Every earnings season, I try to do one thing before I form an opinion:
Slow down.
Not to find the stock that moved the most. Not to chase the loudest AI headline. Not to pretend I can understand an entire sector from one revenue-growth percentage and a celebratory chart.
I slow down to separate fact from feeling.
This half-year review is focused on ten Hong Kong-listed AI names across computing infrastructure, semiconductors, memory, optical communications, PCBs, electronic materials, and foundation models.
These are some of the market’s hottest themes — which is exactly why they deserve colder thinking.
The central question is not simply:
Which company grew the fastest?
It is:
Which company is already earning real money from AI — and which is still selling the market a future that has not yet arrived?
As a Hong Kong Professional Investor, I have learned that an exciting story is not automatically a bad investment.
But an exciting story without cash flow, margins, or a credible return on capital can become an extremely expensive form of entertainment.
Two Different AI Trades
The clearest takeaway from this earnings season is that “AI” is not one investment category.
There are at least two very different businesses hiding under the same label.
The first is the AI application and model layer: companies building large models, APIs, enterprise services, and consumer AI products.
The second is the AI infrastructure supply chain: optical modules, PCBs, server components, electronic materials, foundry capacity, memory interfaces, and fibre networks.
Both can grow quickly.
But they carry very different economics.
For model companies, revenue may be rising at extraordinary rates while R&D, inference cost, sales expenditure, and computing expense remain even larger. The product may be real, demand may be real, and the loss may also be very real.
For infrastructure companies, the picture is more immediate. AI data-centre spending has already flowed into orders, pricing, capacity utilisation, gross margins, operating profits, and sometimes cash flow.
That does not automatically make hardware safer.
It just means the revenue has shown up before the dream.
MiniMax: Growth With A Bill Attached
MiniMax reported first-half revenue of approximately USD 117 million, up 283.1% year on year. Gross profit rose faster, by 464.8%, and gross margin improved from 12.1% to 17.9%.
That is meaningful progress.
But then I look at the bill.
The statutory loss was roughly USD 358 million. Adjusted net loss was around USD 293 million. R&D expenditure rose to approximately USD 297 million.
This is the central question I would keep asking myself: for every additional dollar of revenue, how much computing, research, and customer-acquisition cost is required?
High revenue growth is encouraging.
But high growth does not mean a company has escaped the “spend heavily first, explain profitability later” phase.
For MiniMax, I would watch gross-margin expansion, inference cost, customer retention, and whether losses narrow as revenue scales.
The business appears to be moving.
It is just not yet clear how much cash it will require to keep moving.
Z.ai: The API Engine Is Starting
Z first post-listing interim results were more encouraging at the commercial level.
First-half revenue reached RMB 954 million, up 399.7% year on year. Its open-platform and API business generated RMB 825 million, up 2,735.7%, representing 86.5% of total revenue. Token usage rose by more than 40 times, while average API pricing increased by 101%.
That is not merely a slide deck.
That is usage.
More importantly, the API segment’s gross margin moved from negative 0.4% to positive 24.6%. It may only be a first step, but moving from “every transaction loses money” to positive gross profit is a meaningful one.
Still, the consolidated gross margin fell from 50.0% to 26.4%, while R&D expense reached RMB 2.13 billion. The statutory net loss remained substantial at RMB 2.07 billion.
So I would not call the business proven yet.
I would call it commercially alive.
The next test is whether API scale leads to better overall margins, lower R&D intensity relative to revenue, and a path toward self-funded growth.
Revenue can grow at 400%.
The question is whether the economics eventually grow up with it.
The Hardware Reality
While AI-model companies are still proving monetisation, several hardware names have delivered something the market understands immediately:
Revenue.
Profit.
Cash flow.
And, in some cases, dividends.
建滔積層板（01888）reported first-half revenue of HKD 14.9 billion, up 55%, with net profit rising 209% to HKD 2.89 billion. AI servers, high-speed networking, electronic fibreglass, copper foil, and copper-clad laminates all benefited from stronger demand and pricing.
The company’s vertically integrated structure mattered. When competitors faced material shortages and delivery constraints, Kingboard could continue producing at near-full utilisation.
That is what an operating moat looks like in an industrial cycle: not a glossy presentation, but the ability to deliver when everyone else is scrambling.
長飛光纖光纜（06869） also posted striking numbers: first-half revenue rose 53.6%, while attributable profit jumped 889%. Gross margin reached 53.4%.
That sort of profit growth tells me product mix, pricing, and capacity utilisation have moved together.
But it also tells me to become more cautious.
When profits rise nearly ninefold, I do not only ask, “How good is this?”
I ask, “How much of this is repeatable?”
High-cycle earnings often look permanent at precisely the moment they are most vulnerable to competition, supply expansion, and price normalisation.
Cash Flow Is Harder To Fake
勝宏科技（02476） stood out because its earnings were supported by cash generation.
Revenue rose 28.8% to RMB 11.63 billion. Net profit rose 33.3% to RMB 2.86 billion. Operating cash flow was approximately RMB 2.91 billion.
For a manufacturing company expanding capacity rapidly, operating cash flow keeping pace with earnings matters.
It reduces the risk that growth is merely being booked on paper while working capital quietly becomes the real business model.
The AI server, high-density interconnect, and high-layer-count PCB opportunity appears genuine.
But I would still watch customer concentration, capacity additions, depreciation, and pricing carefully. A strong order book is useful. A strong order book at an inadequate return on capital is less useful.
The market tends to remember the first one and forget the second.
Semiconductors: Better, But Still Cyclical
華虹宏力 (01347) Hua Hong Semiconductor delivered a recovery that was hard to ignore.
First-half revenue reached USD 1.378 billion, up 24.5%. Gross profit rose 83.2%, while gross margin improved from 10.1% to 14.8%. Profit attributable to owners of the parent rose more than fourfold.
The drivers were clear: expanding 12-inch capacity, improved product mix, rising wafer shipments, and higher average selling prices.
That looks more substantial than a simple accounting rebound.
But foundry economics remain cyclical. New capacity can arrive. Customers can adjust inventory. Price competition can return. Margins are improving, but they are not yet close to historic peaks.
A good recovery is not the same as a permanently better industry.
I keep reminding myself of that because the market usually stops reminding me precisely when I need it most.
Optical Modules: The Market’s Favourite Exam
中際旭創（03308） may have produced one of the most powerful reports in the AI optical-communications supply chain.
Revenue rose 182.5% to RMB 41.78 billion. Net profit rose 241.7% to RMB 13.65 billion, driven by demand for 800G and 1.6T optical modules and higher cloud-capex spending from overseas customers.
Those are exceptional operating results.
The question is not whether the business is performing.
It is whether the valuation already assumes that it will perform exceptionally forever.
That is a much harder standard.
For companies like Zhongji, I would watch technological migration from 800G to 1.6T and eventually 3.2T, customer concentration, product leadership, gross margins, and the pace of hyperscaler capex.
The business can remain excellent while the stock still suffers if the market’s expectations were even more excellent.
That is one of investing’s less charming features.
Memory: Great Results Can Still Lose Money
兆易創新（03986）reported a spectacular first half: revenue up 178.7%, net profit up 1,091.5%, adjusted net profit up 796.9%, and operating cash flow up 531.5%.
Then the share price reportedly fell sharply after the result.
That is not irrational.
It is the market asking whether the memory-price cycle has already been priced in.
In memory, earnings can look strongest when supply discipline, pricing, and inventory dynamics all align. But the same forces can reverse when capacity expands or demand slows.
I would focus less on the headline profit percentage and more on adjusted profit quality, memory-price sustainability, inventory, and the durability of domestic-substitution demand.
A 1,091% profit increase is spectacular.
It is also a reminder that percentages can become theatrical when the starting base was small.
The Quietly Important Name
瀾起科技（06809） numbers were less explosive, but still strong.
First-half revenue grew 26.7%, gross profit grew 36.9%, and attributable profit rose 72.3%. Its interconnect-chip and server-platform businesses continued to benefit from AI-computing demand.
This is the sort of report I find useful because it shows operating leverage: profit growth outpacing revenue growth.
But the market’s reaction was cautious after the results, partly because the share price had already performed strongly and because investors were scrutinising dividend momentum and adjusted-profit trends.
That is the uncomfortable truth of a hot sector:
A company can report good news.
A stock can still fall.
The market does not only price what happened. It prices the gap between what happened and what had already been expected.
What This Earnings Season Says
Three themes stand out to me.
First, AI infrastructure is already producing visible earnings across the supply chain.
Optical modules, fibre networks, PCBs, electronic materials, memory interfaces, and semiconductor capacity are seeing direct demand from data centres and accelerated-computing investment.
Second, AI-model companies and AI-hardware companies are different investment questions.
Model companies such as MiniMax and Zhipu may be building important platforms, but they remain in a period of high R&D, high computing costs, and incomplete monetisation.
By contrast, many hardware suppliers are already reporting real revenue, real profit, and real cash flow.
Third, cyclical profits are not permanent profits.
When I see record margins in electronic materials, optical communications, foundry services, or memory, I do not dismiss the numbers.
I simply ask:
If industry demand falls by 30%, does this company still earn an acceptable return on capital?
That question is much less exciting than saying “AI is changing everything.”
It is also more likely to protect me when everything starts changing again.
Becoming Worthy Of Good Stocks
Charlie Munger once expressed an idea I keep returning to: good partners are rare, so the best way to find one is to become someone worthy of a good partner.
I apply a version of that to investing.
The best way for me to find a good stock is to become an investor worthy of holding one.
That does not mean predicting every quarterly result.
It means doing the quiet work:
Understanding what the company actually sells
Distinguishing revenue from cash flow
Looking at margins rather than headlines
Understanding the supply-and-demand cycle
Checking customer concentration
Tracking capital expenditure and depreciation
Asking whether growth increases per-share value
Remaining willing to revise my view when the facts change
Good companies are rare.
Good companies at sensible prices are rarer.
And investors who can hold them calmly through noise may be rarer still.
I do not need to have an answer to every market headline immediately.
Some questions require the next earnings report. Some valuations need sentiment to cool. Some investment theses need time to prove themselves.
For now, I will keep doing the work: reading, writing, recording, and updating my view.
Not because I expect to be right every time.
But because I would rather make a decision based on logic than allow the market’s mood to make it for me.
Views are personal and for discussion only. They do not constitute investment advice.
