Nepal's New Government Wants AI Data Centers. Its Hydropower Has a Six-Month Asterisk.

I lived through load-shedding at my own home in single-digit hours a day a decade ago, and I still catch myself being surprised every time I write the sentence "Nepal exports electricity." So when Nepal's newly elected government announced it wants to stop exporting electrons and start exporting AI compute instead, I wanted to know whether the hydropower math actually supports that ambition, or whether it's aspiration running ahead of infrastructure the way the FCC just approved a satellite before anyone had jurisdiction to weigh its side effects. The short answer: the surplus is real, the ambition is genuine and now has an elected mandate behind it, and the seasonal asterisk on Nepal's hydropower is the single most important fact missing from almost every headline about this story.

What actually changed: an election, then a budget line

Nepal held snap elections on March 5, 2026, following the September 2025 "Gen Z protests" that forced Prime Minister K.P. Sharma Oli's resignation. The Rastriya Swatantra Party (RSP) — three years old at the time — won a landslide 182 of 275 seats, the second-best result in Nepal's electoral history. Balendra Shah, the 35-year-old former Kathmandu mayor better known by his stage name Balen, was sworn in as prime minister on March 27, becoming both the country's youngest PM in decades and its first from the Madhesi community.

RSP's economic platform puts digital infrastructure at the center: the party has stated an ambition to turn Nepal into a data-center and AI-compute hub within five years, explicitly reframing the country's hydropower story from "exporting electrons" to "exporting compute." The first concrete line item showed up in the FY2026/27 budget, presented in late May 2026: a "Sovereign AI Compute Center" at Syuchatar in Kathmandu, intended to house thousands of AI processing units offered to domestic startups at subsidized rates. As of that budget, the commitment exists in two clauses — with, per reporting on the budget, no specified allocation, no timeline, and no named operator. That's a real signal of intent from an elected government, and it is not yet a funded, scheduled project. Both things are true at once.

The hydropower Nepal actually has

The scale-up is genuinely fast. Nepal's installed hydropower capacity grew from around 2,800 MW to a reported 3,983 MW by December 2025 (Poush 2082 in the Nepali calendar) — roughly 39% growth in under two years — with the government targeting 4,800 MW by the end of FY2025/26. I've written before about how Nepal got here: Kulman Ghising's NEA turnaround ended load-shedding by 2018, Upper Tamakoshi (456 MW) came online in 2021, and by FY2024/25 Nepal was exporting roughly 2.35 billion units of electricity to India worth about NPR 17.46 billion. That part of the story is well established and I'm not re-deriving it here.

What's less discussed is how much further there is to go, and how uncertain the ceiling actually is. Two credible estimates of Nepal's total hydropower potential disagree by a meaningful margin: the long-cited figures — 83,000 MW theoretical, roughly 43,000 MW economically feasible — date back to older assessments and still circulate widely. Nepal's own Water and Energy Commission Secretariat revised that down in 2019, to roughly 72,500 MW gross theoretical and 32,700 MW techno-economic potential. Both numbers describe the same country; they just don't agree, and I'm showing both rather than picking one.

Either way, Nepal is using something between 5% and 12% of its own estimated economic potential today. There's room to grow. The question is what kind of power that growth actually produces, hour by hour.

The asterisk: six months of surplus, six months of imports

Here's the fact that gets lost in "Nepal is a net electricity exporter" headlines, including some of my own past coverage: nearly 90% of Nepal's hydropower is run-of-river, meaning output tracks river flow directly rather than being stored and dispatched on demand. River discharge runs roughly 84% wet-season and 16% dry-season. The practical result, reported by Nepali Times and consistent with NEA's own public statements, is a genuine monsoon surplus — around 1,200 MW currently — and a genuine dry-season deficit that Nepal still covers with imports from India. Nepal is a net annual exporter. It is not, today, a continuous exporter, and the difference matters enormously for anything that needs power 24 hours a day, 365 days a year — which is exactly what a data center needs.

No source I found puts a single, confirmed megawatt figure on the current dry-season shortfall — it's consistently described qualitatively as a deficit covered by imports, not quantified the way the monsoon surplus is. I'm not going to manufacture a precision the sourcing doesn't support. What is clear: any data-center strategy built on Nepal's hydropower has to either build real storage capacity, sign hybrid contracts that accept seasonal curtailment, or plan around dry-season imports — none of which are impossible, but none of which are what "clean hydropower to power AI" pitches usually mention.

What an AI data center actually needs

For scale, a typical hyperscale data center draws around 100 MW, per the IEA. Large AI-specific campuses now under construction run 500 to 750 MW, and the newest "AI gigafactory" tier of facility is designed around roughly 1 GW — comparable to a nuclear reactor's output. Regular readers of this site will recognize the shape of that curve from xAI's Colossus site, which I covered nearing 2 GW, and from the 120-140kW a single modern GPU rack draws — about 13-15 times an industry-average server rack. These are not small numbers, and they're not going down.

Put Nepal's reported 1,200 MW monsoon surplus against that scale and the arithmetic is genuinely interesting — not because it's huge by global standards, but because it's a real, specific, comparable number for once. It's more than ten typical hyperscale facilities' worth of power, or a bit over two large AI campuses, for roughly eight months of the year. It is a rounding error against what a single frontier-scale "AI gigafactory" campus would need, and it evaporates in the dry season without storage or backup.

That last card is the honest ceiling on this story right now: zero data centers Nepal's domestic hydropower alone can guarantee power to, twelve months a year, without either new storage infrastructure or an accepted dry-season import/curtailment arrangement. None of that makes the ambition foolish — a country that went from 18-hour blackouts to a real, if seasonal, power surplus in under a decade has earned some benefit of the doubt on execution. But "clean hydropower" and "reliable, contractable baseload for a hyperscale customer" are not automatically the same claim, and any serious data-center pitch to a foreign operator will get asked about the six-month gap before it gets asked about the electricity price.

The regulatory gap is real, and it predates the ambition

The Kathmandu Post's reporting on this surfaces a second, separate problem: Nepal's environmental-assessment framework doesn't currently list data centers under its Schedule 3 requirements, meaning a full Environmental Impact Assessment isn't automatically mandatory for one. A ministry official quoted in that reporting put it plainly: "these specific IT infrastructures haven't been envisioned in the current legal framework," which as written lacks explicit provisions for a data center's electronic waste, water consumption, or power-consumption disclosure. The same reporting cites a real, already-operating case — an Ncell data center — where nearby residents reported vibration-causing noise and visible smoke emissions, and had no clear legal standing to object during the approval process because they weren't informed the facility was a data center during initial consultations. Nepal is not alone in regulating a new infrastructure category after the first facility gets built rather than before, but it's worth stating plainly rather than assuming a "sovereign AI compute" pitch comes with a mature permitting regime already attached.

The honest read

Nepal's hydropower story over the last decade is genuinely one of the better infrastructure turnaround stories anywhere, and the RSP government inherited real, growing, underused generating capacity, not an empty promise. The specific pivot toward selling compute instead of electrons is a sound instinct — exported electricity earns a commodity price; exported AI compute, if the demand and the reliability are both real, earns more. But the two hardest parts of that pivot are exactly the parts current coverage undersells: a seasonal generation profile that makes "24/7 clean baseload" a claim Nepal can't fully back yet without new investment, and a regulatory framework that, by the government's own admission, hasn't caught up to the infrastructure category it's now trying to attract. The Sovereign AI Compute Center at Syuchatar is a real, named, government-committed first step. It is also, as of this writing, two budget clauses with no money attached yet. Both of those sentences are true, and a Kathmandu-based writer covering this story owes readers both of them, not just the more exciting one.

Sources

• NBC News — Balendra Shah sworn in as Nepal prime minister
• Al Jazeera — Nepal's youngest premier sworn in
• Wikipedia, "2026 Nepalese general election" (cross-checked against news wire coverage, not treated as a primary source alone)
• Kathmandu Post — "Nepal wants to become a data centre hub. It has no rules for how" (April 27, 2026)
• English Nepalnews — "Nepal's AI Compute Center plan draws scrutiny over feasibility and risks"; "Everything You Need to Know About Nepal's Budget for FY 2026/27"
• IPPAN, Energy Statistics of Nepal 2026
• Nepal Economic Forum, "Unlocking the Geopolitics of Hydropower in Nepal"
• Water and Energy Commission Secretariat (WECS), 2019 hydropower potential assessment (via Nepal Economic Forum, Spotlight Nepal reporting)
• Nepali Times, "Nepal has to plan beyond megawatts"
• IEA — hyperscale data center power draw benchmarks
• This site's own prior reporting: global-electricity-electronic-era-2026 (Nepal's load-shedding-to-export turnaround, NEA figures) and global-tech-landscape-2026 (xAI Colossus capacity)

Written by Abhishek Kushwaha, founder and writer at Global Tech Search, based in Kathmandu, Nepal.