Don't Just Be the Landlord: The Data Centres Are Here — What Do We Keep?

Drive north out of Johor Bahru, turn off around Senai and Kulai, and grey blocks start appearing along the road. No windows. High fences. A sign at the gate with a company name you've almost certainly never heard of. You'd assume warehousing. In fact you're looking at the kind of building that has pulled in more than a hundred billion ringgit of approved investment in the past few years — data centres.

The first time I drove past, the question that came to mind was a plain one. All that land — how many people work in there?

I ended the last column by saying the world is racing to put up dormitories for a new workforce, and that some of the busiest building sites are in Johor. In early August, Singapore's Energy Market Authority gave two companies conditional approval to import a combined 900 megawatts of electricity from Peninsular Malaysia.

Our land, our power, somebody else's tenants. What I thought of first was palm oil. We've run this script before — tin, rubber, crude palm oil, crude oil. For a century the thing Malaysia has been best at is selling the top of the chain, then watching the profitable end of it happen in someone else's country.

So are data centres a blessing or a curse? My answer is that they're neither. They're a question. And it happens to be a question we once got right, on palm oil of all things.

What drew them here wasn't our brilliance. It was our endowment. There's nothing shameful in that. A large endowment is an advantage, and the neighbours would dearly like one.

This business needs four things. Land, power, water, and a fat, reliable undersea cable. Singapore is tight on the first three, having frozen approvals for new data centres in 2019 and only reopened the door, conditionally, in 2022. On the fourth it's spoilt for choice. So the demand that couldn't get in crossed the strait and settled in Johor, where land is cheap, power is spare, water is adequate, and the network is quick enough that you can't feel the difference.

How big has it got? The Ministry of Investment, Trade and Industry told Parliament last year that 143 data centre projects have been approved since 2021, twenty-five of them with tax incentives attached. We've been dealt a good hand. Playing a good hand is a separate skill.

Electricity is the worst thing in the world to export

A shipload of palm oil can sail to Rotterdam. A container of chips can fly to Mexico. Electricity can only travel down a cable, shedding a little of itself the whole way, and the thickness of that cable decides how much you can sell in a day.

It's also a crowded trade. Singapore intends to import roughly 6 gigawatts of low-carbon electricity by 2035, six or seven times the batch just approved, and the queue of suppliers includes Indonesia, Vietnam, Cambodia and Australia alongside us. That's a textbook commodity contest, settled on price and reliability. The 900 megawatts still has to clear power purchase agreements and financing, and nothing actually flows until around 2029.

Set against that, the electricity we've already approved for data centres at home is more than double it. The balance, in other words, has quietly tipped the other way. Rather than send the power abroad, we're converting it into compute here.

That instinct is the right one. Exported electricity is as raw as a commodity gets. Turned into compute, it has at least been through one stage of processing. It's the equivalent of refusing to remain a seller of crude palm oil and insisting on building a refinery.

Except that once the refinery is up, you still have to ask what comes out of it, who buys it, and at what price.

Compute may be a raw commodity too

Renting racks and selling compute is a landlord's business, and what a landlord can charge depends on how scarce the goods are. Compute is becoming unscarce at a startling rate.

Stanford's AI Index put a figure on it. When ChatGPT arrived at the end of 2022 and astonished everybody, getting an answer of that quality cost about $20 per million tokens, somewhere around RM80. Two years later the same quality cost seven cents, under thirty sen — a 280-fold drop.

That's not a sale. It's chips getting faster, models getting leaner and software getting sharper. Whatever compute impresses you today will most likely go the same way within two years, which is why an advantage built on electricity being a few sen cheaper doesn't hold for long.

Look further upstream and it gets plainer still. The most valuable thing in the building is the silicon, and the silicon isn't ours. A good deal of the rest of the kit is imported as well. Last year Malaysia also began requiring a Strategic Trade Permit for American high-performance AI chips leaving the country or passing through it. Even the rules for handling the things are written elsewhere.

So the ledger reads like this. The land is ours, the water is ours, the power is ours. The chips are bought in, the customers are abroad, and the money settles abroad.

Start with the question from the roadside. How many people work in there? The same parliamentary answer has it. Those twenty-five projects, the ones with incentives, account for RM144.4 billion of investment and are expected to create 1,429 jobs. Nobody's at fault for that. A machine hall is built for machines.

Then the power. Behind the figures I quoted last time — 7% of the peninsula's electricity today, 31% by 2035 — sits a blunter way of saying the same thing. A rough sum on the minister's own numbers suggests that between 2026 and 2035, something like seven of every ten units of new electricity demand on the peninsula gets swallowed by data centres.

And then the number nobody quotes. These data centres have reserved close to twice the electricity they actually draw. The government would like to see 85% of the declared maximum actually being used. As of April this year the real figure was 54%.

Some of that is projects still under construction, ramping up as they go. But the Energy Commission itself has said that some of those applications may simply be holding a place in the queue. The awkward part is that the grid and the water mains get built to the size of the application rather than the size of the usage, and that spending ends up spread across everybody's bill. If the demand never turns up, we'll have built half an empty dormitory on somebody else's expectations.

None of which makes data centres a bad thing. It makes them a business you pay for in instalments, rather than a cheque you applaud on arrival.

Where's the downstream?

Which brings us back to the thing we got right. In the 1970s and 80s Malaysia refused to stay a seller of crude palm oil and used policy to keep the processing onshore. That's where today's refining and oleochemicals industry came from. Compute needs precisely the same move. Somebody here has to be placing the orders.

So who places them?

The customers who keep a data centre alive aren't people asking ChatGPT the occasional question. They're institutions that keep the machines busy all year, where a single model run burns a few hundred thousand ringgit. Banks pricing risk. Plantations forecasting yield. Hospitals reading scans. Government services of every kind. All of it eats compute. The trouble is that in Malaysia the institutions willing to spend seriously on any of it are still thin on the ground.

The most direct way to grow that kind of buyer is research. Malaysian R&D spending has hovered around 1% of GDP for years, well short of the government's own 2.5% target — Korea is close to 5%, Singapore around 2%. And the money that does get spent leans hard towards the practical, towards deliverables and applications and something to show within three years. Those are exactly the questions AI is now quickest to answer for you. The work that genuinely consumes compute is the open-ended, high-failure kind, and that's the kind we're least willing to fund.

The machines are next door. That isn't the same as affordable.

There's a second route, and I think it's the more underrated one. Let people here start companies.

Cheap power won't sustain a business selling compute. What it can sustain is a group of people willing to burn compute. This field is won by trying things, and a team that can afford ten attempts beats a team that can afford one.

The catch is that having the machines next door doesn't mean anyone here can afford them — a founder in Kuala Lumpur is one thing, the price list is another. A founder in Kuala Lumpur renting compute today is mostly paying the international rate, much as they would in San Francisco. But that's a design choice, not a law of nature — the palm oil downstream didn't appear by accident either. The equivalent move now is to set aside a share of that capacity, at a price people here can actually pay, for local companies, universities and laboratories.

And this is where it bites. The World Bank's Malaysia Economic Monitor in April put roughly 1.86 million Malaysians abroad, more than half of them in skilled or semi-skilled occupations, and called the outflow structural — there simply aren't enough high-value jobs at home. A machine hall doesn't produce those jobs. Companies grown on top of one do.

Treat a data centre as a rent-collecting business and it becomes a new plantation. Vast acreage, handsome export figures, and the profitable stretch happening somewhere else. Rent is a fine thing to collect — but the ceiling on rent is always set by the tenant.

Treat it as a newly commissioned power station and the story changes entirely. Getting it built is only the first step. What decides the outcome is whether the country plugs anything into it.

Whatever those windowless grey blocks eventually come to mean, it won't be decided by the buildings.

It'll be decided by whether the rest of us switch anything on.