# BloombergNEF just revised its US data center power forecast 83% higher in seven months > Source: > Published: 2026-07-22 03:39:44+00:00 *A July 21 BloombergNEF report projects US data centers will consume 194 gigawatts of electricity by 2035, one-fifth of the entire country's power supply, after the firm's own December estimate put the figure at roughly 106 gigawatts. AI is moving faster than the models built to track it.* Seven months. That's all it took for BloombergNEF to nearly double its own forecast. The firm's July 21 report projects US data centers will draw 194 gigawatts by 2035, up from 5.9% of national electricity today to roughly one-fifth of everything the country produces. EPRI and S&P Global have made similarly sharp upward revisions in recent months, with EPRI more than doubling its 2024 estimate and S&P's forecast rising by more than a third between October and April. When three serious forecasters all revise this aggressively in the same direction within the same window, the story isn't the number. It's the rate of surprise. Nearly half of that 194-gigawatt figure will go to AI training and inference alone, and by 2033 the US will host 64% of global AI chips by power demand. Four companies - AWS, Google, Meta, and Microsoft - already control 42% of US data center capacity, and collectively they're expected to spend around $530 billion on AI infrastructure this year. The scale of what's being built has simply outrun the modelling. For anyone with money in energy or compute-adjacent infrastructure, the 83% revision in seven months is the investment thesis in one number. It means the forecasts most portfolio decisions were based on in late 2025 were structurally too low, and the positions sized off those numbers may be too small. Data center REITs like Equinix and Digital Realty benefit directly from the build. But the more interesting trade now sits one layer upstream: power generation. Nuclear has become the answer everyone in this space is converging on. As of May 2026, 13 announced projects have committed over 9.8 gigawatts of nuclear capacity specifically to AI data center infrastructure, and every major US hyperscaler has signed at least one deal. Constellation Energy, which operates the largest nuclear fleet in the country, has locked in 20-year power purchase agreements with major tech firms. Vistra and Talen have similar arrangements. One gigawatt is roughly the output of a single traditional nuclear reactor. Next-generation hyperscale campuses are expected to consume between one and five gigawatts each - meaning a single campus could require the equivalent of an entire plant dedicated to it. The math is uncomfortable. BloombergNEF's base-case scenario still projects a 19-gigawatt supply shortfall by 2035 even after accounting for on-site gas generation additions. That gap doesn't close without faster permitting, faster grid interconnection, and the kind of capital commitment to new generation that the utility sector hasn't seen in decades. Uranium prices are reflecting some of this: they've been holding near $86 per pound, sustained by data centre demand rather than traditional utility procurement cycles. ## What founders and VCs building in compute-adjacent markets should take from this Here's the thing about an 83% forecast revision: it's not a signal that analysts got unlucky. It's a signal that the physical infrastructure supporting AI is being built faster than the frameworks used to model it can track. For founders and investors betting on compute-adjacent startups, that gap between model and reality is where opportunity lives - and where risk hides in equal measure. Power availability is quietly becoming the binding constraint for new data centre development, ahead of land, fibre, and even capital. Regions with grid capacity - whether Texas with its deregulated market or areas near nuclear plants in the Mid-Atlantic - are attracting a disproportionate share of new hyperscale announcements. That matters for where capital flows next. Startups building in cooling, grid software, demand response, or distributed generation are not in niche markets. They're in the critical path. The same goes for the software layer: energy management platforms, AI workload scheduling tools that optimise for off-peak grid pricing, and carbon accounting infrastructure for hyperscalers facing ESG scrutiny are all getting a structural tailwind that didn't exist in quite this form two years ago. What's harder to size is the second-order effect on the compute market itself. If power is the constraint, efficiency becomes the premium. Full stop. Chips and architectures that do more inference per watt get priced differently than those that don't, and the competitive dynamics between Nvidia and its rivals - AMD, custom silicon from the hyperscalers themselves - look different when the cost of running a chip includes its share of a $10 million, 256-chip rack that burns through grid capacity like a small town. TechCrunch noted on July 21 that data centres are expected to use four times more electricity by 2035, framing what BloombergNEF is projecting as a structural transformation of the US power sector, not a temporary spike. Frankly, the scariest implication of the revision isn't the number itself. It's what a second revision of similar magnitude would mean. If BloombergNEF's July forecast is as conservative as the December one turned out to be, the US will need to build energy infrastructure at a pace and scale that has no recent precedent. That's not a disaster scenario. But it is a very different world from the one most energy policy, most utility investment cycles, and most VC energy theses were designed for. **Also read:** [Claude Fable 5 helped crack the Jacobian Conjecture after 87 years of failure](https://startupfortune.com/claude-fable-5-helped-crack-the-jacobian-conjecture-after-87-years-of-failure/) • [Curative CEO canceled a $600,000 Salesforce contract after vibecoding a replacement CRM in two months](https://startupfortune.com/curative-ceo-canceled-a-600000-salesforce-contract-after-vibecoding-a-replacement-crm-in-two-months/) • [OpenAI's Codex and ChatGPT Work just crossed 10 million weekly users and the coding wars have a new shape](https://startupfortune.com/openais-codex-and-chatgpt-work-just-crossed-10-million-weekly-users-and-the-coding-wars-have-a-new-shape/)