The Year AI Crossed the Mainstreaming Threshold

If you write about the AI industry, the hardest part is that right up until the very last day of the year, you never really know whether you'll get to do a proper year-end review — because AI could make another giant leap on December 31st and rewrite every argument you'd built. Take early 2025: out of nowhere, DeepSeek rewrote the whole narrative around open-source models. So you genuinely have to wait until the year is completely over before you can write the retrospective. Good news: 2025 is now, officially, done.

Felt in Your Gut: Engineers Get Repriced, Job Listings Shrink Fast

2025 was an extraordinarily hot year for AI. More than $200 billion in capital poured into the industry in a single year — a huge tailwind for Taiwan's stock market, since the root of all this demand is semiconductors and data centers, exactly where Taiwan is strongest.

On top of that, 2025 delivered a string of eye-wateringly valued AI software companies and equally eye-watering AI acquisitions — Windsurf mid-year, Manus at year-end, both $2-3 billion deals. Sky-high valuations became almost routine: two members who split off from OpenAI's board started AI companies now valued north of $10 billion each, never mind the countless smaller AI outfits worth a few hundred million.

But here's the interesting part: at the very same time, engineering job listings were shrinking fast enough that people could feel it. US engineering postings fell to a degree you could sense in your gut — some data put the decline at roughly 40% of technical roles. Corporate reports started citing AI directly as the reason for layoffs. Layoffs used to mean a downturn; now they genuinely mean being replaced by a machine, and companies are restructuring their workforces faster than ever. What stood out most clearly last year was a fork in engineer demand: demand for junior engineers kept falling, while at the very top — the people actually building the models — pay got eye-wateringly expensive.

The market has become distinctly "barbell-shaped": at one end, top talent gets repriced like professional athletes, with signing bonuses running into the tens of millions of dollars; at the other, companies decide that under the 80/20 rule, 80% of staff simply aren't needed — and lay them off outright.

Google's 2025 acquisition of Windsurf is the textbook case: Google effectively poached the talent it wanted through licensing payments, while only actually needing a fraction of the staff. Across companies like this, people are getting sorted into two buckets — "absolutely essential and extremely expensive" versus "relatively cheap and replaceable by AI."

After Big Tech's astronomical 2025 capex, something eventually has to come out the other end — and what came out was full-blown, industry-wide scaling of compute. That compute kept improving the systems, and AI agents got measurably more capable from the start of the year to the end. It's no longer the "copilot" people were picturing at the start of the year — it's shifted decisively toward the AI agent, which can now aim directly at completing multi-step tasks and is steadily edging toward doing the work humans used to do. Anyone who's used the pricier models has probably felt that shift viscerally.

Because starting in 2025 AI agents could directly execute a huge range of tasks, willingness to pay for AI — and annual recurring revenue (ARR) — kept climbing. Take Manus: even though most people still haven't started using it, its ARR hit $100 million in under a year. Given how few people are actually using the service, that number is staggering — which is presumably why Meta acquired it (though Chinese regulators are reviewing the deal, so its fate is still up in the air). Expect more application-layer acquisitions like this over the coming year.

Early 2025's DeepSeek shock touched off a round of debate over whether open-source or closed-source was the smarter bet, but the big tech companies kept building closed models regardless. Looking back over the full year, closed models seem to have come out ahead, because the moat that matters most is data and infrastructure. The leading AI companies' real advantage turns out to be their data and their user base. Elon Musk's xAI (Grok) clearly benefited from the sheer volume of data its huge user base generates, driving very strong growth for Grok in the second half of the year.

Truly Stunning: Hand 30% of Your Work to AI — and It Does It Better

Google Gemini, which already commands the world's largest user base, closed the gap fast in 2025 on the back of world-class infrastructure. Gemini has now caught up and then some, putting it in a genuine head-to-head race with OpenAI.

At the start of 2025, using AI from the big players still felt somewhat optional — a tool you could pick up or not. By year's end, it was viscerally clear how much of one's own work had already been handed off to AI. I'd guess roughly 30% of what I do is now genuinely better left to AI — and relearning how to use a computer, in that sense, is a remarkably strange feeling.

In 2025, AI's scaling and democratization genuinely happened — the entire industry has entered something like the moment smartphones went mainstream. Roughly 60% of American adults have now used AI at least once, and I doubt Taiwan's rate is far behind.

Starting in 2026, AI becomes a mainstream, default tool — nobody's going to specifically ask whether you use AI anymore. In 2026, AI becomes something built in: using a computer or a phone will simply mean using AI. The past year was the tipping point for AI going mainstream, and I'm genuinely bullish on how AI performs in capital markets in 2026. One more thing: starting in 2026, I think AI's value will gradually diffuse outward into the application layer — that's where I'd be watching for investment opportunities.