AI Is Taking Your Memory — and Rewriting How We Work

A friend went laptop shopping the other week at Low Yat Plaza, Kuala Lumpur's great cathedral of consumer electronics, and came back looking personally insulted. Same specs as last year, he said. Why does it cost so much more?

Because something's outbidding you for the memory, I told him. And that something isn't a person.

I wasn't being cute. Back in February, Gartner reckoned rising memory costs would push the average PC this year 17% above last year's price, and the average phone 13%. In June, Counterpoint's memory tracker put close to six in every ten DRAM chips shipped worldwide into data centres and AI systems. Phones got about two. PCs got one.

So a boom that looks like it's happening on the far side of the world has already reached into the price tags at Low Yat.

The wildest ride anywhere this year

The same force sent the South Korean market on what was probably the most spectacular round trip of the year.

On 18 June the KOSPI closed above 9,000 for the first time in its history, driven mainly by Samsung and SK Hynix — the two companies that make the memory. By 28 July it had dropped more than 10% in a single session, tripping the exchange's circuit breaker and freezing all trading for twenty minutes. Next day, same thing again. Three days after that, the index jumped nearly 18% in one session: the biggest one-day gain it has ever recorded.

And even after that rebound, it's still sitting a long way below June.

One trigger, according to press reports, was a sudden fear that China would ramp up memory production and turn a scarce, high-margin business back into a commodity. Though as one analyst put it, the panic around AI right now "appears to be indiscriminate."

I don't give investment advice and I can't predict markets. But underneath that ride sit two entirely coherent arguments, one bullish and one bearish, and neither camp is stupid. What I want to do here is lay them both out.

To follow either one, though, you need to be clear about what these companies are actually selling.

Not software. Labour.

Picture a staffing agency turning up at your office with an offer. We have as many workers as you want. They start immediately, never take leave, never resign, never ask for a raise. You pay by the hour, and only for the hours you use.

That's the business AI companies are really in.

The hundred-odd ringgit a month you hand over for a subscription isn't where the interesting money is. The interesting money is metered: however much work you've got, you call in that many workers, and you pay for exactly what you use.

Regular readers will know I'm an AI dove. But let me be honest about one thing. Since late last year I — an astrophysicist — have essentially not written a line of code by hand. Which isn't to say the thing solves everything. I still have to watch it, check it, keep feeding it corrections, make sure what comes out is what I actually asked for. But the sensation is unmistakable: it's like having hired a software engineer out of thin air.

Two Bets: What Are They Betting On?

The bullish case is simple enough. If most of what humans have historically produced — code, copy, research summaries — can be handed to a workforce you can summon in any quantity, then there's hardly an industry these companies can't reach.

And this isn't hypothetical. Uber reportedly burned through its entire annual budget for AI coding tools in four months, then capped every employee at $1,500 a month — around RM6,100 — for each individual tool they use. What matters about that number isn't how big it is. It's that it now sits in a formal budget, on the same spreadsheet as salaries and office rent.

Dove though I am, I don't dispute the premise. Even if large language models stopped improving tonight, what they've already done to the way people work will probably outlast the arrival of the internet.

And what that creates isn't only a market for AI services. These workers don't eat or sleep, but they do need somewhere to live — not dormitories, but data centres. So memory, processors, electricity and land all get bid up together. Your laptop got dearer and the Korean market went on a rollercoaster for very nearly the same reason.

None of this is a distant story for us. By industry estimates, Johor now has the largest pipeline of data centres under development anywhere in Asia-Pacific: what Malaysia is selling is housing for this digital workforce. The price is electricity. On figures the energy transition minister gave in July, data centres currently take 7% of the peninsula's power, and by 2035 that rises to 31%.

If it's all so unstoppable, why does the thing keep crashing?

Because supply can scale furiously and demand can't. Being able to hire ten thousand workers is not the same as having ten thousand jobs to hand them.

Three things get in the way.

First, this workforce is lopsided. As I argued a couple of columns back, outside of text AI is still a long way behind people at seeing and at moving.

Second, there's only one manager. However many assistants you've got, the hours a human can spend supervising them are finite. The genuinely hard work still falls apart when nobody's in the room.

The third is the interesting one, because it's hiding on the back of the numbers I just quoted. Ramp, which tracks AI spending across tens of thousands of American businesses, found the heaviest-spending 1% laying out several hundred thousand ringgit per employee per year. The median company spends under a thousand ringgit on the same employee over the same year. A few hundred times apart, the two ends.

Read that gap as a bull and you see a median that could grow several hundredfold. Read exactly the same gap as a bear and you see a party almost nobody has been invited to yet. One set of numbers, two readings — which is precisely where the rollercoaster comes from. And the hardest thing to judge is which part of all this is heat and which part is the actual waterline.

Strip it back and the argument comes down to two wagers.

The first: that AI, once it's clever enough, can improve itself. That's the phrase doing the rounds in Silicon Valley at the moment — RSI, recursive self-improvement. In May a startup called Recursive Superintelligence reportedly raised $650 million, about RM2.7 billion, with no public product yet; the industry has even coined a word, neolab, for outfits that do research and are in no hurry to ship. If that wager lands, the lopsidedness sorts itself out.

The second: that new demand gets conjured into existence, the way the internet conjured it. Some directions have no natural ceiling, and research is one of them — from cancer drugs to the far edge of the universe, there's no upper limit on what people want to know. Which is why, between April and June, OpenAI, Google and Anthropic all launched tools aimed squarely at scientific work. I make my living in that field, so I'm glad to see it.

But wagers are wagers, and the bill still lands. Alphabet alone spent $91.4 billion last year, and in July raised its guidance for this year to somewhere between $195 and $205 billion. Moody's reckons six large cloud and infrastructure companies will together approach a trillion dollars of capital spending next year — on returns it politely calls unclear.

So my own position hasn't shifted. AI will generate enormous value; I've no doubt about that. Whether that value happens to equal today's valuations is a different question entirely. What the two camps are really arguing about was never whether AI works. It's when the accounts balance.

So are the jobs going?

This time last year, more or less everybody was announcing mass unemployment. This year, plenty of them have quietly changed their minds.

The data is duller than either slogan. A Stanford group that traced American employment records found that in occupations where AI is best at replacing people, employment among 22-to-25-year-olds has dropped noticeably — while in occupations where AI mostly assists people, employment has grown across every age group.

So what we're looking at right now isn't every job going at once. It's a shock landing on particular ages and particular tasks. Inside a single industry, the person being replaced and the person being amplified may be sitting at adjacent desks. The on-ramp has undeniably got steeper, which is hard on anyone just out of university and deserves to be taken seriously.

With anything new there's always more noise than signal, and the only defence against noise is to go and understand the thing yourself. The internet rewrote how everybody worked, and society found a new equilibrium in the end. My own research practice has been rebuilt from the ground up — different questions, different tools — and the upshot is that I've never been busier in my life.

So the next time memory prices climb at Low Yat, don't rush to file it under good news or bad. What you're watching is the whole world racing to put up dormitories for a new workforce.

The dormitories will keep going up. That much is barely in question. Whether they ever fill, and whether there turns out to be enough work to go round — that's the bet underneath the whole ride.