On October 29, 2025, Nvidia became the first company to reach a $5 trillion market capitalization. Its stock closed more than 3% higher, just three months after the company reached $4 trillion. CEO Jensen Huang disclosed more than $500 billion in AI chip orders through the end of 2026, with plans to ship 14 million Blackwell units over the next five quarters.

That demand creates a practical problem for the companies buying the hardware: securing enough electricity to run it. Chip deliveries can be scheduled well before a utility can provide a new grid connection. A deployment with funding, servers, and an engineering team can still wait years for power.

Blackwell shows how quickly the hardware is advancing. The architecture has 208 billion transistors, 2.5 times the previous generation, and delivers up to 20 petaFLOPS of FP4 AI compute. That performance figure describes calculations using four-bit floating-point numbers. It doesn't mean every workload will achieve that speed, and buying the chips doesn't establish that a site can power them.

The scale of the electricity requirement

A 20-megawatt deployment needs roughly enough electricity to supply 15,000 homes continuously, depending on household consumption. For an AI facility, that is a capacity requirement the operator must secure before the deployment can run at its intended scale.

The demand forecasts are much larger. Estimates for global AI data centers put additional power capacity needed in 2025 at 10 gigawatts, described as more than Utah's entire power capacity. The cited projections rise to 68 GW by 2027, approaching California's total power capacity, and 327 GW of AI data center demand by 2030.

For comparison, total global data center capacity in 2022 was estimated at 88 GW. On those figures, projected AI demand in 2030 would be nearly four times the capacity of all data centers in 2022. These are forecasts, and the comparison depends on how capacity and demand are defined. Even so, they point to a substantial expansion requirement within a short period.

Individual facilities are growing too. A 30 MW data center was considered large ten years ago. Current large-scale developments can reach 200 MW, with some facilities approaching 1 GW. Training a large language model such as GPT-4 has been estimated to require approximately 30 MW of continuous power.

At the equipment level, an Nvidia DGX H100 server consumes about 10,200 watts. GB200-based systems may require rack power densities of up to 120 kilowatts per rack. A site's total electrical capacity therefore isn't the only consideration. Its electrical distribution and cooling systems also need to support that much equipment in a concentrated space.

Grid construction takes longer than deployment planning

Nvidia and its manufacturing partners can expand chip supply, although new fabrication plants also take years to build. Electricity supply has a separate set of constraints. Generating power, moving it across the grid, and delivering it to a data center require infrastructure that often takes much longer than an AI deployment plan allows.

The main barriers are:

  • Grid connections. Wait times in key regions such as Virginia can run four to seven years. Some projects face connection dates in 2030 or later.
  • Transmission lines. New lines take four to eight years to build in advanced economies. Wait times for critical components such as transformers have doubled over the past three years.
  • Generation capacity. A connection alone doesn't resolve a shortage of electricity supply. Some utilities warn that their electrical infrastructure cannot meet the aggressive targets AI companies are announcing.
  • Permitting and local approval. Environmental reviews, local opposition, and permitting processes can add years to power projects.

Analysts estimate that around 20% of planned data center projects could face significant delays unless these risks are addressed. That estimate concerns planned projects, rather than every expansion a company might want to pursue.

These constraints help explain why power can become the limiting factor even when chips are available. A hardware delivery date is only one part of the schedule. The grid connection, supporting transmission work, and generation supply must be ready as well.

How power changes site selection and costs

Available capacity can rule out an otherwise suitable site

Proximity to customers, existing infrastructure, and tax incentives still matter. But none can compensate for a grid that cannot support the proposed load on the required date.

Power availability is becoming a primary filter for site selection. Regions with strong grids and utility policies that support new connections may attract projects that would previously have gone to established technology centers. A historically attractive location can lose investment if its next usable power allocation is several years away.

This also changes what counts as a credible deployment plan. A site with enough land and a favorable tax agreement is not ready simply because servers can be delivered there. The capacity commitment and connection schedule need to be part of the decision.

Cloud providers face the same physical limits

Moving workloads to AWS, Google Cloud, or Microsoft Azure shifts responsibility for operating the facility, but it doesn't remove the electricity requirement. Hyperscalers need power allocations for new regions and expansions just as other operators do.

They have advantages, including scale, established utility relationships, and long planning horizons. Even so, unavailable power can stall their growth. Some providers are reportedly paying premiums for capacity or prioritizing regions largely on the basis of power availability.

For customers, the implication is that a preferred cloud region cannot always be treated as an unlimited source of future capacity. Regional options belong in infrastructure planning, particularly for large AI workloads.

Electricity changes the hardware calculation

Power pricing is becoming an important part of AI workload economics. In some markets, electricity costs are reported to exceed even the amortized cost of expensive GPU hardware. That comparison will depend heavily on utilization, electricity rates, hardware prices, and the period over which costs are measured.

Raw compute performance and purchase price are therefore incomplete measures of value. Evaluation also needs to consider how much useful work a system completes for the electricity it consumes. The fastest chip may not be the best choice if its power requirement limits the number of systems a facility can operate.

Power efficiency can offer a lasting advantage where electrical capacity is scarce. Reducing the energy required for a workload can lower operating costs and make better use of an existing site's allocation.

Practical steps for AI infrastructure planning

Improve the equipment and workloads already in service

  • Measure power at the rack level. Facility totals alone provide limited insight into which equipment and workloads are driving consumption. Rack-level measurements give optimization work a more useful starting point.
  • Schedule flexible workloads. Some training jobs can run during off-peak hours or move to locations with better power availability. Workloads that need continuous service offer less flexibility.
  • Evaluate smaller or more efficient models. Compression, quantization, and distillation can reduce resource requirements. Quantization uses lower-precision numbers, while distillation trains a smaller model to reproduce capabilities of a larger one. Their value depends on whether the resulting model still meets the workload's requirements.
  • Include power efficiency in hardware selection. Compare performance with power consumption, rather than ranking GPUs only by peak speed. The relevant result is useful workload performance within the available power budget.

Plan sites and utility relationships together

Organizations planning new capacity should develop options in multiple regions rather than assume the preferred location will have electricity available when needed. That requires starting utility discussions before the deployment is close to launch.

Some organizations are co-developing generation capacity, investing in local grid improvements, or partnering with existing power-intensive industries. These arrangements take time and capital, but they can become part of a capacity plan rather than a response to a missed launch date.

System architecture also affects the available options. Workloads that can be distributed across multiple sites may have more room to adapt to regional power constraints. Power availability then becomes a placement consideration alongside performance requirements.

Assess longer-term supply and efficiency options

Some operators are exploring on-site generation as a hedge against grid uncertainty. Options under discussion include natural gas, solar paired with battery storage, and small modular nuclear reactors. Exploring an option doesn't mean it will be available on the deployment's schedule, and on-site supply should not be assumed to replace a workable grid plan.

Efficiency investment extends beyond the GPU. Algorithms, model architectures, and system design all affect the electricity required to deliver a capability. A system that could deliver the same AI capability with 50% of the power would have a substantial advantage at a power-constrained site. That is an efficiency target, not an established result across AI workloads.

Deployment commitments need a power schedule

Nvidia's valuation and reported order book reflect strong expectations for AI demand. Turning those orders into operating infrastructure depends on construction and utility work that chip vendors cannot complete on their own.

As of November 2025, the central planning risk is a mismatch between AI deployment schedules and the time required to expand electricity supply. Some projects may need to be delayed, relocated, or redesigned. Funding and engineering capability help, but they cannot make an unavailable grid connection ready immediately.

Early utility agreements and regional alternatives could give organizations years of lead time. Before a large deployment date is treated as firm, the plan needs a credible answer for how much power will be available, where it will come from, what it will cost, and when it can be delivered.

References

  1. Nvidia Corporation. NVIDIA Newsroom: financial results and announcements.
  2. McKinsey & Company. (2025). AI power: Expanding data center capacity to meet growing demand.
  3. International Energy Agency. (2025). Energy and AI.
  4. Pilz, K. F., Mahmood, Y., & Heim, L. RAND Corporation. (2025). AI's Power Requirements Under Exponential Growth.
  5. U.S. Congressional Research Service. (2025). Data Centers and Their Energy Consumption.