Try filing your tax returns in the final week of Asar, and you will get the taste. You log in to the portal of the Inland Revenue Department after lunch. It hangs. You reload. It still hangs. During the few moments of agony between logging in, paying the taxes and printing the receipt, you have consumed two cups of tea, had a few small meltdowns and got the old, familiar feeling that the government is not on your side. Now, visualise several million people doing the same thing, on the same day. You now know how we sit on the digital map.

It is not because of the incompetence of our software developers; the portal hanging is not due to incompetence. It is the result of two decades of our engineers building Nepali facades over systems built elsewhere, in which we do not hold a share or ownership. The server, the database and even the payment gateway belong to someone else. When it breaks down, it gets fixed by whoever has time on that day. While this strategy may have worked in the days of websites or mobile wallet applications, it will not work for artificial intelligence. AI is not a simple software product. It looks much closer to an electricity grid, and the one who does not own the foundation must rent it forever.

The world’s data centres need a cheap and clean electricity source, and we happen to have a mountain of it. After mulling over it for a while, thinking that this idea might never see daylight in our political landscape, I heard Finance Minister Swarnim Waglé deliver pretty much the same message in his address to Parliament on May 29. It was included in the budget announcement: Nepal is going to have its first-ever Sovereign AI Compute Centre to be built at Syuchatar, Kathmandu. Funds are allocated to purchase hundreds of AI processor units to make their use cheap for local startups. It speaks of turning our hydro power to compute power, and of inviting 15 Nepali experts abroad to return as fellows. The budget also introduces a new fund, Matribhumi Kosh, which proposes to move a part of the reserves to investments in ‘strategic assets’, among which there is supposed to be an ‘AI factory’.

Despite my reservations, this makes me happy. This is the first time the government has treated raw compute as a national asset. However, there is a difference between training a machine learning model and using a ready-trained model. Training a model requires expensive warehouses filled with specialised chips, each of which is more expensive. They require a tremendous amount of electricity, and training a single AI model takes many days or weeks. But using an already trained model is easy. It can analyse your X-ray, translate your sentences and summarise documents. The inference process is very cheap, fast and low-power, and all practical AI use is performed through inference.

When applying this distinction to the budget, the thing starts to annoy me. ‘AI factory’, ‘thousands of GPUs’ is the training language. It is a vocabulary we find in countries investing tens of billions to create frontier machine learning models. If Syuchatar is going to be just another quiet and efficient inference centre, a place where our hospitals’ diagnostic systems or farmers’ inquiries about potato blights are going to be computed under the Nepali legal system and on Nepali hardware, it will probably be one of the best moves of our government in recent times. But if it turns out to be our entrance fee to the training competition, it will lead to nothing good at all. In five years, the technology used there will be obsolete. Best chips are exported under strict restrictions, while the rest are inference chips, and the budget should explicitly say so. But it does not.

Then comes the data centre’s location. AI computations occur whenever there is a cheap, abundant electricity supply—as in Iceland, Norway, Quebec and parts of Bhutan—wherever electricity is available. Our advantage is water. So why are we building our data centre within the Syuchatar area, a valley prone to earthquakes and lacking sufficient infrastructure? There is a secondary concern hidden beneath the primary one. During winter, we import electricity from India. The data centre does not turn off its power during the rainy season; it requires electricity non-stop, all year round. Being advertised as the energy centre for the AI industry without guaranteeing annual power is just telling half the story, and the buyers will figure out the other half soon enough.

And then there is the cost. Using some portion of our foreign exchange to import GPUs is far different from using it to invest in a power project. An electricity generation station will pay back for decades. A GPU becomes obsolete within three or four years, cannot be manufactured in-country and needs replacement every single time at an even higher cost that we cannot print ourselves.

None of this is to say that the budget is wrong. But it must be made to live up to its own rhetoric, because it is precisely the danger that we understand so well. Sovereignty does not lie in the physical placement of a box. Building a compute centre using foreign GPUs, running models developed elsewhere, sustained through foreign services, optimised by individuals we never trained, is simply old economics in new clothes.

Real sovereignty is owning the whole stack: Our own Nepali-language data, engineers who can fix the thing at three in the morning, models tuned for the people who speak Maithili, Tharu and Nepal Bhasa, and never show up in any dataset put together in California. Fifteen fellowships are a kind gesture in that direction. It is not a workforce. The bottleneck has always been people, not hardware, and people is the line in this budget with the smallest figure next to it.

We have been late to every technological turn of the last two hundred years. This budget is the first time I can remember us arriving early, with the right hunch and a real advantage in hand. That is rare, and it is worth protecting.

The IRD portal will keep crashing either way, Syuchatar or no Syuchatar. The real question is the next portal, the one that will diagnose, decide, translate and govern. Whether it runs on hardware we understand and control, or on one more expensive wrapper around a machine that belongs to someone else, is the choice this cabinet makes over the next two years, on purpose or by drift.