Chapter 3: The Hydrological Compute Engine

Rivers as Currency

Every major technology company in the world has the same problem, and they are running out of ways to solve it.

Training a frontier AI model (the kind that powers ChatGPT, Gemini, or Claude) requires an amount of electricity that would have seemed absurd a decade ago. A single training run for a model like GPT-4 consumed an estimated 50 gigawatt-hours of electricity, roughly equivalent to powering 4,600 American homes for an entire year. The next generation of models will consume more. The generation after that, more still.

The companies building these models are desperate for power. Microsoft has signed a deal to restart a nuclear reactor at Three Mile Island. Google is investing in small modular reactors. Amazon has purchased a data center campus adjacent to a nuclear plant in Pennsylvania. The message is clear: the AI industry has outgrown the grid. It needs dedicated power sources, preferably ones that are clean, cheap, and abundant.

Nepal has 43,000 megawatts of economically viable hydropower sitting in its rivers. It has tapped less than 3,600 MW of that potential. And unlike nuclear reactors (which take a decade to permit and build), hydropower projects in Nepal are being constructed at speed. As of early 2026, 259 projects are under construction, poised to add over 10,600 MW to the grid. The government's target is 15,000 MW of installed capacity by 2030 and 28,500 MW by 2035.

The math is straightforward. Nepal is building far more generating capacity than its domestic economy can absorb. The question is not whether there will be surplus power. It is what Nepal does with it.

The Seasonal Problem

There is a catch, and it is a significant one.

Nepal's hydropower sector relies overwhelmingly on run-of-river projects, which are facilities that generate electricity from the natural flow of rivers without large storage reservoirs. This design is cheaper to build and less environmentally disruptive than large dams, but it comes with a fundamental vulnerability: output depends on how much water is flowing.

During the monsoon months, from June through September, Nepal's rivers swell with rainfall and snowmelt. Power generation surges. The country produces more electricity than it can use or export, and surplus energy is effectively wasted as water spills over dams that are already generating at capacity.

During the dry season, from November through April, the opposite happens: river flows drop, and generation capacity can plunge by 40 to 60 percent. Consequently, Nepal, which was exporting power during the monsoon, becomes an importer, buying electricity from India at prices that strain the national budget.

Climate change adds another layer of uncertainty. A weak monsoon, an increasingly plausible scenario as weather patterns shift, can sharply reduce generation across the entire hydropower fleet, affecting domestic supply, export commitments, and any industrial operations that depend on cheap power.

Any strategy that treats Nepal's hydropower as a constant, year-round resource is ignoring basic hydrology. The strategy has to work with the seasons, not against them.

The Interruptible Load Model

This is where the nature of AI workloads becomes strategically important.

Not all computing is created equal. A bank's transaction processing system needs to run 24 hours a day, 365 days a year, with zero tolerance for downtime. A consumer web application needs consistent uptime or users leave. These are the kinds of workloads that require rock-solid, year-round power, which is exactly what Nepal's seasonal hydro grid cannot reliably provide.

AI model training is different. Training is a batch process. You feed a model enormous volumes of data, run computations across thousands of GPUs for weeks or months, and produce a trained model at the end. The process can be paused. It can also be checkpointed (saved mid-computation so that it resumes exactly where it left off) and scheduled around power availability.

This characteristic makes AI training an ideal candidate for what energy economists call an "interruptible load": a large power consumer that agrees, contractually, to reduce or cease operations when the grid is under stress, in exchange for significantly lower electricity rates during peak availability.

The model works like this: during the monsoon months, when Nepal's hydropower is generating surplus electricity, data centers run at full capacity. GPUs train models around the clock, consuming power at industrial rates. When the dry season arrives and generation drops, the data centers throttle down or shut off entirely. The hardware sits idle for three to four months. When the monsoon returns, operations resume.

For a traditional industry, shutting down for a quarter of the year would be economically ruinous. For AI training, it is not, because the value is captured in the trained model, not in continuous operations. A global AI company can schedule its largest training runs for Nepal's wet season, extract the trained model, and deploy it from data centers elsewhere for the rest of the year.

The Bhutan Precedent

This is not a theoretical model. Nepal's neighbor has already proven it works.

Bhutan, a country with an even smaller economy and a similar hydropower profile, entered into a partnership with Bitdeer Technologies, a major computing infrastructure company. The deal: Bitdeer would build and operate a 500 MW high-performance computing facility in Jigmeling, expanding from an earlier 100 MW facility in Gedu.

The Power Purchase Agreement includes explicit seasonal flexibility. The computing center scales down during winter months when domestic power is scarce and imported electricity is expensive. Despite operating for only part of the year, the facility is highly profitable. Bhutan's cheap hydropower, combined with the natural cooling provided by altitude, results in production costs so low that the margins remain substantial even after accounting for hardware depreciation and seasonal downtime.

There is an important distinction to acknowledge: Bhutan's facility is primarily focused on cryptocurrency mining, not AI training. Crypto mining is a simpler computational task, as it does not require the same interconnect bandwidth or memory capacity as training a large language model. Scaling from crypto mining to AI training infrastructure requires more sophisticated hardware, faster networking, and engineers capable of managing GPU cluster operations.

But the energy economics are identical. The fundamental insight that seasonal hydropower surplus can be monetized through compute-intensive workloads that tolerate interruption applies regardless of whether the workload is hashing Bitcoin transactions or training neural networks.

The Regulatory Unlock

For years, Nepal's energy market operated as a near-monopoly. The Nepal Electricity Authority (NEA) controlled generation, transmission, and distribution. Independent power producers existed, but they sold electricity to the NEA at fixed rates and had limited ability to negotiate directly with end consumers.

That changed in January 2026.

The Electricity Regulatory Commission issued the Open Access Directive, which fundamentally restructured the market. Under the new rules, any entity requiring 5 MW or more of power can access the national transmission and distribution networks on a non-discriminatory basis. In practice, this means a foreign technology company can now sign a direct, bilateral Power Purchase Agreement with an independent hydropower producer, paying only standard wheeling charges to the NEA for use of the transmission infrastructure.

This reform matters enormously for the data center strategy. It allows for customized pricing structures (such as time-of-day rates, seasonal indexing, and interruptible load discounts) that a monolithic state utility would never have offered. A hyperscaler negotiating a PPA with an independent producer in Nepal's western hills can structure a deal where monsoon electricity costs as little as NPR 5 per kilowatt-hour ($0.037/kWh). For comparison, industrial electricity in Singapore costs roughly $0.15/kWh. In Northern Virginia, the US data center capital, rates range from $0.06 to $0.10/kWh.

At $0.037 per kWh, Nepal is not just competitive. It is among the cheapest clean power sources available to the global AI industry.

The Case Against Exporting Raw Power

Nepal is already moving to export electricity on a significant scale. Long-term agreements are in place to export up to 10,000 MW to India, and a tripartite agreement facilitates exports to Bangladesh via the Indian grid. Major transmission corridors are under construction, including a 315-kilometer, 400 kV line funded by the Millennium Challenge Corporation.

Exporting electricity is straightforward and low-risk. It generates foreign exchange. It utilizes surplus capacity. It is the path of least resistance.

But it is also the path of least value.

When Nepal exports a megawatt-hour of electricity to India, it earns a few cents per kilowatt-hour. The value chain ends at the border substation. India uses that electricity to power its own data centers, its own AI companies, and its own IT sector, capturing the downstream economic value within its own borders.

When Nepal uses that same megawatt-hour to power a GPU cluster training AI models for global clients, the economic calculus changes entirely. The electricity becomes an input into a service, "Green GPU Compute," that can be sold at rates orders of magnitude higher than raw power. The revenue stays in Nepal. The jobs stay in Nepal. The technical expertise stays in Nepal. The tax revenue stays in Nepal.

This is the core argument of this chapter, and indeed of this entire book: Nepal's hydropower is worth far more as compute power than as exported electricity. Exporting megawatts builds transmission lines. Converting those megawatts into compute exports builds a technology economy.

What Could Go Wrong

It would be dishonest to present this strategy without acknowledging its risks.

Monsoon variability. Climate models suggest increasing unpredictability in South Asian monsoon patterns. A sequence of weak monsoon years could reduce the surplus power available for data centers, straining both the interruptible load model and Nepal's domestic energy security.

Transmission bottlenecks. Hydropower plants are often located in remote river valleys. Data centers need to be close enough to these plants to minimize transmission losses, but also accessible enough for hardware logistics, maintenance crews, and fiber-optic connectivity. Nepal's road network and transmission grid are not yet designed for this kind of distributed infrastructure.

Geopolitical risk. Nepal's energy strategy operates in the shadow of its two giant neighbors. India's influence over cross-border power trade, and China's growing involvement in Nepali hydropower development, introduce political variables that pure engineering analysis cannot resolve.

Environmental and social impact. Large-scale hydropower development has real costs, including displaced communities, altered river ecosystems, and seismic risks in a geologically active region. The narrative of "clean, green hydropower" is broadly true, but it should not be used to paper over legitimate concerns about environmental stewardship.

Execution risk. Nepal's track record on large infrastructure projects is mixed at best. Bureaucratic delays, procurement challenges, corruption, and political instability have derailed projects that looked viable on paper. The strategy outlined here requires sustained political will across multiple election cycles, which has historically been difficult to maintain.

These risks do not invalidate the strategy. They condition it. The interruptible load model, the Open Access Directive, and the favorable energy economics are real. But translating those advantages into operational data centers requires a level of institutional execution that Nepal has not yet demonstrated at scale. The following chapters address the physical and technical infrastructure that would make it possible.

Key Takeaways

Nepal has 43,000 MW of viable hydropower potential; less than 3,600 MW is tapped, with 10,600+ MW under construction.
AI training workloads are uniquely suited for an interruptible load model: they can be checkpointed, paused, and scheduled around seasonal power availability.
Bhutan has proven this model works: its 500 MW Bitdeer facility operates profitably despite seasonal shutdowns.
The January 2026 Open Access Directive allows foreign companies to sign direct PPAs with independent power producers, unlocking customized pricing.
Converting hydropower into compute exports captures orders of magnitude more economic value than exporting raw electricity.