The artificial intelligence boom is no longer primarily a semiconductor story; it's becoming an electricity and infrastructure crisis. While investors have focused on chip makers like NVIDIA, the actual bottleneck constraining AI expansion has quietly migrated downstream to power generation, transmission equipment, and cooling systems that the U.S. grid was never designed to handle.

By the end of 2025, the four largest U.S. hyperscalers,Microsoft, Alphabet, Amazon, and Meta Platforms,are expected to spend more than $700 billion on AI infrastructure, with that number climbing toward $1 trillion by 2027. But that money isn't flowing into software or even chips anymore. It's flowing into football-field-sized buildings, gigawatt-scale power systems, and transmission lines that don't exist yet.

Why Is Electricity Suddenly the Limiting Factor?

A single modern AI data center now consumes the power of a mid-sized city. Meta's Hyperion facility is currently operating at 2 gigawatts and is expected to scale to 5 gigawatts, according to recent reports. A single AI rack with NVIDIA's latest GB200 NVL72 system draws about 120 kilowatts, roughly equivalent to the continuous power consumption of 100 average U.S. homes. The next-generation Rubin platform is being designed for 250 kilowatts or more per rack.

According to Goldman Sachs analysis cited in recent market commentary, global data center power demand could rise roughly 160 percent by 2030, while U.S. electricity demand from data centers alone may nearly triple during that period. By 2030, U.S. data centers could account for 17 percent of all national electricity consumption, up from just 4 percent today. The nation's power grids were not built for this kind of demand growth.

The constraint isn't just about generating enough electricity. It's about delivering it reliably, 24 hours a day, seven days a week. Transformer lead times have stretched from 12 months to two and a half years, while heavy gas turbines are running up to seven years out. In some northern Virginia utility territories, grid hookups for projects filed after 2024 won't be available until at least 2028.

Where Is the Investment Opportunity in This Crisis?

The shift in who's buying power tells the story. A decade ago, the biggest buyers of nuclear and baseload power were utilities. Today, they're software companies. Microsoft signed a long-term commercial power agreement in September 2024 to restart one of the reactors at the former Three Mile Island nuclear facility, with Constellation Energy investing $1.6 billion to revive the plant and deliver 835 megawatts of dedicated capacity. Amazon signed a 17-year, $18 billion power purchase agreement with Talen Energy, underscoring just how aggressively Big Tech is moving to lock up baseload supply.

The infrastructure buildout unfolds in two distinct phases, each with different investment implications:

• Near-term (2025-2027): Natural gas, on-site generation, and grid infrastructure are the layers most directly benefiting from the constraint, because hyperscalers need immediate, scalable electricity for data centers that cannot wait the five to eight years a typical utility interconnection now takes.
• Medium-term (2027-2028): The binding constraint will increasingly be generation itself, which is why behind-the-meter nuclear and small modular reactor (SMR) development have become live investment categories rather than science projects.
• Long-term (2028 and beyond): The constraint bifurcates into three parallel tracks: water and nuclear fuel and electrical-steel inputs; the structural shortage of skilled trades labor; and the workload shift from training to distributed inference at the edge.

How to Position for the Power Infrastructure Supercycle

Investors tracking this shift have identified several critical chokepoints in the power delivery chain where supply constraints create investment opportunities:

• Cooling Systems: Direct-to-chip liquid cooling can carry heat 3,500 times more efficiently than air at the same flow rate. Vertiv Holdings already supplies most of the large hyperscaler build-outs with its cooling distribution units and has a market cap above $100 billion with annual revenue over $10 billion.
• Optical Interconnects: When data leaves one chip, it has to communicate with thousands of others, and copper wire cannot keep up. Silicon photonics converts electrical signals into pulses of light at speeds copper simply cannot match. Coherent Corp. is the leading supplier of the optical transceivers that make this happen, and NVIDIA took a $2 billion stake in the company earlier this year as a signal of how central it is to the next phase of AI buildout.
• Power Distribution Equipment: Transformers, switchgear, and power distribution infrastructure connect every data center to the actual grid. Eaton Corporation has been solving exactly this problem for more than 100 years, and data centers are now the company's fastest-growing end market, with multi-year backlogs and a $148 billion market cap.
• Memory Supply: High-bandwidth memory (HBM) is what allows a GPU to actually function. Micron Technology is the only American producer of HBM, and its output for 2025 was completely sold out before the year began, with capacity for 2026 already largely committed.
• Nuclear Generation: Constellation Energy is the largest nuclear operator in the U.S., with 21 reactors and a 20-year deal signed with Microsoft to bring Three Mile Island back online. The company generated about $4.2 billion in operating cash flow during 2025 while continuing to return capital to shareholders through dividends and share repurchases.

The broader thesis mirrors previous technology cycles. The internet era created massive winners in fiber optics, semiconductors, cloud computing, and wireless infrastructure. AI may now be creating a similar investment cycle centered around electricity generation, transmission infrastructure, cooling systems, battery storage, and grid modernization.

Memory spend as a share of hyperscaler budgets has shifted from 8 percent to 30 percent in just two years, according to market analysis. This reallocation reflects the reality that the marginal dollar in AI infrastructure is increasingly migrating out of accelerators and into the physical machinery that keeps those accelerators running.

The stocks in this infrastructure space may have already moved higher as investors recognized the connection between AI expansion and electricity demand, but the actual infrastructure buildout is still in its early stages. The gap between current valuations and the physical assets that still need to be constructed is exactly where the opportunity lives, according to market observers tracking the shift.