio

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

  Strengths

  Approximately 80% of its current data-center equipment backlog is tied to prime power.

  Weaknesses

  Its modeled contribution is substantially smaller than Caterpillar's.

  Comparison

  UBS estimates up to approximately 18GW of cumulative installed base for data-center behind-the-meter solutions through 2026-30.

  Risks

  Capacity expansion and greater allocation to data centers are required to achieve upside.

Key data

• Incremental compute capacity, 2026-30~340GWUBS estimate across GPUs and ASICs
• Implied data-center shell capacity, 2026-30~380GWBased on a PUE of 1.12
• Potential global powered-shell capacity~415GW~307GW grid support plus ~108GW behind-the-meter supply
• Visible global data-center capacity~276GWUnder construction or planned; implies ~100-105GW additional shell requirement
• 2027 powered-shell gap~19GW~57GW implied capacity versus ~38GW visible excluding early-stage projects
• 2028 powered-shell gap~26GW~76GW implied capacity versus ~50GW visible excluding early-stage projects
• Implied NVIDIA data-center revenue~$408B in CY26E to $1.1T in CY28EUBS power-based framework

Impact & implications

UBS argues that the available power pipeline can support its AI semiconductor outlook through 2030, preserving potential upside to semiconductor estimates. The conclusion is conditional on conversion of early-stage power projects, expansion of behind-the-meter generation, sufficient data-center construction and the avoidance of regulatory or equipment-delivery disruptions.

Risks

• State and local restrictions, including moratoriums and stricter cost-allocation rules, could delay or cancel data-center development.
• Early-stage grid projects may fail to obtain permits, be completed, or connect to the grid at the required rate.
• Grid interconnection timelines exceeding four years could cause a mismatch between available power and projected compute deployment.
• Long lead times for transformers, switchgear, backup systems and gas turbines could constrain buildouts.
• Power procurement, local opposition, labor availability and construction delays could extend greenfield data-center timelines.
• Semiconductor and data-center demand depends on hyperscaler monetization efforts and continued infrastructure investment.

What to watch

• Conversion rates for approximately 50GW of early-stage grid projects planned in each of 2027 and 2028.
• Growth in planned and under-construction data-center capacity, particularly in the United States.
• Grid reserve margins, plant completion rates, retirements and alternative generation additions.
• Allocation of Caterpillar, Cummins and Innio equipment capacity toward data-center prime-power applications.
• US state and local data-center legislation, permitting and interconnection policy.
• Lead times for transformers, switchgear and on-site gas turbines.