BloombergNEF raised its 2030 US data center capacity forecast to 118 GW on July 21, 2026, a 52% upward revision from its December 2025 outlook, with a 2035 figure of 194 GW representing an 83% jump [S3].
Goldman Sachs Research models global data center power demand climbing about 50% to roughly 84 GW by 2027 and as much as 165% by the end of the decade versus 2023, with AI workloads shifting from about 14% of today's consumption to 27% by 2027 [S2]. Cross-checking against ABI Research's regional build, US AI-dedicated active capacity is forecast to grow from 8.2 GW in 2026 to 26.4 GW by 2031, more than tripling in five years [S4].
Headline numbers across the three credible 2026 forecasts
Goldman Sachs Research's baseline scenario puts 2030 global data center power demand at roughly 84 GW (a 50% rise by 2027 from ~55 GW today) and as much as 165% above 2023 by 2030, with AI moving from 14% of the load today to about 27% by 2027 [S2]. The same bank's scenario tree brackets 2030 outcomes at 14% CAGR in a muted case, 17% baseline, and 20% in an upside case, producing a 9-13 GW swing on either side of baseline [S2].
BNEF's July 2026 update is more aggressive on US installed megawatts: 118 GW by 2030 and 194 GW by 2035, with a separate AI-chip-demand scenario implying 325 GW of global capacity additions between 2023 and 2033 and 207 GW of that landing in the US [S3]. The 63 GW gap between BNEF's US forecast for 2033 and what AI chip shipments imply is being treated by analysts as a binding power-and-energy constraint, not a market-size constraint [S3].
ABI Research's country-level model tracks active AI-dedicated capacity specifically, not total facility power: US 8.2 GW in 2026 to 26.4 GW in 2031, China 3.1 GW to 10.3 GW, Europe 3.5 GW to 9.7 GW, and Japan 1 GW to 2.6 GW, with South Korea nearly quadrupling from 599 MW to 2.5 GW [S4]. For US peak summer grid share, the Marcus/Goldman Sachs report shows data center load rising from 4.1% of total US peak summer demand in 2025 to 5.3% in 2026 and 8.5% in 2027 [S5].
AI training vs inference: how the workload mix shifts the build
Goldman Sachs Research expects inference to overtake training as the dominant AI workload in data centers by 2027, with AI then handling roughly half of all data center workloads by 2030 [S2]. The shift changes the deployment pattern: training is bursty and concentrated in a few hyperscaler mega-sites, while inference needs regional distribution and edge presence to keep latency down, which is the main reason neocloud and sovereign operators are scaling outside the US [S2][S4].
Power density is the second-order effect. Average AI data center density is projected to climb from 162 kW per square foot today to 176 kW per square foot by 2027, which breaks the conventional hot-aisle/cold-aisle air-cooling envelope and forces direct-liquid or rear-door heat-exchanger designs on new builds [S1]. Operators that have not retooled their mechanical rooms will see per-rack headroom cap out well below the GPU-roadmap requirements spelled out in the AI chip demand 2026-2030 ramp curve discussion.
Who is building: hyperscalers vs the new 150-project cohort
BNEF's split-out of developer type shows hyperscaler capacity (Microsoft, Meta, Alphabet, AWS) tripling over the next decade while non-hyperscaler capacity nearly quintuples, reaching 138 GW versus 57 GW for hyperscalers by 2035 [S3]. Of the 100 largest US projects tracked, 52 are being announced by companies that have never built a data center before, accounting for 87 GW of announced capacity and an outsized share of execution risk [S3].
The US-side $500 billion Stargate Project (OpenAI, Microsoft, Oracle, SoftBank) targets 10 GW of dedicated power capacity, anchoring the hyperscaler bucket [S4]. US construction spend has tripled in the last three years, and occupancy rates remain near record highs in most US markets even as new capacity lands [S1]. The Stargate commitments and the giga-campus announcements are the upstream signal that NVIDIA's continued accelerator-stack dominance in 2026 is being built into physical real estate, not just paper allocations.
Regional split: where the megawatts actually land
North America and Asia Pacific hold most of the installed power and floor area today, with hyperscale and wholesale operators controlling about 60% of global capacity and the remaining 40% split across corporate and telecom-owned sites [S2]. ABI Research's regional table makes the geographic rebalancing explicit: by 2031, the US still leads on absolute AI-dedicated capacity (26.4 GW), but Rest-of-Asia-Pacific reaches 8.9 GW and Europe 9.7 GW, with the Middle East and Africa each growing 4-6x from a small base [S4].
Sovereign-AI policy is a real siting driver in China, Europe, South Korea, the Middle East, and parts of Africa, with data-residency rules pushing governments to fund domestic capacity rather than rent overseas hyperscaler regions [S4]. Practical consequence for suppliers: liquid-cooling distribution, flow-meter skids, and pressure-sensor instrumentation on secondary loops are now being specified at the campus, not the rack, level on new sovereign builds. The same logic applies to industrial-valve packages for chilled-water and dielectric-fluid distribution.
Power and grid: the binding constraint, not capex
US data centers' share of national peak summer power demand is projected to climb from 4.1% in 2025 to 5.3% in 2026 and 8.5% in 2027, a 4.4 percentage-point jump in two years that no other load class matches [S5]. BNEF's central estimate now sits 63 GW below what AI chip shipments alone would imply for the US by 2033, which is the cleanest single number in the research for the size of the energy bottleneck [S3].
Inside the campus, operators are responding with on-site generation, fuel cells, and behind-the-meter PPAs; Bloom Energy's 2026 Power Report documents an increasing share of campuses exceeding gigawatt scale and the corresponding shift in pressure-transmitter and data-logger requirements on the gas and steam side [S6]. Goldman Sachs Research's caveat is worth carrying forward: efficiency gains like DeepSeek's reported training-cost reductions could shift the curve down by 9-13 GW, but the baseline still points to a tight supply-demand balance through 2026 with occupancy projected to peak above 95% in late 2026 before moderating in 2027 [S2].
What this means for industrial procurement and plant engineers
For anyone sizing PLC cabinets, chiller skids, or bus-duct for a 100 MW+ campus, three numbers from the research set the spec floor: 162 kW/sq ft today rising to 176 kW/sq ft by 2027, hyperscaler capacity tripling to 57 GW by 2035, and non-hyperscaler capacity nearly quintupling to 138 GW in the same window [S1][S3]. Equipment that is rated only for legacy 8-12 kW-per-rack envelopes is already obsolete for greenfield AI builds, and lead times on 1500V DC bus-duct and liquid-cooling CDUs are the practical gating items behind the announced project dates.
The two near-term tracking signals worth watching are BNEF's next US pipeline update, where the December 2025 to July 2026 jump added 101 GW of announced capacity, and any quarterly revision to the Goldman Sachs baseline that moves the 2030 demand band outside the current 9-13 GW muted-to-upside range [S3][S2]. The 2035 hyperscaler/non-hyperscaler split (57 GW vs 138 GW) is the cleanest forward indicator of whether the new-entrant developer cohort is converting announcements into energized capacity on schedule.