Who builds what in Nepal today

The country is small enough that the AI ecosystem fits, roughly, on a single mental map. The map is also changing fast — what is true in 2026 may not be true in 2028. This section sketches it as it stands now: not as a directory, but as a way of seeing who tends to do what, and why.

Five clusters, doing different jobs

You can roughly group the people building AI in Nepal into five clusters. They overlap, talk to each other, and sometimes share staff — but their incentives, their skills, and the problems they pick differ.

Banks and digital-payments firms. Khalti, eSewa, IME Pay, Nabil, NIC Asia, and the big commercial banks. They use AI mainly for fraud detection, transaction risk-scoring, customer-support routing, and increasingly for credit-decision support. Their advantage is the only thing that matters in finance: they have transactional data, lots of it, on lots of people. Their disadvantage is that they don’t usually publish anything — what they learn stays inside.

Startups and product companies. Two flavours here. The first is outsourcing-style firms — Fusemachines is the most famous — that train and contract out engineers to foreign clients while building selective local products. The second is product-first startups: Paaila Technology (robots and conversational AI in Nepali), Genese (cloud and AI services), Cloudfactory (data labelling at scale), Docsumo (document AI), and a long tail of smaller teams. The product companies are where most public Nepali AI work actually ships.

Universities and research groups. Pulchowk (Institute of Engineering, Tribhuvan), Kathmandu University, NAAMII (Nepal Applied Mathematics and Informatics Institute), and a handful of smaller programmes. They produce the people most other clusters hire from. The published research is uneven but improving — NepaliBERT, Nepali speech datasets, and several papers on Devanagari OCR have come from this circle.

NGOs, donors, and multilaterals. ICIMOD, UNDP, WFP, WHO, and the major donors fund AI work mostly in three domains: climate and disaster modelling, agricultural advisory, and health information systems. They bring two things to the table that the private sector usually does not: tolerance for long timelines and a willingness to fund public datasets. Some of the most useful Nepal-specific datasets in existence — glacial inventories, crop calendars, malaria-risk maps — came from this cluster.

The diaspora. Several thousand Nepali engineers work at major AI labs, big tech, and AI-heavy startups in the US, UK, Australia, and the Gulf. Many of them keep one foot in Nepal — they advise startups, contribute to open-source projects, teach summer workshops, send remittances that fund a sibling’s CS degree. The diaspora is the country’s largest unused AI asset; turning attention back home is one of the more interesting recent trends.

What each cluster is good at

A rough working map:

If you want money pulled out of high-volume noisy signals — fraud, risk, credit — go to the banks and payments firms.
If you want a shippable Nepali-language product in your hands by end of quarter, go to a product startup.
If you want a peer-reviewable claim about a specific Nepali corpus, dataset, or model, go to a university research group.
If you want a long-running public-good dataset built and maintained, work with an NGO or donor.
If you want frontier-grade engineering judgment on a specific architectural choice, reach into the diaspora.

The mistake to avoid is asking the wrong cluster for the wrong thing. NGOs are bad at shipping software products; banks are bad at publishing; universities are bad at maintaining production systems past a PhD’s graduation; startups are bad at long-horizon public-good work. The country gets best leverage when each cluster does what it is good at and the others get out of the way.

What is missing

Two structural gaps stand out. The first is a public Nepali AI dataset office — an entity whose job is to clean, schema, version, and publish Nepal-specific datasets the way Hugging Face does globally. Several proposals exist; none has launched at scale.

The second is a credible domestic compute provider. Renting from foreign clouds is fine for most projects, but it leaves Nepali sovereignty questions unanswered and ties costs to currency rules. A regional compute cooperative — even a modest one — would change which problems are economically tractable here.

These gaps are the reason for Chapter 6, where we return to what the country should build.

Check your understanding

Quick check

—

A donor agency wants a Nepal-specific dataset of monthly cropping patterns built and maintained over five years, with quarterly updates. Which cluster is best suited to lead this work?

A commercial bank.
A consumer product startup.
An NGO or multilateral, possibly in partnership with a university group.
A diaspora engineer working evenings.

What comes next

Chapter 1 has stayed at altitude — context, substrate, ecosystem. Chapter 2 drops into the first real domain: language. Nepali NLP, Devanagari OCR, and speech are where the country’s needs and the available technology meet most directly. We start with what large language models actually do, and do not, with Nepali.