AI’s Next Bottleneck Is Power, Not Just GPUs

FutureIsNow Editorial
1 Min Read
Futuristic scene with scientists analyzing data on transparent screens near large, glowing, high-tech machines connected by cables at sunset.

The signal: AI infrastructure is no longer constrained only by access to accelerators. Power supply, grid connections, cooling, construction schedules and local permitting increasingly determine when computing capacity can become productive. For enterprise leaders, the important question is not simply how many GPUs can be purchased, but how much reliable compute can be brought online—and at what total cost.

The bottleneck is moving from chips to infrastructure

The AI build-out has created an unusually tight coupling between software demand and physical infrastructure. Training and serving models require accelerators, but those accelerators need electricity, cooling, networking, data-centre space and dependable operations. A delayed grid connection can hold up a facility even when hardware has been ordered. A cooling constraint can limit the density at which servers operate. A shortage of trained staff can slow commissioning after the building is ready.

That is why AI capacity plans increasingly look like energy and infrastructure plans as much as technology procurement. The International Energy Agency’s 2026 electricity analysis says data-centre electricity use rose sharply in 2025, with demand from AI-focused data centres growing faster still. Its outlook expects global electricity demand to continue expanding through 2030, driven by a mix of industry, electrification, cooling and data centres—not AI alone.

Sources: the IEA’s 2026 analysis of data-centre electricity use and its Electricity 2026 executive summary.

The scale of the challenge—and the uncertainty

Forecasts vary because analysts use different assumptions about construction schedules, utilisation, efficiency and which projects will actually be completed. Goldman Sachs Research estimated in May 2026 that US data-centre power demand could rise from 31 GW in 2025 to 66 GW in 2027. The same analysis warned that only around 50–60% of capacity scheduled for the following one to two years might come online on time. These are forecasts, not guaranteed outcomes.

The US Department of Energy’s 2025 update to its data-centre energy report provides a separate scenario range: data centres could account for 9.5% to 15.3% of US electricity consumption by 2030, with a central estimate of 11.8%. The range matters. It reflects genuine uncertainty around deployment, hardware efficiency and future demand, rather than a single inevitable trajectory.

Sources: Goldman Sachs Research on US data-centre power demand and the US Department of Energy’s 2025 Data Center Energy Usage Report.

Why a GPU order is not a capacity plan

Accelerators draw substantial power, and their heat must be removed reliably. As rack densities increase, operators may need different cooling designs, stronger electrical distribution, new substations or upgrades to existing facilities. These requirements are location-specific: a region with available land may still lack transmission capacity, while a site near a power source may face water, permitting, community or network constraints.

Capacity planning should therefore model the entire path from utility connection to useful model output. Leaders need to ask when power will be available, how much can be contracted, what redundancy is required, whether cooling supports the planned rack density and how quickly the facility can be commissioned. They should also assess workload efficiency: better model selection, quantisation, batching, utilisation and software optimisation can reduce the amount of compute needed for a given result.

FutureIsNow has explored this broader infrastructure problem in the grid bottleneck behind the AI compute race and in our analysis of NVIDIA’s AI infrastructure advantage beyond chips.

Renewables, storage, gas and nuclear will all play roles

It is too simplistic to conclude that intermittent renewables cannot support data centres or that nuclear power is an immediate, universal prerequisite for AI. The practical answer depends on the grid, the generation mix, storage, transmission, demand response and the reliability requirements of each facility. Renewable power can contribute materially, particularly when paired with storage, grid capacity and complementary generation. Firm low-carbon power—including nuclear where available—may also be valuable for large, continuous loads.

The IEA’s analysis of energy supply for AI projects global electricity generation serving data centres to grow from about 460 TWh in 2024 to more than 1,000 TWh in 2030 in its base case. It expects renewables to meet nearly half of additional demand over the next five years, with natural gas and coal also contributing and nuclear playing an increasing role toward the end of the decade and beyond. The mix will differ significantly by country and region.

Source: IEA: Energy supply for AI.

Small modular reactors and other advanced nuclear projects may become part of longer-term supply strategies, but licensing, financing, construction and deployment timelines mean they should not be treated as an instant fix for near-term grid constraints. Buyers need a credible energy plan for the date their capacity is required, not only a promising technology announcement.

What this means for India

India’s opportunity is to build AI infrastructure alongside its wider digital economy while avoiding the assumption that global demand automatically translates into locally available power and capacity. The country’s fast-growing electricity needs, climate conditions, water availability, transmission infrastructure and regional differences all shape where data centres can be built efficiently.

For Indian enterprises, cloud regions and data-centre operators, location decisions should consider grid reliability, the timeline for additional power, renewable procurement options, water and cooling design, connectivity, disaster resilience and applicable data requirements. For public policy, the challenge is to coordinate digital infrastructure investment with generation, transmission, distribution and local planning so that data-centre growth does not undermine affordability or reliability for other users.

For a related view of domestic infrastructure choices, read our analysis of IndiaAI compute access and sovereign infrastructure.

A practical checklist for CIOs, CFOs and infrastructure leaders

  • Secure power before capacity: validate the utility connection, firm capacity, energisation date and upgrade responsibilities.
  • Model total cost: include electricity, cooling, networking, buildings, hardware, maintenance and financing—not only accelerator prices.
  • Test the workload: measure cost per useful output, latency, utilisation and reliability on representative production tasks.
  • Design for resilience: assess backup power, redundancy, heat rejection, water exposure and recovery procedures.
  • Compare energy pathways: evaluate grid supply, renewables, storage and firm generation based on local availability and realistic timelines.
  • Stress-test schedules: build scenarios for delayed connections, equipment lead times, lower utilisation and demand that falls short of forecast.
  • Track efficiency: revisit model choice, serving architecture and hardware utilisation before assuming that every workload requires more accelerators.

India’s own infrastructure outlook makes the planning issue tangible. The Ministry of Electronics and IT told Parliament in March 2026 that Indian data-centre capacity had grown from about 375 MW in 2020 to around 1,500 MW in 2025, while electricity demand from data centres was estimated to reach 13.56 GW by 2031–32. Separately, CEEW’s 2026 assessment projects capacity of roughly 6.5 GW by 2030 and estimates annual data-centre water use at about 150 billion litres in 2025. These are different measures and forecast horizons—not directly comparable capacity and electricity figures. Together they underline why Indian site selection must assess grid access, cooling water, transmission, renewable procurement and local climate risk, not just land and GPU supply. Sources: MeitY’s March 2026 parliamentary response and CEEW’s data-centre power and water analysis.

The FINA view: compute is only useful when it can run

The AI infrastructure race is not a simple contest to accumulate GPUs or announce the largest planned data centre. The strategic advantage belongs to organisations that can turn power, hardware, software and operational capacity into reliable output at an acceptable cost.

Energy constraints do not mean AI growth must stop, nor do they make one generation technology the universal answer. They force more disciplined planning. The best infrastructure decisions will align workloads with available power, improve efficiency, diversify supply where sensible and make delivery timelines part of the investment case.

Bottom line: In the AI economy, access to compute increasingly depends on access to dependable energy and the systems that deliver it. The signal is not that silicon no longer matters; it is that silicon alone is no longer enough.

Share This Article
3 Comments

Leave a Reply

Your email address will not be published. Required fields are marked *