# Discussion export — from “藕断丝连” to a dependency map

Date crystallized: **2026-07-25**

## The observation

Robin noticed a mismatch between political language and business reality.
Government competition between the United States and China is plainly
non-cooperative, yet companies on both sides remain connected through chips,
tooling, software, cloud platforms, contract manufacturing, vehicle systems and
consumer distribution.

The relationship is **藕断丝连**: the lotus root may be cut, but the fibers remain.

That observation led to a more useful research question than “Who wins the AI
race?”:

> Where do physical capability and services cross the boundary, where does money
> return, and which policy action could make that particular bridge fail?

## Why Robin finds this unusually natural

The project sits at the intersection of Robin's identities and instincts:

- a U.S. and Hong Kong citizen looking at both systems from the inside;
- an engineer who sees systems as pipes, interfaces, constraints and failure
  modes;
- a long-duration allocator more interested in architecture than daily noise;
- a believer that AI/robotics is a competitive capital race, not a globally
  coordinated efficiency project.

The same systems instinct appeared in the discussion of efficient AI models.
Cheaper inference does not automatically reduce chip demand. When national and
corporate competitors face a strategic race, efficiency savings can be reinvested
into more models, more agents, more simulations and more physical deployment.
The relevant frame is therefore not only Jevons' paradox; it is also a
prisoner's-dilemma-style capacity race.

## The Matrix joke that became less silly

Robin joked that *The Matrix* may have misunderstood its own engineering: human
bodies are poor electrical batteries, but biological tissue or organoid systems
can be imagined as unusual information-processing substrates.

The serious insight is not that a “human compute farm” is imminent. It is that
the substrate assumption matters. As long as frontier AI and world models remain
implemented primarily on silicon, gains in algorithmic efficiency do not remove
the strategic need for physical compute. A truly substitutable biological,
optical or neuromorphic substrate would be a different kind of shock—closer to
electric vehicles displacing oil demand than a more fuel-efficient combustion
engine.

## The proposed research system

The first idea was to scan “the core 100” on each side and draw a relationship
chart. We refined that into a reproducible design:

1. Freeze an official 100-company universe on each side.
2. Select 20 AI/robotics-relevant pilot companies from each frozen universe.
3. Build an evidence ledger before drawing any edge.
4. Separate three things that are often collapsed:
   - physical goods and services;
   - reverse commercial consideration;
   - policy-fracture risk.
5. Let blank cells remain blank when public evidence is insufficient.
6. Preserve source links and evidence dates so the map can be audited and later
   refreshed.

## Universe decision

### United States side

Use the official Nasdaq-100 constituent/weight snapshot. The official file
contains 101 securities but 100 companies because Alphabet has two share classes;
combine those two classes at the issuer level.

### Hong Kong side

Use the official Hang Seng China (Hong Kong-listed) 100 Index as the “HKEX
Core-100.” It is an exact 100-company, rules-based universe of large and liquid
Mainland companies listed in Hong Kong. This is more reproducible than inventing
a “top 100” from a live market-cap screen.

## The map's ontology

Each company is classified by role rather than by an overly broad “technology”
label:

- compute and semiconductor IP;
- chipmaking equipment and EDA;
- cloud/model platform;
- device and systems manufacturing;
- EV/autonomy and robotics;
- sensors/components;
- battery/energy infrastructure;
- consumer/distribution platform.

Each relationship records:

- provider and recipient;
- what crosses the boundary;
- whether evidence is current, announced or historical;
- the source and evidence grade;
- commercial consideration direction;
- whether the amount is disclosed;
- the relevant policy fracture;
- what the source does not prove.

## Early thesis

The U.S.–China AI/robotics economy is neither integrated in the old globalization
sense nor cleanly decoupled. It is a selectively constrained network whose most
important fibers are:

1. compute architectures and accelerators;
2. cloud and AI software stacks;
3. electronic-design software and semiconductor tools;
4. autonomous-vehicle compute;
5. contract manufacturing, optics and batteries.

Policy acts on specific edges, not on an abstract “relationship.” The useful unit
of analysis is therefore the **dependency edge plus its substitution path**.

## What this pilot does not claim

- It does not measure undisclosed dollar flows.
- It does not prove that every historical supplier remains active.
- It does not treat a corporate announcement as proof of shipment volume.
- It does not assume that all network activity creates shareholder value.
- It does not rank stocks or recommend trades.

The purpose is to make the hidden architecture visible enough to ask better
questions.