The expansion of data-center infrastructure, propelled by artificial-intelligence workloads and broader digitalization, is confronting the physical limits of electricity grids, water systems, and local communities. Global data-center electricity consumption is estimated at approximately 415 TWh annually 5,11,14,51,52,54, with estimates ranging from 240–340 TWh in 2022 47 to 447 TWh in 2025 23. Most forecasts point to at least a doubling by 2030. The International Energy Agency projects consumption of roughly 945 TWh 2,3,4,5,6,7,8,9,10,16,17,33,51,52,54,59, while other estimates reach 1,050 TWh by 2026 14,47 and as much as 1,200 TWh by 2035 43.
This is not merely an infrastructure question. It is a constraint on the pace and sustainability of future compute deployment—and therefore a material consideration for NVIDIA, whose GPUs underpin the majority of AI compute.
The Scale and Concentration of Demand
In the United States, data-center electricity consumption tripled from 58 TWh in 2014 to 176 TWh in 2023 32, representing approximately 4–5% of total U.S. electricity use 1,13,32,49. Projections for 2028 range from 325 to 580 TWh, equivalent to 6.7–12% of national electricity consumption 10,12,15,32,33,42. Growth is also accelerating: the annual rate increased from 12% to 17% in 2025 51,54,59, while AI-specific data-center demand could expand by more than 30% annually through 2030 35.
At the global level, data centers still account for only approximately 1.5% of electricity consumption 14,23,51,54. Moreover, projected incremental demand from 2024 through 2030 remains below the combined increase expected from industry, electric vehicles, and air conditioning 23. These figures provide useful perspective, but they do not dissolve the engineering problem. A load need not dominate the global system to overwhelm a local one. In Texas, for example, data centers may represent as much as 90% of proposed new load on the ERCOT grid 53. The local concentration of demand therefore matters more than the global average.
Water as a Parallel Constraint
Electricity is only one side of the balance. Data centers require substantial freshwater for cooling 24,25, and the volume consumed depends on facility size and cooling technology 28. Facilities in drought-prone regions face heightened scrutiny 30,52. Critics further argue that cooling requirements can shift costs to households through higher water bills 24.
Thus, the relevant site-selection equation is not simply land plus power. Water availability, cooling design, permitting, and community tolerance are coupled variables. Consider the circuit: changing one branch alters the operating conditions of the whole network. Data-center development is no different.
Grid Capacity, Reliability, and Cost Allocation
Power availability, grid capacity, and transmission bottlenecks are increasingly viewed as the principal determinants of project feasibility, often preceding land or hardware availability 28,40,46,55. Grid operators face growing reliability challenges 41,45,49, while utilities must invest in generation, transmission, and system resilience. Those costs may ultimately be passed to all ratepayers 45,49,56.
The concentration of very large loads in particular regions creates disproportionate local effects 27,32 and introduces potential cascade risks across the sector 50. A grid should not be treated as a simple bus to which one may attach unlimited load. Every interconnection has impedance, every transmission path has a limit, and every large step change produces a transient response. Whether a constraint is truly negligible depends on the operating conditions; in these systems, that assumption is increasingly difficult to defend.
Fossil fuels currently supply much of the incremental electricity demand 51,54. This reliance can undermine corporate climate goals and increase carbon emissions 21,44,54, intensifying environmental and social concerns 39,48,57. Governments and communities are consequently scrutinizing data-center energy and water consumption, grid impacts, and permitting more closely 26,39. These pressures may slow projects, increase costs, or require additional investment in renewable generation and supporting infrastructure.
Implications for NVIDIA
NVIDIA’s data-center GPU business is directly exposed to the ability of cloud providers and enterprises to power and cool their facilities. Electricity access is now identified as a primary bottleneck, in some cases a greater constraint than chip supply 22,36. If power or water scarcity slows data-center construction, demand for high-performance accelerators may be capped even when customer demand for compute remains strong.
The same constraint, however, strengthens the economic case for efficient compute. Rack densities are moving toward 10 kW and beyond 29, while accelerator thermal design power can exceed 1,000 watts 31. Energy costs are among the largest operating expenses for data centers 18,58, so improvements in performance per watt affect both environmental performance and project economics 33. NVIDIA’s architectural emphasis on efficiency—including the Grace Hopper Superchip and liquid-cooled Blackwell systems—therefore positions the company as an enabler of more sustainable scaling.
This is the favorable side of the constraint. When power is scarce, an efficient computation is not merely a greener computation; it is a more valuable one. The practical advantage is measured in useful work delivered per unit of electrical and thermal capacity.
Rising Power Density and System-Level Demand
The movement toward higher rack densities, liquid cooling, and DC power distribution 34,37,60 expands the market for power, thermal, and interconnect solutions 38,61. It also reinforces NVIDIA’s system-level position. GPUs, networking products, and software are increasingly integrated into the compute fabric of next-generation, power-dense data centers.
This development benefits adjacent suppliers and ecosystem partners, but it does not insulate GPU sales from an industry-wide slowdown if power constraints become acute. The circuit remains whole: efficient accelerators can reduce the energy required per unit of computation, but they cannot by themselves create transmission capacity, water availability, or permitting approval.
Regulatory and Social-License Risk
Hyperscalers’ reliance on fossil-fuel backup and the indirect emissions associated with purchased electricity 54 could attract regulatory mandates. Such requirements may increase costs or compel further investment in renewable generation, influencing site selection and potentially delaying deployments. NVIDIA’s customers, including Meta and Alphabet, explicitly identify data-center electricity consumption and sustainability risks 19,20, demonstrating that the exposure extends across the industry.
Community opposition, regulatory scrutiny, and legal challenges related to resource consumption therefore represent a growing social-license risk. NVIDIA and its customers may need to engage earlier in site-selection decisions, prioritizing energy and water availability alongside the traditional criteria of land, connectivity, and hardware access.
Practical Implications
NVIDIA’s growth trajectory is inseparable from the buildout of power-intensive AI data centers. Electricity availability and grid reliability could moderate GPU adoption in the medium term, even as demand for AI compute continues to rise.
Energy efficiency has become a critical competitive differentiator. NVIDIA’s performance-per-watt advantages and investments in cooling-optimized architectures, including direct-to-chip liquid cooling, should strengthen its position as operators seek to control operating costs and meet sustainability targets.
At the same time, rising power density creates opportunities for NVIDIA’s networking and interconnect businesses and for ecosystem partners supplying advanced electrical and thermal infrastructure. These opportunities are substantial, but they do not remove the central risk: if grid constraints become binding, the entire deployment cycle may slow.
The proper Ansatz is therefore neither alarm nor complacency. Data-center expansion remains a powerful structural demand driver, but its continuation depends on the coupled behavior of generation, transmission, cooling, water supply, regulation, and community acceptance. NVIDIA stands to benefit from the need to compute more efficiently; it remains exposed to the harder question of whether the surrounding system can deliver the required power at all.