Chapter 5: High-Performance Computing: From Academic Experiment to National Infrastructure

The Machines at Banepa and Kirtipur

Nepal has two supercomputers. That sentence alone would have been unthinkable a decade ago.

The first sits at Kathmandu University's IT Park in Banepa (the same facility whose troubled history was described in the previous chapter). The High-Performance Computing cluster was donated by CERN, the European Organization for Nuclear Research, and began operations in 2019. The system consists of 184 CPU servers, 16 disk servers, 12 network switches, over 2,500 processor cores, and 8 terabytes of memory.

The numbers sound impressive, and for Nepal they are. Computational tasks in scientific modeling that previously took ten days on a desktop workstation now finish in two hours. Researchers at Kathmandu University have used the system for projects in climate modeling, disaster prediction, and bioinformatics, representing work that was previously impossible without sending data abroad for processing.

The second system sits at Tribhuvan University's campus in Kirtipur, operational since January 2022. It is smaller but still significant: 500 CPU cores, dedicated GPUs, 5 terabytes of memory, and 20 terabytes of storage. The TU-HPC supports parallel processing for physics simulations, machine learning experiments, and climate modeling.

Between them, these two facilities represent something more important than their raw specifications suggest. They prove that Nepali institutions can install, operate, maintain, and productively use high-performance computing infrastructure. The administrative competence, cooling, power management, user training, and security protocols all had to be built from scratch, and they were.

What the Machines Actually Do

It is one thing to list hardware specifications. It is more useful to describe what these machines have actually produced.

The KU HPC unit, operated by the Kathmandu University Information Technology Center (KUITC), has focused on several research domains that align directly with national priorities. Climate modeling is perhaps the most consequential: Nepal is one of the most climate-vulnerable countries on Earth, with glacial lakes that threaten downstream communities, monsoon patterns that are growing less predictable, and agricultural systems that depend on rainfall timing. Running climate models locally, rather than requesting compute time from foreign institutions, means Nepali researchers can focus on the variables that matter most to Nepal.

The facility also serves as a training ground. As described in Chapter 2, the National AI Policy mandates 5,000 trained AI professionals by 2030. The HPC clusters at KU and TU are the only facilities in Nepal where students can gain hands-on experience with parallel computing, distributed systems, and the kind of workloads that professional AI development requires. A student who has managed jobs on a 2,500-core cluster has fundamentally different skills from one who has only worked on a laptop.

But the honest assessment is this: the research output, while valuable, remains modest relative to what these machines are capable of. Utilization rates for both clusters could be higher. The user base is concentrated among a relatively small group of researchers who have the training to submit jobs and interpret results. Expanding access to more universities, private sector researchers, and government agencies is a governance and education challenge as much as a technical one.

The Gap Between CPUs and GPUs

Here is where the story gets complicated.

Nepal's existing supercomputers are CPU-heavy systems. They are built around general-purpose processors, the kind of chips that Intel and AMD have manufactured for decades. CPUs are excellent for a wide range of computational tasks: scientific simulation, data analysis, numerical modeling.

The modern AI industry runs on something different: GPUs, or Graphics Processing Units, which were originally designed for rendering video game graphics and turned out to be extraordinarily well-suited for the matrix mathematics that underpin neural networks. Training a large language model requires thousands of GPUs operating in parallel, connected by high-speed networking that allows them to share computation across the cluster.

The scale of this distinction matters. The KU cluster has 2,500 CPU cores. A single Nvidia H100 GPU, the current standard for AI training, has the equivalent parallel processing capability of hundreds of CPU cores for machine learning workloads. A serious AI training cluster would require thousands of H100s (or their successors) operating together. The hardware investment for a competitive GPU cluster starts at tens of millions of dollars and scales quickly into the hundreds of millions.

Nepal does not currently have a single GPU cluster at this scale. But Nepal is not alone in this deficit. A European Parliament study revealed that the European Union captures only 7 percent of global AI investment, while the United States and China collectively capture 80 percent. If Europe, representing some of the wealthiest economies on Earth, is falling this far behind in the compute race, the urgency for a geographically advantaged country like Nepal to secure its own infrastructure is obvious. The TU-HPC has some GPU capacity, but nothing approaching what commercial AI training demands.

This gap between academic CPU-based supercomputing and commercial GPU-based AI infrastructure is the central challenge of this chapter.

The GPU Supply Chain Problem

Even if Nepal had the capital to purchase thousands of advanced GPUs tomorrow, it would face a problem that is frustrating many far wealthier nations: availability.

The global GPU market is shaped by geopolitics in ways that directly affect Nepal's options. The United States has imposed export controls on advanced AI chips (specifically targeting Nvidia's A100 and H100 GPUs and their successors) restricting their sale to China and a growing list of other countries. While Nepal is not currently subject to these restrictions, the regulatory landscape is evolving rapidly. The US CHIPS and Science Act, passed in 2022, aims to reshape global semiconductor supply chains in ways that prioritize American strategic interests.

Nvidia effectively controls the market for high-end AI training chips. Its order backlog stretches months into the future. Companies like Microsoft, Google, and Meta have placed orders worth billions of dollars. A Nepali entity entering the queue for H100s is competing for allocation against the most well-capitalized technology companies on Earth.

AMD offers alternatives. Intel is developing its own AI accelerators. Chinese companies like Huawei are building domestic alternatives. But the ecosystem (the software libraries, developer tools, and training frameworks) is overwhelmingly optimized for Nvidia hardware. Choosing a different chip vendor means accepting compatibility risks and a smaller support community.

None of this is insurmountable, but it shapes the strategy. Nepal is unlikely to acquire leading-edge GPU hardware through direct purchase at scale in the near term. The more viable path involves attracting foreign companies (such as hyperscalers, cloud providers, or specialized AI compute companies) that already have established relationships with GPU vendors and can deploy their own hardware within Nepal's green data centers.

Inference Over Training: The Open Source AI Strategy

Because of these hardware constraints, Nepal must be ruthless about where it competes in the AI ecosystem.

A common fallacy in national AI strategies is the belief that every country must train its own "frontier model" from scratch. Training a model on the scale of OpenAI's GPT-4 requires tens of thousands of H100 GPUs running continuously for months, consuming megawatts of power and hundreds of millions of dollars in capital. This is a billionaire's game. Developing nations that attempt to train base models from scratch will burn their limited capital playing catch-up.

Instead, Nepal's strategy must be strictly focused on Inference and Fine-Tuning of open-source models.

We are currently living through a golden age of open-weight AI. Global innovators have realized that open-sourcing model weights commoditizes the base layer of AI, allowing developers everywhere to build upon their work.

Mistral AI in France has become the European champion of this approach, proving that highly efficient, smaller open-weight models (like Mixtral) can rival the performance of massive proprietary models.
DeepSeek, a Chinese AI initiative, sent shockwaves through the industry by releasing open-weight models with highly efficient architectures. DeepSeek proved that frontier-level intelligence could be delivered at a fraction of the traditional cost, triggering a massive drop in global API pricing and proving that efficiency can beat brute-force spending.
Moonshot AI (Kimi) demonstrated how pushing the boundaries of specific features, such as massive multi-million-token context windows, can revolutionize how we process massive document repositories, which is a capability crucial for digitizing legal and bureaucratic systems.

For Nepal, the playbook is clear. You do not need 10,000 GPUs to train a base model. You only need a fraction of that compute to host and run inference on DeepSeek or Mistral models. A Nepali tech startup can download these world-class open weights, deploy them in Nepal's green data centers, and fine-tune them using localized datasets to understand the Nepali language, legal codes, and cultural nuance perfectly.

By doing so, Nepal can offer a sovereign AI API to its government ministries and local startups for pennies on the dollar, while entirely bypassing foreign dependency on OpenAI or Anthropic. Let the hyperscalers spend billions training the base models; Nepal should focus on hosting them efficiently using cheap hydropower.

The Economics of Compute-as-a-Service

The business model that makes this work is not one where Nepal buys GPUs and resells compute time. That approach requires capital that Nepal does not have and vendor relationships that take years to build.

The model that works is Compute-as-a-Service: Nepal provides the real estate, the power, the cooling, and the connectivity. A foreign partner provides the hardware, the software stack, and the customer relationships. Revenue is shared. Jobs are created locally for facility operations, hardware maintenance, and system administration. Over time, Nepali engineers gain experience managing GPU clusters, and domestic capacity grows.

This is how most emerging compute markets develop. It is how several countries in the Middle East and Southeast Asia are building their AI infrastructure: partnering with established cloud providers who deploy hardware in local data centers, rather than trying to build the entire stack domestically from day one.

The advantage Nepal brings to these partnerships is described in Chapters 3 and 4: electricity at $0.037/kWh and natural cooling that eliminates 35 to 45 percent of operating costs. For a hyperscaler evaluating where to deploy its next GPU cluster, those economics are compelling, provided the supporting infrastructure (power reliability, connectivity, physical security, legal frameworks) is credible.

The 100 percent FDI allowance under the 2025 IT Ordinance removes one of the traditional barriers to these partnerships. Foreign companies can now fully own and operate computing facilities in Nepal without requiring a local majority partner.

The Bridge from Academic to Commercial

The transition from two university supercomputers to a national compute infrastructure is not a single leap. It is a series of steps, each building on the last.

Step 1: Maximize existing capacity. Before building anything new, the KU and TU HPC clusters should be operating at full utilization. This means expanding access to researchers across all universities, not just KU and TU. It means creating shared compute programs where government agencies can submit modeling jobs like weather prediction, infrastructure planning, and public health analysis, rather than contracting these to foreign providers.

Step 2: Establish a national HPC governance body. Both clusters currently operate under their respective university administrations. A national coordinating body, possibly housed within the proposed National AI Centre, could manage shared access, set priorities, and plan capacity expansion.

Step 3: Pilot a commercial GPU facility. A small-scale facility (500 to 1,000 GPUs) co-located with a hydropower substation and designed for the interruptible load model described in Chapter 3. This facility would serve as a proof of concept for foreign partners and a training ground for the Nepali engineers who would eventually manage larger deployments.

Step 4: Scale with foreign partnerships. Once the pilot demonstrates reliable operations, the economics of Nepal's power and cooling advantages become a concrete sales pitch rather than a theoretical one. The data (actual PUE metrics, actual uptime statistics, and actual cost per GPU-hour) becomes the basis for attracting hyperscale investment.

The Cost Question

Transparency about costs is essential if this strategy is to be credible.

Operating a supercomputer is not cheap, even when the electricity is inexpensive. Hardware depreciates. Components fail. Cooling systems require maintenance. System administrators need salaries. Software licenses, even for open-source platforms, require support contracts for mission-critical workloads.

The KU HPC cluster, despite being donated hardware, has ongoing operating costs that the university absorbs. The TU system faces similar constraints. Neither facility has published detailed cost-per-compute-hour metrics, which makes it difficult for potential users (whether academic or commercial) to evaluate whether running workloads on these systems is more cost-effective than using foreign cloud providers.

This lack of cost transparency is a problem that needs to be solved. If Nepal wants to position itself as a destination for commercial compute, it needs to publish real numbers: cost per GPU-hour, uptime percentages, average job queue times, maintenance schedules. The competitive advantage only matters if it can be demonstrated with data, not just asserted with optimism.

The foundation exists. Nepal has proven it can operate high-performance computing systems. The next step is proving it can do so at a scale and reliability level that the global AI industry takes seriously. The hardware chapter of Nepal's story is still being written, but at least the first pages are real.

Key Takeaways

Nepal's two supercomputers (KU Banepa: 2,500 cores; TU Kirtipur: 500 cores + GPUs) prove domestic competence in HPC operations.
The critical gap is not CPU-based scientific computing: it is GPU-based AI training infrastructure, which requires a fundamentally different hardware profile.
GPU supply constraints and US export controls mean Nepal should attract foreign partners with established supply chains rather than attempting direct procurement.
The Compute-as-a-Service model (Nepal provides power, cooling, connectivity; foreign partner provides hardware) is the realistic path for scaling.
The transition from academic to commercial HPC must follow four steps: maximize existing capacity, establish national governance, pilot a commercial GPU facility, then scale with foreign partnerships.