Preface Executive Briefing

Week of June 27, 2026

Beyond the US-China binary: who else is building AI power

The next AI winners may own the layer, not the leaderboard.

See Key Takeaways
Explore the briefing

The sovereignty gap

Four countries, four different layers of AI power

AI power is no longer concentrated in two countries. It is spreading because each of these nations chooses to own the one layer of the stack that matters most to it, instead of trying to win the whole thing.

Brazil

South America

Owns language relevance

Making AI understand Portuguese, local institutions, and public services matters more here than topping a global benchmark.

Japan

East Asia

Owns orchestration

The bet is on coordinating many specialised models and agents into reliable workflows, not on training one giant model.

South Korea

East Asia

Owns memory

A sovereign model sits on top, but the real leverage is HBM, the high-speed memory that AI chips depend on to run.

UAE

Middle East

Owns compute architecture

Research, sovereign cloud, compute partnerships, energy, and state demand are being assembled into one national deployment stack.

Reference points: Sakana AI (Japan), Upstage and SK Hynix (Korea), MBZUAI, G42 and Core42 (UAE), and the Latam-GPT initiative.

The stack trap

The model is the tip. The dependencies are the iceberg.

Benchmark scores are easy to compare, so leaders treat the model as the whole map. But using a model is not the same as controlling the stack beneath it, and that stack is where cost, availability, and resilience are decided.

Above the waterline

The model and its benchmark

What every leaderboard and launch headline shows you

  1. Cloud region and data law

    Where the workload runs, and which jurisdiction's rules govern the data.

    Sovereign cloudOracle

  2. Memory (HBM)

    The high-speed memory that keeps AI chips fed and busy.

    SK HynixSamsungMicron

  3. Advanced packaging

    Bringing compute and memory close together so they work as one.

    TSMC

  4. Lithography

    The machines that can print the most advanced chips at all.

    ASMLZEISS

  5. Networking and optics

    The switches, fibre, and optics that wire AI clusters together.

    BroadcomAristaCorning

  6. Power and cooling

    The energy and data-centre capacity that everything above needs.

    Grid and siting

The executive tool

An AI dependency map has six layers

The old question was which AI tool to buy. The strategic question is what each critical workflow quietly depends on. Trace it down the stack, one layer at a time.

01

Workflow criticality

Which business process is the AI actually supporting, and how much rides on it?

02

Data sensitivity

Is it public, internal, customer, regulated, or cross-border data?

03

Model and provider

Closed or open, hosted where, and could we switch providers if we had to?

04

Cloud and deployment

Does it run on global, sovereign, regional, private, or on-premise infrastructure?

05

Infrastructure and supply chain

Which chips, memory, packaging, networking, power, or export controls could affect cost or access?

06

Governance and accountability

Who approves, audits, and reviews the output, and who owns it when the model fails?

Use sovereignty selectively

Sovereign AI is a ladder, not a switch

Most organisations should not build their own model. Climb only as high as the risk justifies, and keep the freedom to switch everywhere else.

  1. Level 1

    Global AI SaaS

    Low-risk drafting, brainstorming, and public-information tasks.

  2. Level 2

    Enterprise cloud AI

    Internal workflows with contractual controls and enterprise data policies.

  3. Level 3

    Regionally governed AI

    Sensitive work needing local data residency or regulated-sector compliance.

  4. Level 4

    Sovereign or private deployment

    High-risk public-sector, financial, healthcare, legal, or national-data work.

  5. Level 5

    Locally controlled model

    Only where the organisation must control model, data, and governance end to end.

Do not let one model become your operating model. The goal is control where risk demands it, and optionality everywhere else.

Closer to home

For Hong Kong, the edge is integration, not ownership

Hong Kong is unlikely to own the full stack, and it does not need to. Its opportunity is to be the smart adopter and regional integrator, strong at bilingual, cross-border, compliance-heavy AI across the US, Mainland China, and global cloud.

US AI
Mainland China AI
Global cloud

Hong Kong

Bilingual, cross-border, compliance-heavy

The questions that decide the advantage

  • Which data can leave Hong Kong, and which cannot?
  • Which workflows can safely use global AI tools?
  • Which require China-compatible deployment?
  • Which need local or sovereign cloud?
  • Which AI systems can switch models if regulation, pricing, or access changes?

Key Takeaways

Own the layer, not the leaderboard

AI power is no longer only about model leadership. It is about stack position, dependency control, and deployment architecture.

01

Sovereign AI is selective, not full-stack.

Countries and companies build advantage by owning or influencing the one layer that matters most to them, not by winning every layer. Brazil picks language, Japan orchestration, Korea memory, and the UAE compute architecture.

02

The leaderboard is visible. The dependency map is strategic.

Benchmark scores make headlines, but resilience is decided below the model: data, cloud region, memory, packaging, lithography, networking, power, and law. Access to a tool is not control of the stack beneath it.

03

Design for optionality, governance, and chokepoint awareness.

Do not let one model become your operating model. Build the ability to switch providers, govern how data moves, and know which suppliers, regions, and regulations you quietly depend on.

FAQ

No. Most organisations should stay on the lower rungs of the sovereignty ladder and reserve full control for genuinely high-risk workloads. Sovereignty is selective, not the default. The cost, talent, and maintenance burden of a foundation model is rarely justified.