Is AI a Bubble? Start With a Plate of Fried Rice

When I was a kid, my mother could not eat out without auditing the entire table. "Twenty ringgit for fried oyster mushrooms? Oyster mushrooms aren't even expensive right now, and there's barely any labour in it. This is daylight robbery." In a food paradise like Malaysia, my mother's tongue is the fairest yardstick there is.

The logic is simple. If some uncle's fried rice is overpriced and lifeless — no wok hei, none of that smoky char that comes off a screaming-hot pan — the stall next door, cheaper and smokier and better, takes his lunch, literally. A loud signboard counts for nothing. In the end the customer's tongue and wallet answer only to the food, not to the name above the stall.

Underneath that yardstick sit the two most important words in any market — premium and arbitrage.

A premium is the gap between what something is priced at and what it is actually worth. My mother clocking that the mushroom dish is wildly overpriced — that's her smelling a premium. Arbitrage is what clever people do to pocket that gap. The stall next door sees your inflated price, undercuts you with something more honest, and walks off with your whole table. Bit by bit, the inflated price gets sanded back down to reality.

Here's the thing. In these last few years of AI sprinting flat out, nothing has swung more wildly than premium and arbitrage. So today, let me take those two yardsticks to three "dishes" of the AI era.

Dish one — what does your subscription actually buy?

Anyone vaguely keeping up has an AI subscription on their phone by now — ChatGPT, Claude, Gemini, take your pick — and the entry tier runs around a hundred ringgit a month. People like me, who write code all day, just max everything out; a thousand ringgit a month is nothing unusual.

But when you hand over the money, how many of us can actually tell whether it's worth it? How much electricity does a single AI query burn? What do the data centres and the chips behind it really cost? Put a plate of fried oyster mushrooms in front of my mother and she sees the markup instantly. Put AI in front of most people and they can't even tell you how the thing was built, let alone price it.

And the models come in sizes, with wildly different costs. A big model is the pricey abalone; a small one is the two-ringgit handful of enoki at the wet market. The catch is that the menu just says "AI" — you rarely get to see which one actually landed on your plate.

Which is how a whole new business was born. It's called model routing, and stripped of the jargon it's really just a head waiter who orders for you. Most people don't care which model answers, as long as the answer is good enough — and that indifference opens up an enormous premium, and an enormous space to arbitrage it. A platform like OpenRouter plugs hundreds of models into a single socket, picks whichever one gives you the most bang for your ringgit, and takes a cut for its trouble — a business now worth tens of millions of dollars a year. By industry estimates, simply routing the easy questions to smaller models can shave up to 80% off the bill.

All of that, at least, is out in the open. The subtler situation is the other one. Even when you're paying for the very same subscription, your hardest question might quietly get handed to a cheaper model without your ever knowing. Put it this way. You think the head chef is cooking, but the plate that lands in front of you was knocked out by the kid on the line. OpenAI's GPT-5 stumbled over exactly this last year — it is really several models stitched together, and for a while the router was shunting even hard questions to the small, cheap one. Users started shouting that "GPT-5 got dumber," and the company scrambled to fix it. Nor is this any one company's specialty. Anthropic, the maker of Claude, took its own beating over Claude "getting dumber" too — though it insisted the cause was engineering bugs and flatly denied swapping in a weaker model. Either way, the diner's nagging doubt is the same. Did the head chef actually cook this, or not?

That said, it isn't entirely their call. Order lazily and the kitchen reheats you yesterday's rice; order with intent and the head chef finally picks up the wok. Only when you're sharp yourself can you ask the kind of question that squeezes the full value out of that subscription. (How to ask a good question is its own sermon — I delivered it in an earlier column, so I'll spare you here.)

Dish two — the startups that "look" valuable

Scale the same yardstick up to investing, and the story gets more exciting.

AI is rewriting everything, including itself. It already writes vast amounts of its own code, it iterates at a frightening clip, and the engineers building it lean on it more by the day. The whole field reinvents itself so fast that almost nobody can honestly claim to know what it can and can't do right now — and wherever understanding runs out, premium breeds.

When AI is hot, valuations drift badly out of focus. Plenty of companies have no real moat, nothing a rival couldn't copy, and yet in front of hot money even seasoned investors get seduced by a shiny story — because the field moves so fast that a judgment formed three months ago may already be stale.

People come to me all the time, excited about something they're about to back. Often enough I'll spend half a day building a rough prototype, and they go pale. The company they were ready to pour serious money into turns out to have a shockingly shallow moat, because even a half-competent outsider like me can reproduce its "core feature" in one afternoon. That gap is the premium, naked.

Warren Buffett has a famous rule he calls the circle of competence — only invest in what you genuinely understand. The trouble is that in AI, people with both the technical depth and a real read on the market are vanishingly rare, and each of those takes the better part of a lifetime to build. This isn't about anyone being less clever; the combination is simply scarce, and in a market as young as Malaysia's it needs more time to grow.

None of which means the moatless startups are all doomed. Riding a premium is a perfectly legitimate way to make money — provided you know when to get off. An investment with no real value underneath it is a bit like bitcoin. At bottom, it's a game of chicken. As long as you're not the one left holding the bag when the tide goes out, you win, and that comes down entirely to your nerve and your luck.

Dish three — the thing being mispriced is also you

Look at it coldly and, in the job market, a person is a commodity too.

What worries me is how many students think they can run the same premium play. A nice degree, plus a hand from AI, and the transcript gleams, the output looks prolific, the surface impressiveness is genuinely hard to match.

In the short run, it works. I've sat on plenty of admissions and review panels over the years, and I've watched it happen too many times. A glossy personal statement, a row of pretty grades, a long stack of publications, and the panel is dazzled into a wildly generous "price," especially before people have worked out just how good AI is at surface polish. But that value only counts when there's something real underneath.

As more people catch up, the tide goes out, and what you're actually worth becomes visible to everyone. So I tell my students as much, plainly. If you want to bluff your way through with ChatGPT, I honestly don't mind, because in my grading, and especially in how I pick PhD students, anything I could produce myself with AI starts at zero by default.

It sounds harsh, but that's simply how the market clears. This surface premium will eventually turn on anyone who refuses to take themselves, and the world, seriously. Malaysia's particular trouble is that a credential like mine — which ought to be a foot in the door — gets treated as a free pass. A mature market should judge a brand-name halo more harshly, not more gently, the same way the bar at Din Tai Fung is rightly higher than at the stall down the road.

In the end, a name and a degree are both just signboards, a premium. However loud the signboard, if there's no real wok hei in the pan, the premium won't hold for long. A person, it turns out, isn't so different from a plate of fried rice.

Coda — after the tide goes out

Buffett has another line, even more famous than the first. "You only find out who is swimming naked when the tide goes out."

The AI tide is repricing everything — ideas, companies, people — at a speed we have never seen. A premium can hold for a while, but not forever; sooner or later the arbitrageurs turn up and sand the inflated part back down.

So whether you're using AI, investing in AI, or being repriced by it, what survives the bubble is never the loud signboard. It's the real wok hei at the bottom of the pan.

Because in Malaysia, in the end, whether a dish is any good is still my mother's call.