The Hong Kong Situational Awareness Portfolio: 19 Stocks Behind China’s AI Build-Out

Below the Models, Part 2 |
HKDS Original Research |
By Petra van der Vegt |
Published 15 September 2026

At the end of

Part 1: Leopold Aschenbrenner’s Portfolio and the AI Infrastructure Thesis ,

I left you with one question:

How much of an AI portfolio really needs to be invested in AI-model companies at all?

I expected the answer to be “quite a lot.”
After going through the Hong Kong market, with some AI assistance to widen the search and organize the research, I no longer think that.
The first 60% of this simulated portfolio is now allocated, and almost all of it sits in the chips, servers, networks and data centers needed to build and run AI.
Optical connections. Printed circuit boards. Memory-interface chips. Semiconductor manufacturing. AI processors. Servers. Data centers. Cloud infrastructure. Telecom networks.
Nineteen Hong Kong-listed companies made it through this first stage.
Some names I expected to include did not make the final cut. Six of the twelve companies in the hardware section were not even available as Hong Kong-listed stocks a year ago.
That last point matters.
The investable Hong Kong AI infrastructure universe changed significantly in 2026.
This is the first published 60% of the HKDS Hong Kong Situational Awareness Portfolio.

HKDS Research Definition

The HKDS Hong Kong Situational Awareness Portfolio is a hypothetical portfolio of Hong Kong-listed companies selected for their exposure to the infrastructure required to build and operate AI computing capacity.
The research is inspired by the infrastructure logic in Leopold Aschenbrenner’s Situational Awareness thesis. Company selection and portfolio weights are original HKDS research.

Where Part 1 ended

In Part 1, Leopold Aschenbrenner’s Portfolio and the AI Infrastructure Thesis, I started with Aschenbrenner’s 2024 Situational Awareness essays.
The chapter that kept pulling me back was Racing to the Trillion-Dollar Cluster.
The scale of the argument is difficult to ignore.
If increasingly capable AI requires vastly more computing, then the spending moves into processors, memory, networking, data centers and electricity.
Then I looked at the public holdings of Aschenbrenner’s Situational Awareness fund.
Its Form 13F for 30 June 2026 showed something very similar.
SanDisk and Micron together represented approximately 55.77% of the disclosed long-stock portfolio when I calculated the positions from the filing.
Other holdings included TSMC, Bloom Energy, Nebius, CoreWeave and Applied Digital.
That gave me the basic research direction for Hong Kong:

Follow the requirements created by AI.

Part 1 mapped the idea.
Part 2 is where I start putting money into the simulation.

How this research was built

HKDS used AI-assisted research to widen the candidate universe, organize filings and identify possible infrastructure links.
Company roles, listing dates and material financial claims were then checked against company disclosures, HKEX filings and other primary sources.
Final company selection and portfolio weights are HKDS decisions.

What exactly are we building?

The portfolio rules are deliberately simple.

The weights reflect how directly I think each business connects to the infrastructure chain, how concentrated that exposure is, and how much uncertainty sits around the case.
They are research allocations.
They are not recommendations.

Hong Kong suddenly became much more interesting

This was one of the first things that changed my view of the exercise.
Six companies in the hardware section listed in Hong Kong during 2026:

• Shanghai Biren Technology on 2 January
• GigaDevice on 13 January
• Montage Technology on 9 February
• Victory Giant Technology on 21 April
• Luxshare Precision on 9 July
• Zhongji InnoLight on 30 July

A year ago, half of this hardware sleeve could not have existed in a Hong Kong-only portfolio.
That gives us access to areas that were previously difficult to reach through HKEX: optical interconnects, advanced PCBs, memory-interface chips and domestic AI accelerators.
It also explains why I wanted to revisit this now rather than simply look at the same familiar internet companies again.

The first 60%

Here is where the portfolio currently stands.

Looking at this table, one thing becomes obvious.
This does not look much like a conventional “AI stocks” list.
That is exactly what interested me.

38% in Data Movement & Hardware

3308.HK Zhongji InnoLight — 6%

The largest single position is Zhongji InnoLight.
This was one of the easier decisions.
Zhongji makes optical interconnect products that move data between servers, switches and other equipment inside cloud and AI infrastructure. Its current products include 400G, 800G and 1.6T optical modules.
In 2025, revenue reached RMB38.24 billion.
The business sits close to a simple problem created by larger AI clusters: more processors produce more traffic between processors, memory and networking equipment.
The data has to move.
That earns Zhongji the full 6%.

Victory Giant changed how I looked at this section

2476.HK Victory Giant Technology gets 5%.
Before doing this research, printed circuit boards would not have been the first place I looked for AI exposure.
They are now hard to ignore.
AI servers and high-performance switches require increasingly complicated PCBs connecting processors, memory, networking components and power systems.
Victory Giant produces high-layer-count multilayer PCBs and high-density interconnect boards, and its Hong Kong listing documents specifically identify AI and high-performance computing as important application areas.
Its 2025 operating revenue reached approximately RMB19.29 billion, up 79.8% from 2024.
That puts the company very close to the physical build-out.

Montage was another one I did not expect to like this much

6809.HK Montage Technology receives 4%.
Montage makes chips that help server processors communicate with memory.
That sounds like a small detail until you think about what happens inside an AI server.
More memory only helps if the processor can access it quickly enough.
Montage supplies DDR5 memory-interface chips, supporting chips and PCIe/CXL interconnect products.
The company itself points to AI training and inference as drivers of higher server-memory capacity and bandwidth requirements.
It sits between processing and memory.
That became a more interesting position the longer I looked at it.

The direct Chinese AI chip exposure

6082.HK Shanghai Biren Technology receives 4%.
This is the most direct AI-compute business in the hardware section.
Biren develops GPGPU chips and computing systems for AI training and inference.
Its BR106 is designed for large-scale AI training and inference workloads.
The attraction is clear.
So is the risk.
Biren is still investing heavily and remained loss-making in the first half of 2026 even as revenue grew sharply.
Technology changes quickly here. Financing matters. Export restrictions matter.
Four percent feels enough.

Luxshare is useful, but much broader

2475.HK Luxshare Precision gets 4%.
The communications and data-center part of Luxshare is what earns the position.
It supplies high-speed copper and optical connections, power products, thermal-management systems and server components.
For the first half of 2026, communications and data-center revenue reached RMB16.61 billion, up 49.7% year on year.
I originally considered weighting Luxshare more heavily.
Consumer electronics remain a major part of the group.
So 4% feels more representative of the exposure we are actually trying to capture.

GigaDevice deserves careful treatment

3986.HK GigaDevice Semiconductor gets 3%.
I do not want to overstate this one.
GigaDevice is not a Chinese equivalent of an HBM specialist such as Micron or SK Hynix.
Its memory business is mainly Flash memory and niche DRAM. It also produces microcontrollers, sensors and analog products.
Its product range includes NOR and NAND Flash, DDR3L, DDR4 and LPDDR4X niche DRAM.
The company has also introduced low-power storage products aimed at AI computing applications.
It earns a place.
The exposure is broader and less concentrated than the memory positions that stood out in Aschenbrenner’s portfolio.
That is why it receives 3%.

SMIC and Hua Hong: the foundry layer

I ended up thinking about 0981.HK SMIC and 1347.HK Hua Hong Semiconductor together.
Someone still has to manufacture the chips.
SMIC receives 3%.
In 2025, revenue reached US$9.33 billion and monthly capacity moved above one million standard 8-inch-equivalent wafers.
Its AI exposure is broad because the company manufactures chips for many industries.
What matters here is its role in China’s domestic semiconductor manufacturing capacity.
Hua Hong gets 2%.
Its platforms include embedded and standalone non-volatile memory, power discrete semiconductors, analog products and power-management chips.
One number caught my attention.
Revenue from its Analog & Power Management platform rose 41.4% in 2025, and the company identified AI supporting-power applications as one of the drivers.
That is a good example of what I mean by looking below the obvious AI names.
An AI cluster needs much more than the headline processor.

Packaging, fiber and servers

0522.HK ASMPT receives 2%.
As chips become more complicated, packaging becomes more important.
ASMPT supplies equipment used in semiconductor and electronics manufacturing. Its 2025 results showed momentum in advanced packaging, including thermo-compression bonding, with AI-related investment contributing to demand.
I see this as supporting exposure rather than a core AI holding.
Two percent is enough.
6869.HK Yangtze Optical Fibre and Cable also gets 2%.
YOFC moves us from connections inside the server toward the fiber between racks, buildings and computing locations.
Its 2025 reporting discusses rising demand for optical-fiber products used inside and between AI data centers.
The company has also shown connectivity systems incorporating specialized fiber and 800G/1.6T optical transceivers for AI computing centers.
This one is easy to understand once you see the architecture.
The machines need to talk to each other.
0992.HK Lenovo gets 2%.
I expected Lenovo to rank higher when I started.
Then I looked at the size of the whole business.
Its Infrastructure Solutions Group supplies servers, storage and high-performance computing systems, and the AI server business is clearly growing.
By the end of fiscal 2025/26, Lenovo said its AI server order pipeline had reached US$21 billion. Annual server manufacturing capacity exceeded 70,000 racks, including more than 11,000 direct-liquid-cooled racks designed for AI workloads.
Those are meaningful numbers.
Lenovo is still a very broad technology company.
So the AI infrastructure exposure gets 2%.
6088.HK FIT Hon Teng receives the final 1%.
Its cloud end-market revenue increased 37.6% in 2025, driven largely by AI server demand and higher shipments of copper-based components.
FIT also produces CPU/GPU socket products and liquid-cooling connectors.
The link is real.
The company also has large smartphone, automotive and electronics businesses.
One percent reflects that.

What are we really buying here?

When I put the companies next to each other, the chain became much clearer:
Foundry → processors and memory → memory interfaces → circuit boards → connectors → optical links → servers
That helped me separate the companies with very direct exposure from the ones where AI is one demand driver among many.
It also stopped the word “AI” from becoming a shortcut.
I want to know what the company sells.
I want to know where it sits.
And I want to know what has to happen for AI spending to reach its income statement.
That is much more useful.

Then the machines need somewhere to live

The next 22% goes into data centers, cloud infrastructure and networks.
This section contains some very different businesses.

GDS is the clearest data-center position

9698.HK GDS Holdings gets 5%.
GDS develops and operates high-performance data centers in China.
At the end of 2025, the company reported 668,283 square meters of net floor area in service, with 93% committed.
More computing eventually means more physical capacity.
That means buildings, electricity, cooling and connectivity.
GDS sits directly in that layer.
Five percent felt justified.

SUNeVision gives the portfolio something very Hong Kong-specific

1686.HK SUNeVision gets 4%.
SUNeVision operates carrier-neutral data centers in Hong Kong through its iAdvantage business.
Its latest FY2025/26 results show growing demand from hyperscale customers.
MEGA IDC Phase One reached 70% committed occupancy after a major customer commitment. The planned combined Phase Two and Three development is expected to add around 700,000 square feet of high-density, liquid-cooling-ready capacity.
SUNeVision plans total available power capacity of approximately 280MW as the campus expands.
For a Hong Kong portfolio, I wanted this exposure.
It gives us actual AI and cloud infrastructure physically located in Hong Kong.

China Telecom and China Mobile bring scale

0728.HK China Telecom receives 4%.
Its businesses now include mobile and fixed communications, cloud computing, computing power, big data and AI.
The company is building around an integrated cloud and computing network.
That gives the portfolio access to a large national network with computing infrastructure layered on top.
AI is only part of the company, so the weight stays at 4%.
0941.HK China Mobile gets 3%.
Its 2025 sustainability report stated that total intelligent-computing capacity had reached 92.5 EFLOPS on an FP16 basis across national and regional computing nodes.
That is difficult to ignore.
The telecom business is still enormous compared with the AI component.
So I kept it below China Telecom.

Kingsoft Cloud is smaller, but the AI link is much clearer

3896.HK Kingsoft Cloud gets 2%.
This is a more concentrated cloud-computing position.
Its 2025 revenue increased 22.8% to RMB9.56 billion.
In the second quarter of 2026, the company reported that gross billings from its AI cloud business increased 82% year on year and represented 56% of public-cloud revenue.
That is the kind of number I was looking for.
The business sits close to actual spending on AI computing resources.
The company is also financially less mature than China Telecom or China Mobile.
That keeps the weight at 2%.

China Tower was one of the harder calls

0788.HK China Tower receives 2%.
I would not describe China Tower as an AI data-center stock.
Its core business is telecom infrastructure.
At the end of 2025, it managed around 2.15 million tower sites.
The interesting part is what else those sites offer: power connections, distributed physical infrastructure and a developing edge-computing business.
China Tower has said it intends to expand its edge-computing activities.
That gives it a possible role further out in the infrastructure chain.
I am less convinced by this link than I am by GDS or SUNeVision.
That is why it gets 2%.

ZTE closes the network section

0763.HK ZTE receives 2%.
ZTE supplies telecom-network equipment, servers and computing infrastructure.
For 2025, the company reported revenue of RMB133.9 billion and identified computing as one of the main drivers of growth.
It sits across both connectivity and computing.
Geopolitical and export-control risks also deserve attention here.
I kept the allocation small.

And what happened to China Unicom?

China Unicom appeared in the preliminary list in Part 1.
It did not make this version.
That was one of the more useful parts of doing the exercise properly.
At the beginning, I had a broad list of companies that looked connected to AI infrastructure.
Once I started comparing them, some links were clearer than others.
Within the 22% available for data centers and networks, I found stronger or more differentiated exposure elsewhere.
China Unicom may still benefit from increased computing demand.
It simply did not earn a place in this version of the portfolio.
I like that.
A research list should be allowed to change.

So, where are Alibaba, Tencent and the AI models?

We are now 60% through the portfolio.
And we still have not allocated a meaningful part of it to AI-model companies.
That answers the question from Part 1 better than I expected.
A model may create the demand.
Then the spending travels.
More computing requires more accelerators.
More accelerators require more memory and interconnection.
Those systems need servers.
The servers need data centers.
The data centers need networks.
And all of it needs electricity.
That last part is where this gets even more interesting.

Why I kept leverage at zero

There is another lesson sitting in the background of this research.
Situational Awareness suffered a 67% decline in portfolio value during July 2026 after heavily leveraged AI positions moved against the fund.
Reuters later reported, on 11 September, that JPMorgan had ended its lending relationship with Situational Awareness following those losses.
I do not think that tells us whether Aschenbrenner’s long-term AI infrastructure thesis will eventually prove right.
It does tell us something about portfolio construction.
This simulation uses no leverage.
The maximum position is 6%.
A company can sit directly in the path of rising AI spending and still be a terrible investment at the wrong valuation.
That is something I want to keep separate from the infrastructure thesis itself.
First we map the companies.
Later we decide which ones deserve financial research.

Three things I learned building the first 60%

The first is that Hong Kong’s AI infrastructure universe is much better than it was a year ago.
The 2026 listings changed that.
The second is that some of the most interesting companies sit in places that barely sound like AI at first: optical connections, PCBs, memory interfaces and power management.
The third is that “AI exposure” needs context.
For Biren, AI training and inference sit close to the center of the business.
For Zhongji InnoLight, AI and cloud infrastructure are driving demand for high-speed optical connections.
For China Mobile, AI sits inside a much larger telecom operation.
For China Tower, the connection is further out again.
Putting all four under the same “AI stock” label would hide more than it reveals.
The weights are my way of showing those differences.

We have reached 60%. Now we need electricity.

This is where Part 3 begins.
The next 28% of the HKDS Hong Kong Situational Awareness Portfolio goes into Power, Grid & Thermal.
That takes us into a completely different part of the Hong Kong market.
Electricity generators.
Nuclear power.
Grid infrastructure.
Power equipment.
Cooling.
Thermal management.
The machines in Part 2 cannot run without them.
After that, 7% is reserved for Models & Applications and 5% remains in cash.
Then the full portfolio will finally be visible.
And I suspect the finished 100% will look even less like a conventional AI portfolio than I imagined when I started.

Coming next: Who Powers Chinese AI?

Part 3 of Below the Models will build the Power, Grid & Thermal sleeve.
Once that is complete, we can put the entire HKDS Hong Kong Situational Awareness Portfolio on one page.
Then comes the harder question:

Which of these companies actually deserves more research at today’s financials and valuation?

That is where the usual HKDS work starts again.
Safety.
Growth where it is relevant.
Value.
And probably a much shorter list.

Quick reference

What is the HKDS Hong Kong Situational Awareness Portfolio?

It is a hypothetical portfolio created by HongKongDividendStocks.com to map the infrastructure logic in Leopold Aschenbrenner’s Situational Awareness thesis onto companies listed in Hong Kong.

How many companies are included in Part 2?

Nineteen.

How much of the portfolio has now been allocated?

60%.
Data Movement & Hardware receives 38%.
Data Centers & Networks receive 22%.

Is this Leopold Aschenbrenner’s portfolio?

No.
The company selection and weights are original HKDS research.
The research is inspired by the infrastructure logic in Aschenbrenner’s Situational Awareness essays and by the shape of the publicly disclosed Situational Awareness investment portfolio.

Are these stock recommendations?

No.
At this stage, inclusion means a company has earned a place in the infrastructure map.
It says nothing yet about whether the shares are attractively valued.
That work comes later.

Sources and primary documents

1. Hong Kong Dividend Stocks.
   
   Leopold Aschenbrenner’s Portfolio and the AI Infrastructure Thesis
   ,
   Part 1 of Below the Models.
2. Leopold Aschenbrenner.
   
   Situational Awareness: The Decade Ahead
   ,
   June 2024, including Racing to the Trillion-Dollar Cluster.
3. U.S. Securities and Exchange Commission.
   Situational Awareness LP Form 13F, quarter ended 30 June 2026, filed 14 August 2026.
4. Reuters.
   11 September 2026 reporting on JPMorgan ending lending to Situational Awareness following the fund’s July AI-related losses.
5. Zhongji InnoLight.
   Corporate and investor disclosures; HKEX July 2026 listing documents.
6. Victory Giant Technology.
   2025 disclosures and Hong Kong offering documents.
7. Luxshare Precision.
   2025 Annual Report, H1 2026 results and Hong Kong listing documents.
8. Montage Technology.
   2025 Annual Report, product disclosures and Hong Kong listing documents.
9. Shanghai Biren Technology.
   Hong Kong prospectus and 2025 Annual Report.
10. GigaDevice Semiconductor.
    Company product information and Hong Kong listing documents.
11. SMIC.
    2025 Annual Report and HKEX disclosures.
12. Hua Hong Semiconductor.
    2025 Annual Report.
13. ASMPT.
    2025 annual-results announcement.
14. Yangtze Optical Fibre and Cable.
    2025 Annual Report and 2026 AI connectivity product disclosures.
15. Lenovo Group.
    FY2025/26 Annual Report and full-year results.
16. FIT Hon Teng.
    2025 Annual Report and annual-results disclosures.
17. GDS Holdings.
    2025 Form 20-F and investor disclosures.
18. SUNeVision.
    FY2025/26 final-results disclosures.
19. China Telecom.
    2025 Annual Report and 2026 interim-results disclosures.
20. China Mobile.
    2025 corporate reporting and 2026 HKEX disclosures.
21. Kingsoft Cloud.
    2025 Form 20-F and Q2 2026 results.
22. China Tower.
    2025 Annual Report and company investor disclosures.
23. ZTE.
    2025 annual-results announcement.

Important: This article is for informational and analytical purposes only. It does not constitute financial advice or a recommendation to buy, sell or hold any security.
The HKDS Hong Kong Situational Awareness Portfolio is hypothetical. Portfolio weights are research allocations created by HongKongDividendStocks.com and may change as company disclosures, business exposure and the research develops.