US data centers will burn through one-fifth of the nation's electricity by 2035 — four times what they use today — according to a fresh forecast from BloombergNEF that underscores just how much power the AI boom demands.
A surge in AI compute will push total data center capacity to nearly 200 gigawatts over the next decade, the report projects. Roughly half of that capacity will be dedicated to training and inference workloads. The vast majority will stay concentrated in the US, which BloombergNEF expects to host 64 percent of the world's AI chips by power demand as soon as 2033.
For context, 200 gigawatts is roughly equivalent to the entire installed electricity generation capacity of the United Kingdom and France combined. The sheer scale of this buildout has caught even industry insiders off guard. Just two years ago, most analysts predicted data center capacity would hit around 120 gigawatts by 2035 — the new forecast blows past that by more than 60 percent.
"These numbers might even be conservative," the analysts warn. BloombergNEF's own 2035 estimate stands 83 percent higher than what the consultancy predicted just six months ago in December 2025. The revision mirrors a broader trend: EPRI, an electrical-industry nonprofit, has more than doubled its 2024 projection. S&P Global's forecast jumped more than a third between October and April. The pattern is clear — every new forecast gets revised upward as hyperscalers break ground faster than anyone expected.
The Grid Isn't Built for This
The coming wave of data centers will land on electrical grids that are already straining under the load. BloombergNEF's regional breakdown paints a stark picture.
The PJM Interconnection — the sprawling grid that covers 13 states from Virginia to Illinois — will have to devote 34 percent of its electricity to data centers. Texas's ERCOT grid will need to set aside 22 percent. PJM already hosts more data centers than any other US grid region and has struggled to keep up with connection requests from both giant generators and giant loads. It paused applications for new power sources to connect for four years, creating a backlog that has only worsened as demand kept climbing.
Though PJM reopened its interconnection queue to new generating sources in April 2026, the situation has grown so dire that American Electric Power — one of the country's biggest utilities — has threatened to pull out of the grid entirely. The supply-demand imbalance has driven wholesale electricity prices up 76 percent over the past year.
Even with that congestion, data centers still can't get enough. They represented 38 percent of charges in PJM's most recent capacity auction, meaning the grid manager now treats server farms as its single biggest category of new demand.
What the Cloud Giants Are Doing About It
All three major hyperscalers are scrambling to secure power for their expansion plans, but each is taking a different approach.
Amazon Web Services, the largest cloud provider with a 29 percent market share, has been signing direct power-purchase agreements with wind and solar farms at a furious pace. It's also pushing into nuclear, announcing deals to back small modular reactors that could supply carbon-free baseload power to its data center campuses. AWS now operates 123 Availability Zones across 39 regions worldwide, and its capital expenditure on infrastructure has ballooned as it races to secure both GPUs and the electricity to run them.
Microsoft Azure is betting on confidential computing and efficiency. It recently turned on confidential live migration for its secure VMs, allowing enterprises to patch and reboot without the usual trade-offs. But the company has also acknowledged that its climate goals are under pressure from AI demand. Microsoft's water consumption and carbon emissions have both risen as it builds out the data centers needed to support OpenAI's models and its own Copilot products.
Google Cloud, which holds roughly 13 percent of the cloud market, is taking a hardware-first approach. Reports suggest it's designing a custom "Frozen v2" AI chip — a follow-up to its Tensor Processing Units — specifically to slash the power cost of running Gemini inference workloads. Google has also invested heavily in clean energy, including its biggest-ever clean power purchase agreement, a project located just miles from xAI's controversial gas-powered data center campus.
The Supply Chain Squeeze Reaches Substations
The problem isn't just how much power data centers need — it's how fast the grid can deliver it. Transformer lead times have stretched to two years or more in some regions. Substation equipment is backlogged. Grid interconnection studies, which used to take 12 to 18 months, now routinely stretch past three years in PJM territory.
This has created a secondary market for existing industrial sites with strong grid connections. Several hyperscalers have quietly acquired old steel mills, aluminum smelters, and paper plants specifically for their substation capacity and high-voltage transmission access. The strategy: retrofit the building shell for servers and skip the years-long wait for a new grid connection.
Edge Computing as a Safety Valve
The power crunch is also accelerating interest in edge computing. If centralized mega-campuses can't get connected fast enough, smaller facilities closer to population centers start to look more attractive. Metro data centers — located in or near major cities — can often connect to distribution-level grids that have spare capacity even when the high-voltage transmission lines are jammed.
This feeds into the broader trend of AI inference moving closer to the end user. A model doesn't always need to run in a giant Iowa server farm — it can run on a local edge node that sits inside a city's existing electrical infrastructure. That shift is already underway, with companies like Scale Computing (recently acquired by Acumera) and Cloudflare building out platforms designed for exactly this kind of distributed, latency-sensitive workload.
What Happens Next
The BloombergNEF forecast isn't an outlier. Multiple independent analyses now converge on the same conclusion: data center power demand will at minimum double, and potentially quadruple, within a decade. The question is whether the grid can keep up.
The financial stakes are enormous. Goldman Sachs estimates that cumulative investment in US data center infrastructure will exceed $1 trillion by 2032, with more than half of that going toward power-related equipment — transformers, switchgear, backup generators, and on-site solar or battery storage. The bank upgraded its outlook for electrical equipment manufacturers in June, citing the data center buildout as a structural demand driver that will last at least a decade.
"There's no precedent for this pace of demand growth from a single industry," said one grid analyst quoted in the report. "The last time we saw anything like this was post-World War II industrialization, and that took decades. Data centers are trying to do it in under ten years."
For the Cloud & Edge Computing sector, the implications cut both ways. The surge in demand means more business for infrastructure providers, cloud consultants, and edge-computing specialists. But it also creates a bottleneck that could slow down AI deployment — and raise costs for everyone who depends on cloud services.
Read the full BloombergNEF report on TechCrunch here and our prior coverage of New York's data center moratorium for context on how regulation is starting to shape this market.