Robin ChienTaipei

Home / Think

09:00

Thinking ahead

AI, power and buildings are colliding. These are the questions I can’t stop thinking about, and the calls I’m willing to be checked against.

Data-centre electricity, world TWh · IEA, Energy and AI (2025) 415 · 2024 945 · 2030 projected GRID CONNECTION FIRST SITE MEMO · PAGE 1substation · feeder · date N 0 50 100 m Illustration · data: IEA, Energy and AI (2025)

Essays

Where AI, power and buildings meet.

3 min read

My chief of staff is a team of AI agents

Before I wake up, a small company of AI agents has already read the news, checked the facts and drafted my briefing. Building it taught me more about management than about AI.

Read the essay
3 min read

Power is the new location

Taiwan quietly stopped connecting large data centres in its north. Here’s what that says about real estate in the age of AI.

3 min read

An empty field, three years, ten thousand people

What building a semiconductor campus from bare ground taught me about scale. The hard part was never construction. It was sequence.

3 min read

Write the failure first

Pre-mortems and anti-goals are how I make decisions I can defend before I know how they turn out. Here is how I use them, with this website as the example.

3 min read

The dashboard was green

The hardest thing about running AI agents isn’t getting them to work. It’s getting them to tell you honestly when they don’t. I learned how to handle that in buildings long before I needed it for software.

3 min read

Check or go

The standard advice is to keep a human in the loop on everything an AI does. I tried that. It turned me into the bottleneck. Now I put myself only where it counts, and let the rest run.

3 min read

The agent is rented. The data is owned.

AI will change commercial real estate, but not the way the conference slides suggest. The software will get cheap. What stays scarce is clean building data, a referee nobody pays to be optimistic, and people who can turn data into capital decisions.

Polite failures

The dashboard was green. The room wasn’t.

AI agents rarely fail loudly. The expensive failures arrive on time and say everything is fine. Run the audit.

Read: The dashboard was green

FLEET STATUS · 03:50

4 of 6 healthy · as reported

Tap any line to see what was really happening. These are patterns from my own system, simplified.

Check or go

Where should the human actually be?

Ten things an AI assistant might do. Decide which ones need you. Most people put themselves in the loop far too often.

1 / 10

Rename and tidy files in a project folder.

Would you let your AI do this on its own, or make it check with you first?

My rule: Reputation or money: check. Anything that can’t be undone: check. Everything else: go.

Ten decisions. Let’s see where you’d put yourself in the loop.

Read: Check or go

Power, mapped

The site is chosen by the power plant. The zoning follows the substation.

Taiwan’s north stopped connecting large data centres. Tap a fact to see where it happens.

Taoyuan Datan Hsieh-ho Taipei Taichung Kaohsiung 5 MW+ freeze, 2024

Schematic, not to scale. Blue lines show supply sent north from central and southern Taiwan. Sources are listed in the essay.

Read: Power is the new location

Two power numbers

The application is an option. The connection is the asset.

In Taiwan, data centres have asked for fifteen times more power than is actually switched on. Underwrite the smaller number.

Read: The agent is rented
Requested
4,523 MW
Receiving power
303.5 MW

75 data-centre requests to Taipower as of 31 July 2026, against the contracted capacity of the four sites actually receiving power. Source: Taipei Times, 29 Aug 2026.

On the record

Six calls, with due dates.

I’d rather be specific and wrong than vague and safe. Each call says when to check it.

Calls I’m on record with

“On time” means I still hold the view. If I change my mind, the call stays here, marked.

MadeCallDueStatus
Oct 2026InfrastructureIn Taiwan, the first page of every data-centre site memo will be about the substation, not the land price.Taoyuan counted about 624 MW of data-centre demand on roughly 50 hectares in July 2026. The land was the easy part.2027On time
Oct 2026EnergyNight becomes the new peak for Taiwan’s grid.AI loads run around the clock; solar doesn’t. Forecasts show night-time reserve margins below the 15% benchmark through 2029.2028On time
Oct 2026BuildingsMost energy savings in existing buildings will come from software laid over equipment already installed, not from replacing it.Field studies of thousands of buildings show median savings around 9% from fault detection software alone.2028On time
Oct 2026Real estateWell-run buildings will be valued for their data layer as much as their location.When power is scarce, a building that can prove and control its energy use is the one you trust with the megawatts.2030On time
Oct 2026AI‘Can you run a team of AI agents?’ becomes a standard question in management interviews.The skills that make a good manager of people (clear roles, honest review, pruning, knowing when to step in) are the same ones agents need.2028On time
Oct 2026AIVerification becomes the scarce skill. The model that checks the work will matter more than the model that writes it.In my own agent system, the most valuable rule is that the builder never grades its own work.2027On time

Field manual

Eight rules for working with AI.

What running my own agents taught me, in a form you can use with a single chatbot on Monday.

01

Spec first.

Tap for why
Rule 01

Problem, constraints, definition of done. A vague request gets a confident, wrong answer.

02

The builder never grades its own work.

Tap for why
Rule 02

A second model, or a second pass with a different brief, checks the first.

03

Context is a budget.

Tap for why
Rule 03

An agent that reads everything understands nothing. Give it what the task needs.

04

Put the off switch where work leaves.

Tap for why
Rule 04

Guard the point of publishing or paying, not the door the request came in through.

05

Prune memory on purpose.

Tap for why
Rule 05

What an agent remembers forever eventually misleads it.

06

Every new tool is a maintenance bill.

Tap for why
Rule 06

Count the upkeep before you celebrate the capability.

07

“Unverified” is a task.

Tap for why
Rule 07

Never let it sit as a status. Someone goes and confirms it.

08

Enforce rules in code.

Tap for why
Rule 08

A policy the system can ignore will eventually be ignored.

Frameworks I use

Three tools I reach for.

01

Pre-mortem

Assume the project already failed, then explain why.

Hover or tap: when to use it
Use it

before committing money, time or reputation to a plan that’s hard to undo.

How it works →
02

Anti-goals

Decide what you refuse to become while you chase what you want.

Hover or tap: when to use it
Use it

when setting goals for a year, a role, or a big project.

How it works →
03

Capital allocator or technician?

Before you fix it yourself, ask whether a system should fix it instead.

Hover or tap: when to use it
Use it

whenever you catch yourself solving the same problem by hand again.

How it works →