AI Has Entered the Second Half. Are You Ready?

Here's something you probably missed, but really shouldn't have. In February, over the span of 48 hours, the global software-subscription industry lost roughly USD 285 billion in market value — about RM 1.3 trillion. The selloff was savage enough that traders coined a name for it on the spot: SaaSpocalypse.

If "SaaS" — software as a service — sounds like industry jargon you can safely tune out, it really isn't. SaaS is just the business of paying for software by the month or the year. Zoom and Salesforce at work; Netflix and Spotify at home. For the past twenty years, this one business model has been quietly holding up something like half of Silicon Valley. In February, half of it gave way.

What pushed it over the edge was a quieter shift that has been building for about a year: AI's ability to write computer code has finally crossed a line. Several companies are racing on this, but the one out in front is Anthropic — maker of Claude, OpenAI's main rival. Wall Street looked at where Claude was heading, did the math for a few days, and arrived at an uncomfortable conclusion: a lot of these subscription-software businesses don't really have a future.

If markets are the most ruthlessly honest mirror we have for what is actually working in the economy, then this is the mirror saying, plainly: in 2026, AI has properly arrived. Not in the way that makes headlines — kids cheating on homework, scammers cloning voices on the phone. Those are real problems, but they're shallow ones. The deeper hit is the one I just described: an entire industry, uprooted overnight, with the writing on the wall for everyone still working in it.

Longtime readers of this column know I've always considered myself an AI "dove" — the camp that doesn't really buy the idea of AI flipping the world over inside two years. The "hawks" who pop up every few months announcing that the singularity is here have been one of my favourite things to tease. The funny thing about being a dove in the AI era, though, is that AI has a habit of embarrassing its skeptics faster than they can finish their column. Even I have had to raise an eyebrow more than a few times in the last two years. Which is exactly why I'm starting this new series — to step back and look, with a cooler eye and harder numbers, at what is actually happening.

So how did the software apocalypse really happen?

From co-pilot to driver

AI began as a language tool — writing essays, translating, chatting. But people figured out very early that one of its most useful applications was code. The reason isn't mysterious: programming is, after all, a language (we literally call them programming languages), so it falls neatly inside what AI is good at. And the modern world runs on code. Every WhatsApp call, every banking transaction, every budgeting app on your phone was, at some point, typed out a line at a time by a human programmer.

But there's a big difference between "AI can help you write code" and "AI can write code by itself".

For most of 2024 and into the middle of 2025, AI was a useful assistant — but the industry consensus was clear: it was a co-pilot, with a human firmly at the wheel. I can vouch for that. Using it back then was like text-message autocomplete: I'd start a sentence, it would finish it. Helpful. But if you took your hands off the wheel and let it drive, it would just loop around the same junction for an hour — more lost than a brand-new Grab driver in the back streets of old KL.

Somewhere in the second half of 2025, that quietly stopped being true.

The most-respected outfit benchmarking AI coding ability is a group called METR. They use a beautifully simple metric: with no human help, how much of an expert's working time can an AI independently get through? That number has been doubling exponentially for six years — historically every seven months. Over 2024 and 2025, it accelerated to doubling every four months.

In concrete terms: in early 2024, AI could independently handle roughly thirty minutes of an expert's work. In early 2026, that number is ten hours.

This sounds like one more line on one more boring chart. It isn't. It happens to hit software engineering right on its softest spot.

Imagine your junior engineer needs supervising every hour. You — the manager — are not getting anything else done. You're babysitting. Now imagine that same engineer can run for ten hours straight before needing input. The job changes shape entirely. You give them a brief in the morning, hear what they got done at the end of the day, and tee up tomorrow. In the first version of that story, you're training someone. In the second, they're actually doing your job for you.

The same logic applies to AI in code. The moment AI can run for ten hours unsupervised — the moment it works on day-long timescales rather than minute-long ones — it stops being a co-pilot. It's the driver, and the human has been bumped to the passenger seat. For an industry whose entire business model is people paying humans for software, that is an existential problem.

A small experiment of my own

A concrete example from my own desk. Over the past decade I've accumulated a stack of work-related software subscriptions, each running maybe RM 400 to RM 500 a year. Added up, the bill is enough to sting when the credit-card statement lands. Replacing them myself, though? Even if I had the skill, I never had the time.

Until this year. Over the past twelve months I've had AI write knock-off versions of about half of those subscriptions. AI does the bulk of the work; I drop in to patch the rough edges; one afternoon, done. One afternoon, in exchange for RM 500 a year, indefinitely. That is an irresistible deal.

When enough of the market starts doing the same math, you get exactly the headline I opened with: in February, the SaaS market quietly fired itself. I was actually muttering to colleagues late last year that "SaaS is going to crack". But muttering doesn't pay, and I never did get around to shorting any of these stocks. So now I just get to kick myself.

A side note. An "AI agent" is a particular way of using AI: instead of asking it questions and reading the answers, you hand it a whole task and let it go off and finish the job on its own. For that to actually work, the AI needs two things at once — day-long stamina, and a reliable long-term memory of what it has been doing. The breakout example of this new wave is the open-source agent OpenClaw, whose lobster logo has earned the whole category an industry nickname: "lobsters". The same underlying shift powers the recent buzzword "vibe coding" — meaning, roughly, "I don't really know how to code, I just describe what I want and let the AI figure it out".

Where this is going

The curve, as far as anyone can tell, is still holding. The reasonable extrapolation is that fairly soon, AI will be able to handle hundred-hour tasks unsupervised — that is, tasks measured in weeks, not days. In fact, this is quietly already starting. There's a running joke in the industry that, in software, humans aren't even co-pilots anymore. We're more like air traffic controllers, watching several AI pilots come down on several runways at once, hoping none of them collide.

The chain reaction is still rippling outwards. The current frenzy of corporate spending on data centres is essentially one giant bet: that once we cross from ten-hour tasks to hundred-hour tasks, a whole new generation of companies will pop into existence on the other side. I'm not in the business of giving investment advice — but my inner dove can't quite resist a whisper. How much new economic activity can hundred-hour-task AI really sustain? I have my own question mark over that one. The market gets to answer.

So that, at length, is what this column will try to do over the next several months. You've probably already heard fragments of this story — a headline here, a viral post there. What I'd like to offer, across twelve columns, is something a little different: a quieter, more quantitative way to track what is genuinely happening in the AI era, and the logic of what is likely to come next.

Because fear, in the end, is wasted energy. Your real opponent in this era was never some looming abstract "AI". Your real opponent is a group of other humans — standing behind the same cold logic, and just moving faster than you.