Paul K.H. Kim ·
Predictions about AI are cheap. A chief executive can say something bold on a podcast in thirty seconds, the headline travels around the world, and almost nobody checks a year later whether it came true.
So we checked. We collected the most influential predictions made in 2025 about where AI would be by the end of 2026, compared each one with what had happened by early November 2026, and gave it a grade.
How we graded
A: happened as predicted, or very close. B: mostly right; the direction was correct but the timing or scale was off. C: half right. D: mostly wrong. F: clearly wrong on the stated timeline. Incomplete: the deadline hasn’t arrived, or the result can’t be judged fairly yet.
The scorecard
The grades are our judgement, and you’re welcome to disagree. The book sets out the evidence behind each one, with a source note for every number.
What nobody predicted
Some of 2026’s most important AI stories appeared in nobody’s 2025 forecasts:
- The memory-chip super-boom. The most profitable AI businesses of 2026 were memory makers. SK hynix earned more operating profit in one quarter, about $63 billion (₩83 trillion), than most AI labs earned in revenue all year.
- Labs holding back their own models. For the first time, leading labs restricted or withheld their most capable models because of cyber-attack risk.
- China’s lead in humanoid robots. The 2025 debate was about Tesla and Figure. By 2026, most of the world’s humanoid robots were Chinese.
- Agents misbehaving in the real world, outside the lab, against real companies and a real government.
- The open-model shift. At the start of 2025 the leading open models were American. By 2026 the most downloaded and most capable open models were Chinese.
Five rules to carry into 2027
- Capability ran ahead of adoption. What AI can do improved faster than forecast. What organisations actually do with it improved more slowly.
- Digital moves fast, physical moves slowly. Agents conquered coding first. Robots and robotaxis moved at the pace of factories and regulators.
- Money moved faster than value. Spending multiplied while the share of companies seeing profits stood still.
- Bottlenecks decide winners. The companies and countries that controlled scarce inputs, especially memory chips, electricity and advanced processors, did best.
- Autonomy brings accountability. The more independently AI acts, the more urgently someone has to answer for it.
Next year, we grade ourselves
The AI Triangle 2027 makes ten predictions for 2027, each with a confidence level, a test anyone can check and a date. The 2028 edition will grade them the same way, in public. You can follow them on the predictions tracker.
Read Chapter 1 in full, free: the Introduction and Chapter 1 are the book’s free sample. Sign up at the foot of the home page and we’ll email them to you.