I attended the China Q3 Update webinar by KraneShares and decided to share the highlights here, as it answers the most common questions about your China sleeve allocation.
An investor can own Chinese equities and have little exposure to China’s semiconductor industry. Move the allocation into mainland shares, and the gap may remain surprisingly large.
In its October 2026 investment presentation, KraneShares reported semiconductor exposure of just 2% in MSCI China, compared with 9% in both the CSI 300 and MSCI China A 50, and 76% in the STAR Market 50. These are not interchangeable ways of buying the same economic story. They are substantially different portfolios of businesses.
That distinction matters as investors search for opportunities across artificial intelligence. A country allocation tells us where securities belong in a portfolio’s geographical framework. It does not necessarily tell us who buys the companies’ products, what drives their earnings, or whether they are collecting revenue from the AI buildout or financing it.
Yet the answer is not simply to abandon China’s familiar internet platforms for mainland hardware companies. The platforms are changing too. Alibaba, Tencent and others are developing cloud, model and AI-service businesses alongside their established activities. KraneShares estimates that cloud computing and AI services together accounted for 18% of revenue across KWEB* holdings in the second quarter of 2026. That is a combined category, not a measure of AI revenue alone, but it challenges the idea that China’s established platforms belong entirely to yesterday’s economy.
The more useful distinction is between two AI trades: companies supplying the infrastructure, and companies spending to deploy and monetise it.
Their opportunities connect. Their economics differ.
The China inside the benchmark
The starting point is the composition of conventional China exposure.
KraneShares’ presentation, using MSCI data as of September 30, 2026, shows mainland A-shares accounting for 17% of MSCI China, Hong Kong-listed Chinese companies for 80%, and US-listed Chinese companies for 3%. In MSCI China All Shares, the corresponding weights are 52%, 46% and 2%.
Neither composition is a universal prescription for a portfolio. But the comparison demonstrates how strongly benchmark selection shapes the version of China an investor owns.
Benchmark composition by listing venue, as presented in the October 2026 KraneShares webinar deck. (Utilization authorized by KraneShares)
The more revealing comparison comes next. Even the broad mainland indices shown in the presentation have semiconductor weights of only 9%, far below the 76% reported for STAR 50. Moving onshore therefore does not automatically deliver a concentrated allocation to the semiconductor opportunity.
There are three separate decisions here: how much China to own, how much exposure to take through mainland or offshore listings, and which businesses and earnings drivers to own within those markets. The first two do not resolve the third.
A geographical allocation is not the same thing as an economic exposure.
An investor seeking broad Chinese equity participation may reasonably want a diversified country benchmark. An investor seeking semiconductor equipment, memory, chip design, or AI infrastructure is asking a different question. Confusing those objectives can leave a portfolio looking correct on a country-allocation report while failing to express its intended investment thesis.
China’s first AI trade: supplying the buildout
The first trade concerns the physical infrastructure behind AI.
KraneShares places China within an Asian supply chain that includes Taiwanese foundry and packaging capabilities, Korean memory, Japanese semiconductor equipment, and Chinese optical networking and components. Its presentation identifies China as home to three of the four largest global optical-transceiver suppliers, citing LightCounting research from May 2026.
This is a useful corrective to an AI narrative dominated by processors and large language models. The investment opportunity extends to the equipment, materials, and connections required to make the infrastructure work.
The STAR-market portfolio discussed in the presentation illustrates one concentrated expression of that opportunity. Its reported sub-industry breakdown includes 57% semiconductors and 24% semiconductor materials and equipment as of September 30, 2026. Those are portfolio classifications and should not be confused with the separate index-level semiconductor comparison.
The businesses represented include chip designers, equipment manufacturers, storage companies, and component suppliers. Their individual exposures differ, but the investment questions have a common structure.
Who is placing the orders? How durable is the demand? How much pricing power does the supplier retain? What investment is required to expand capacity? And how much revenue ultimately becomes cash available to shareholders?
Those questions matter because growth in AI infrastructure spending does not distribute profits evenly across the supply chain.
The presentation highlights strong first-half revenue growth among KSTR’s top 15 holdings. That establishes a growth story among the selected businesses. It does not, on its own, establish an earnings story, a free-cash-flow story, or an attractive valuation.
A supplier can experience rapidly rising sales while committing substantial resources to research, production, or working capital. To turn an industrial opportunity into an investment thesis, investors must examine the economics between the order book and the shareholder.
They must also distinguish different demand sources. A company benefiting from domestic substitution is not necessarily exposed to the same cycle as one supplying international data-centre investment. Both may be labelled Chinese technology. Their customers, policy sensitivities and competitive environments may differ materially.
China’s second AI trade: monetising the platforms
The second trade begins with businesses investors already recognise.
China’s internet platforms have established positions in commerce, entertainment, advertising and digital services. The AI proposition is that these existing relationships can become channels for distributing and monetising new capabilities.
KraneShares’ presentation describes several potential layers of monetisation: computing capacity, model and API usage, enterprise software, commerce and advertising. It highlights Alibaba’s ability to connect infrastructure, models and applications, and Tencent’s combination of messaging, merchants and payments as a potential foundation for consumer AI agents. These are investment arguments about how the ecosystems could develop, not evidence that every proposed revenue stream has already matured.
The distinction from the infrastructure trade is important. A supplier seeks to earn money by delivering something the buildout requires. A platform may first need to spend money creating the infrastructure and services from which it hopes to earn future revenue.
KraneShares estimates that revenue attributable to cloud computing and AI services across KWEB companies grew approximately 51% year-on-year in the second quarter of 2026. The presentation also estimates that the combined category represented 18% of those holdings’ revenue.
Those figures suggest that newer activities are becoming meaningful. But they require careful interpretation.
First, cloud computing is not synonymous with AI. Combining the categories does not justify describing all the revenue as AI-derived.
Second, the portfolio itself is changing. The presentation notes August additions including Lenovo, GDS Holdings and Kingsoft Cloud, alongside other companies. An internet-labelled portfolio can become more AI-oriented because its existing businesses evolve, because its constituents change, or both.
Third, revenue growth does not settle the return-on-investment question. Investors still need to assess infrastructure costs, operating expenses, utilisation and the durability of customer spending.
The platforms’ existing businesses should therefore not be dismissed as obsolete baggage. Their cash generation, customer relationships and distribution may be part of what makes AI expansion possible. But those advantages must eventually translate into measurable economics.
The capex distinction that changes the argument
One of the presentation’s most consequential slides requires particular care.
Its headline suggests that internet capital expenditure may have peaked. The underlying chart, however, is labelled “China Internet Quarterly Capex Growth” and measures percentage growth rather than absolute spending.
A peak in the growth rate is not a peak in expenditure.
If spending continues to grow at a slower positive rate, the absolute amount still increases. Nor does slower spending growth, by itself, establish that free cash flow or profitability will improve.
This is more than a chart-reading technicality. It goes to the relationship between the two AI trades.
Infrastructure suppliers may benefit while platforms continue expanding investment. Platforms, meanwhile, must demonstrate that the resulting capacity generates sufficient revenue and returns. The same spending programme can support one company’s sales while placing demands on another company’s cash resources.
That does not automatically favour the suppliers. Their valuations may already anticipate strong demand, while capacity expansion or competition may erode the benefits. Equally, a platform carrying a heavy investment burden today could develop valuable recurring revenues later.
The investment task is to identify where each company sits in that sequence, not to assume that every participant benefits at the same time.
An exposure gap is not a valuation gap.
The benchmark argument becomes dangerous when it turns into an automatic buy signal.
A business can be underrepresented in a widely followed index and still be expensive. A sector can be strategically important and deliver disappointing shareholder returns. A correct long-term industrial thesis can coexist with a difficult entry price.
The deck itself supplies a useful reminder. Its performance appendix reports a three-month NAV return of −30.02% for the US-listed KSTR fund to September 30, 2026, alongside a one-year NAV return of +11.95%. These are returns over different periods, not a peak-to-trough drawdown calculation. Nevertheless, they illustrate how sharply the investment experience can vary even within a compelling technology theme.
The offshore platform thesis also needs separation into its component parts. The presentation shows consensus estimates for profit recoveries at companies including Meituan, JD.com, Alibaba, and PDD Holdings. Those are forecasts, not completed recoveries.
An improvement in earnings could reflect recovering established operations, reduced competitive pressure, an easier comparison period, AI-related growth, or some combination. Those explanations have different implications for durability and valuation.
“AI rerating” is too broad a description if investors cannot identify what is actually improving.
Buying a business model, not a label
The question is not whether investors should replace “old China” with “new China.” That division is too neat.
Broad mainland exposure does not necessarily deliver concentrated semiconductor exposure. An offshore internet portfolio does not necessarily exclude AI infrastructure or emerging AI services. And a familiar platform may contain both a mature cash-generating business and an investment-intensive technology expansion.
A more useful assessment starts with four questions:
1. Who is the end customer?
2. Is the company supplying the AI buildout, financing it, or doing both?
3. How does growth translate into sustainable cash generation?
4. What expectations are already reflected in the price?
Only then should investors ask whether their benchmark expresses the intended exposure.
The risk of buying yesterday’s China is not simply owning yesterday’s companies. It is using yesterday’s categories to analyse businesses whose customers, revenue mix and investment requirements are changing.
