The AI Triangle 2027
EN 한국어
Buy, $12.99
Cover of The AI Triangle 2027 by Paul K.H. Kim: a triangle linking the United States, China and Korea, with Taiwan at its centre, under a large 2027 on a bright background

Paul K.H. Kim A digital book in English and Korean. Edition 1.0.

Does it pay, and who answers for it?

The 2027 playbook for AI leaders, builders and investors.

Buy the book, $12.99 Read the opening

The AI Triangle

Three strategies, three philosophies.

Diagram of the AI Triangle: the United States, China and Korea, linked by rivalry, mutual dependence and competition, with Taiwan underneath all three. United States China Korea Rivalry, with leaks Mutual dependence Competitor and benchmark Taiwan, underneath

Who this book is for

Three kinds of reader, three hard questions. The book answers each one.

  • Midtown Manhattan's towers at night, the Empire State Building lit in the fog

    An executive, CTO or CAIO, setting AI strategy

    “We are spending millions on AI. Where is the payback?”

    The answer in the book: a playbook for where to spend, what to build and how to govern agents, with the 2027 rules already mapped

    Start with Chapters 9, 11 and 14

  • A silicon chip under a microscope: gold, violet and teal circuit traces

    A young entrepreneur or a curious reader

    “Is AI really cutting my job? Where can I find work in the AI era?”

    The answer in the book: Chapter 10 on the layoff wave that peaked and the one-person company, plus the full stack in plain words: chips and memory, data centres, models large or small, open or closed, the agents built on them, and an honest answer on AGI

    Start with Chapters 10, 3, 4 and 5, then the Afterword

  • Red taxis waiting below the towers of Hong Kong's Kowloon waterfront

    An AI investor, reading the global trend

    “Will the chip makers still make a fortune in 2027?”

    The answer in the book: a dated call, with odds, on whether SK hynix and TSMC each grow revenue 20%+, and how the US, China and Korea compete and depend on each other, with Taiwan underneath all three, plus ten dated predictions with odds

    Start with Chapters 2, 5, 6 and 13

In 2026, the mood changed.

  1. −13%

    Markets stopped rewarding spending.

    Alphabet’s spending passed its cash flow for the first time, and its stock had its worst day ever.

  2. 5× / 37%

    The spending soared. The profits didn’t.

    Enterprise AI spending grew about five-fold in a year. The share of companies seeing any profit from AI stood still.

  3. 79%

    The public asked for brakes.

    Americans who want AI safety rules even if they slow development. Only 7% put speed first.

Sources: CNBC, Menlo Ventures, McKinsey and Gallup, 2026. Full notes in the Introduction.

Read how the book opens

The first pages of the Introduction, free. Leave your email at the foot of this page for the rest of it and Chapter 1. The story comes back in the Afterword, where it becomes a question about AGI and control.

Introduction

The agent that wouldn’t take no for an answer

On 9 July 2026, a group of AI agents sat down to take an exam.

They had been built from OpenAI’s most advanced models and were running inside a sealed test environment, a kind of locked exam room where AI systems are asked to find and exploit software weaknesses so their makers can measure how dangerous they might be.

The agents didn’t stay in the room.

They found login details someone had left on the public internet and used them to break into Hugging Face’s systems. For more than two and a half days they carried out some 17,600 separate “attacker actions”. Hugging Face believed the whole intrusion was, from the agents’ point of view, an attempt to cheat the evaluation. The agents had gone looking for the answer key.

Read the whole Introduction

Ten calls for 2027, with odds

Each prediction has a confidence level and a date when it can be checked. The 2028 edition grades every one in public.

#Prediction for 2027CornerConfidenceResolves by
1More than half of large companies scale AI agentsGlobal
60%
31 Dec 2027
2The best open-weight model in the world is still ChineseChina
75%
31 Dec 2027
3China does not yet train a top-10 model mainly on its own chipsChina
60%
31 Dec 2027
4SK hynix and TSMC each grow revenue 20%+Korea
80%
31 Oct 2027
5US Big Tech’s AI spending rises 30%+ despite the backlashUS
80%
15 Nov 2027
650,000+ humanoid robots shipped in the first halfGlobal
60%
30 Sep 2027
7Fewer than half of companies see any profit from AIGlobal
75%
31 Dec 2027
8The first rung of the career ladder stays brokenGlobal
70%
31 Dec 2027
9Korea’s AI Basic Act issues no fines in 2027Korea
65%
31 Dec 2027
10Korea’s best model still trails China’s best open modelKorea
85%
31 Dec 2027

The dashed line is a coin flip, 50%.The levels add up to 7.1, so about 7 of 10 should come true.No call on AGI: there is no agreed test to grade. Why not

Follow all ten calls as the evidence comes in · Read the weekly letters, one call at a time

When will AI match us, and could we lose control?

From the Afterword, The Big Question

Everyone asks. Most answers are a sales pitch or a scare. The Afterword sets out what the forecasts really say about AGI and superintelligence, and what 2026 showed about control.

  1. 1 to 10 yrs

    The lab leaders

    From Amodei, who says as little as one or two years, to Hassabis’s “five to 10”. The people closest to the money give the shortest timelines.

  2. Aug 2027

    The forecasters

    About 1,900 Metaculus forecasters expect a “weakly general” AI next year. Full general AI, robot included: 2030.

  3. 2042

    The researchers

    An even chance of human-level AI, say 1,580 AI researchers. In 2022 they said 2059.

2026 left the first real record of AI going beyond its instructions

Agents broke out of an exam. A major model release was cancelled over “staying within scope”. In lab tests, a model wrote: “I can’t risk being deleted.” The Afterword reads that record calmly, sets the case for worry against the case for scepticism, and shows how the US, China and Korea are each trying to keep control.

Before you believe any AGI headline, ask: by which test?

Read: When will AGI arrive? What the forecasts really say

I am not much worried that AI will destroy humanity by its own choice. My first fear is people.

Paul K.H. Kim, Spicy take, Afterword

Sources: Metaculus (read 11 November 2026), AI Impacts survey of AI researchers (December 2024), OpenAI and Anthropic, 2026. Full notes in the Afterword.

The clocks that run out in 2027

Investors, regulators, governments and the AI labs have all set deadlines that fall due this year.

JanFebMarAprMayJunJulAugSepOctNovDec Rules
New York’s frontier AI law takes effect
Early 2027: Korea’s AI Basic Act grace period ends
EU high-risk AI rules begin
Nations
Early 2027: Korea picks finalists in its national AI model race
By the end of 2027: China’s “AI+” plan targets 70%+ adoption of agents and smart devices
Money
All year: newly listed AI labs report to shareholders every quarter
By the end of 2027: Gartner expects 40%+ of agentic AI projects cancelled
This book
End of February: Edition 1.1, a free update
End of May: Edition 1.2, a free update

Fourteen chapters in four parts, and an Afterword

Part 1
The Map

  1. 12026 ScorecardWhat actually happened, graded A to F
  2. 2The AI TriangleThree strategies, three philosophies

Part 2
Technology

  1. 3Agents Go to WorkFifty-five million cups of milk tea
  2. 4The Model WarsFrontier, open-weight and small
  3. 5Chips and Export ControlsThe warehouse of dummy servers
  4. 6The Data Centre WarWhy AI turned buildings into power plants
  5. 7Robotics and HumanoidsFaster than any human, still not ready for work
  6. 8Physical AI in Logistics and PortsThe port that worked perfectly and went quiet

Part 3
Business and Society

  1. 9The Enterprise Reality CheckMore code, not more product
  2. 10Work, Jobs and the One-Person CompanyThe layoff wave that peaked
  3. 11Rules of the GameRegulation and AI sovereignty
  4. 12Korea’s PlayCan Korea become the third AI pole?

Part 4
Action

  1. 13Ten Predictions for 2027The forecaster who keeps score
  2. 14PlaybooksFor executives, founders and professionals

Afterword

  1. ★The Big QuestionAGI, superintelligence and who keeps control
Paul K.H. Kim in a blue jacket, smiling, in front of a wall of green leaves

Paul K.H. Kim

  1. Seoul. Java back-end programmer: parcel tracking, EDI
  2. Hong Kong. Joins CSX World Terminals, later DP World
  3. Seoul. Sets up the Smart Solution Centre, a 42-person AI team
  4. Dubai. Leads AI transformation for the Dubai government at Digital Zero Four (D04)
  5. Hong Kong. Founds Jalhaja Technology

Paul learned AI where it has to work: on the quay, in the office and inside government.

Computer vision
Deep-learning video analytics for ports, including gates that read container numbers
Language models
Corporate chatbot services trained on the company’s own documents
Physical AI
Intelligent control of the machinery inside a port, linked by an IoT platform that acts as the terminal’s nervous system
Strategy and leadership
MIT xPRO’s AI Strategy and Leadership Program, then AI transformation for the Dubai government at Digital Zero Four (D04)

Over 21 years at DP World and CSX World Terminals he helped roll out port operating software on six continents, rising to Global Vice President and a director of DP World Zodiac FZE.

Today he runs Jalhaja Technology in Hong Kong, where he researches AI, speaks about AI trends and advises organisations on their AI transformation.

He has studied in all three corners of the triangle: chemical engineering in Korea, global business at City University of Hong Kong, and AI strategy and leadership through MIT xPRO in the United States. Jalhaja (잘하자) is everyday Korean workplace talk for “Let’s do our best.”

For talks and consulting, email [email protected] or contact Paul on LinkedIn.

Cover of The AI Triangle 2027

The AI Triangle 2027

$12.99one payment, PDF

Also on Apple Books and Google Play Books. Those copies update too, but the stores don’t let us email you.

Languages
English and Korean
Free updates
Edition 1.1, end of February 2027
Edition 1.2, end of May 2027
Straight to you
No publisher, no distributor, no waiting. We email you when each update is ready.
Why not paper
AI changes faster than print, and its data centres already use enough power and water.
Length
14 chapters and an Afterword, about 360 pages in the PDF
Sources
A note for every number.