RM2 billion for the cloud.
RM15 million for the people.

Two national AI programmes landed within days of each other. Read side by side, they are not different in degree. They are different in kind.

Two programmes, same week, different species

Korea. The Ministry of Science and ICT selected three consortia — SK Telecom, KT and Kakao — to deliver a general-purpose chatbot and a public-service AI agent to roughly 52 million residents. Free, with no consumer usage cap, guaranteed through 2028. The state is supplying 512 Nvidia B200 GPUs this year with operating support from 2027, and mandating that at least half the models be Korean-developed. Total national AI spend for 2026: about US$7.2 billion, triple the year before.

Malaysia. From 31 August, 100,000 people aged 18 to 30 who complete modules on the Rakyat Digital platform receive three months of ILMUchat Pro — a product retailing at RM50 a month. Call it about RM15 million of subscription value, in a country of thirty-four million.

We spend a similar share. That is the problem.

Here is where the easy excuse collapses.

Korea’s economy is around four times ours, and its AI spend is roughly 0.38% of GDP. Malaysia’s headline RM5.9 billion works out to about 0.30%. On intensity, we are in the same conversation.

Except that comparison flatters us, and I would rather be accurate than comfortable. Korea’s figure is AI-specific. Malaysia’s RM5.9 billion is described by Bernama itself as a cross-ministry allocation for research, development, commercialisation and innovation — AI is a portion of it, not the whole of it. Count like for like and our AI-specific share sits meaningfully below the headline.

Which carries its own small lesson. When a budget line labelled “AI” turns out to be cross-ministry R&D, the word AI is doing public-relations work. That is the same scoreboard I keep writing about — one that counts announcements rather than transactions.

So where does the money actually go?

This is the question worth asking, and the answer is not scandal. It is philosophy.

Of the RM5.9 billion, roughly RM2.9 billion is publicly itemised: RM2,000 million for the Sovereign AI Cloud via MCMC; RM650 million for internet access across hospitals, clinics and rural health facilities; RM150 million in SME digitalisation grants; RM53 million for the Malaysia Digital Acceleration Grant; RM30 million for Cyber Security Malaysia; RM18.1 million for the National AI Office; and RM7 million for MIMOS.

And the citizen-facing AI programme: about RM15 million of subscription value.

Read the shape of that, not the totals. We fund buildings and institutions. Korea funds users. Two billion for a cloud; eighteen million for the office meant to run national AI policy; fifteen million of capability actually reaching people. Korea paid for the GPUs and fifty-two million fluent citizens.

The remaining three billion or so sits across ministries for research and innovation, without an itemisation I can find published anywhere. I want to be careful, because I am not alleging anything: that is almost certainly ordinary departmental research money. But “what did the other three billion buy, and may we see the list?” is a question any citizen is entitled to ask — and it is a far better question than any accusation.

The three layers, and the only one we can win

Step back and the strategy problem gets simple.

Model training. We lose. So does Korea — nobody there is trying to beat OpenAI. That race costs tens of billions a year and belongs to two countries. Anyone selling Malaysia a frontier-model ambition is selling a monument.

Hosting. That is rent, not a position. I have written it twice: the data-centre wave came for our geology, not our policy, and hosting the hardware means administering somebody else’s rulebook. Racks cannot move, but tenancy can.

Diffusion. This is the winnable layer, and it is the cheapest one. The only durable AI advantage a mid-sized economy can own is how well its people operate the technology. It scales with population rather than with compute. It cannot be revoked by a foreign letter. It compounds locally, inside every business that uses it.

Korea’s seven billion dollars is not mostly an attempt to out-build America. The free-access programme is a purchase of national fluency — fifty-two million people who will be comfortable with these tools by 2028.

We chose the layers in the opposite order. We spent the large number on the layer we cannot own, and rationed the layer we could.

To be fair — the cloud is not stupid, and ILMU is real

Two honest defences, because the cynical version of this argument is lazy.

The Sovereign AI Cloud has a genuine rationale. Government records, health data and anything under data-residency rules legally cannot sit on a foreign hyperscaler. If the law says the data stays in the country, you need somewhere in the country to put it. That is a real requirement, and no amount of citizen fluency substitutes for it.

And ILMU is a genuine achievement. A 100% Malaysian model, tuned for how we actually speak — rojak, code-switching mid-sentence across Bahasa Melayu, English, Chinese and dialect — reportedly at parity with GPT-4o and Llama 3.1 on key benchmarks, and ahead of frontier models on Bahasa Melayu. No foreign lab will ever optimise for Manglish; it is economically irrational for them. That is a defensible moat, and the private sector built it.

So Malaysia already owns the hard half. What we rationed was the easy half: letting people use it.

Or are we buying another white elephant?

Which brings me to the question I cannot shake.

A sovereign cloud is only sovereign if it is used. Two billion ringgit of infrastructure has to be filled — by ministries, by agencies, by regulated industries holding data that cannot travel — and filled at a utilisation rate that justifies the capital. I have not seen a published target for that, or a demand forecast.

Meanwhile the economics are unforgiving, and I have already worked through the arithmetic: the silicon inside a data centre races its own depreciation, losing value on a two-to-three-year clock while the building depreciates over decades. Compute you bought and did not use is not an asset sitting idle. It is an asset actively expiring.

I hope I am wrong, and I would genuinely like to be. But we have built impressive, underused national infrastructure before, and the pattern is familiar enough to name: the ribbon gets cut, the announcement counts, the utilisation is never published, and five years later everyone has quietly moved on.

So the question is not whether we spent enough. We spent a comparable share of our economy to a country doing something far more ambitious. The question is whether we bought capability — or whether we bought another white elephant, this time with a GPU inside it.

The bottom line

For business owners: do not wait for either programme. If you are over thirty you were not invited anyway, and ILMUchat Pro costs about one lunch a month. Start on Monday.

For anyone building policy: publish the utilisation. Publish the itemisation. A national AI programme should be judged on what still exists in month four — not on how many people registered in week one.