Data Centers on Track to Use 4x More Electricity by 2035
New projections show data centers could consume four times their current electricity load by 2035, with facilities built through 2033 alone potentially drawing as much power as all of India does today. The surge is driven almost entirely by the explosive growth of AI training and inference workloads.
Original sourceA new report projects that global data center electricity consumption will quadruple by 2035, a trajectory tied directly to the accelerating buildout of AI infrastructure. The facilities scheduled for construction and commissioning through 2033 could collectively consume as much electricity as India — a country of 1.4 billion people — uses across its entire economy today. That comparison is not a rhetorical flourish; it is a load-planning number that grid operators and energy regulators are now being forced to absorb.
The growth is not evenly distributed. Hyperscaler campuses in Virginia, Texas, Iowa, and the American Southwest are already straining regional grids. Utilities that signed 20-year demand forecasts a decade ago are now tearing up those models. Several regional transmission organizations have disclosed multi-year queues for new grid interconnection requests, meaning new data center capacity is being approved faster than the transmission infrastructure to serve it can be permitted and built.
The energy mix question is equally unresolved. A meaningful fraction of announced data center capacity comes with corporate commitments to procure 100% renewable energy, but procurement commitments and actual grid physics are different things. When a data center draws a gigawatt around the clock, the marginal electron powering it at 2 a.m. in winter is rarely solar. The gap between stated sustainability goals and real-time grid impact is a policy and accounting problem the industry has not solved.
The downstream consequences are starting to materialize in electricity rates, land use policy, and water rights — data centers are also large consumers of cooling water — across the jurisdictions competing hardest for this investment. What was framed as an economic development story in most state capitals is increasingly being re-examined as an infrastructure and resource allocation story, with residential and industrial ratepayers asking pointed questions about who bears the cost of grid upgrades required to serve facilities that pay commercial rates.
Panel Takes
The Futurist
Big Picture
“The thesis embedded in this projection is falsifiable and specific: AI inference workloads will not plateau before they trigger a structural energy transition, meaning the electricity grid becomes a binding constraint on AI scaling before the model architectures do. What has to go right for this to resolve without crisis is a simultaneous acceleration in transmission permitting reform, small modular reactor deployment timelines, and demand-side efficiency gains from next-generation chips — three dependencies that have never moved in coordination before. The second-order effect nobody is pricing correctly is that energy cost becomes the primary competitive moat in AI, not model quality: whoever controls cheap, reliable, carbon-accountable electrons at gigawatt scale controls the next decade of inference pricing.”
The Skeptic
Reality Check
“The 4x figure deserves scrutiny on methodology before it becomes a planning assumption — projections of this kind have historically overstated data center growth because they undercount efficiency gains from hardware generations, and they were also wildly wrong in the opposite direction when they failed to anticipate GPU cluster density. What I'd want to see is whether this model accounts for the possibility that inference efficiency improves faster than deployment scales, which is exactly what happened with cloud storage cost curves. The prediction that kills this projection: Nvidia's next two chip generations deliver the efficiency gains they've been promising, hyperscalers consolidate workloads onto denser hardware, and the 2035 number lands at 2x, not 4x — still consequential, but a very different grid planning problem.”
The Founder
Business & Market
“The real business story here is not the data center operators — it is every company in the energy infrastructure stack that now has a decade-long demand signal with a named buyer and a non-discretionary use case, which is the rarest thing in infrastructure investing. The constraint is not capital or even technology; it is permitting velocity and transmission interconnection queues, which means the companies with existing grid relationships and pre-permitted sites are sitting on a genuinely defensible moat that has nothing to do with software. Any AI company that did not lock in long-term power purchase agreements in 2023 and 2024 is now negotiating from a position of structural disadvantage that compounds every quarter.”
The PM
Product Strategy
“The job-to-be-done that this story surfaces for the AI infrastructure layer is not 'build more compute' — it is 'make every watt of compute do more work,' and the products that solve that problem have a clearer buyer and a more defensible position than any additional GPU cluster. The teams building inference optimization tooling, workload scheduling that reduces idle GPU time, and energy-aware routing for distributed inference are now selling directly into a constraint that is visibly binding on every hyperscaler's roadmap. The gap between what's shipped today in energy-aware ML infrastructure tooling and what the 2035 demand curve requires is large enough to build a real business in.”