China’s new AI application index has 30 companies—and no AI revenue test

The Shanghai Stock Exchange building in Pudong, photographed in 2022. The exchange and China Securities Index launched the new STAR AI Application index in September 2026; the photograph shows the venue, not a constituent or trading session. 钉钉 / Wikimedia Commons, CC BY-SA 4.0; resized.[8]

On September 2, the Shanghai Stock Exchange and China Securities Index launched code 950398, the SSE STAR Artificial Intelligence Application Premium Selected Index. The name sounds like a narrow bet on software. The actual basket is more revealing: 30 STAR Market companies drawn from the methodology’s three eligibility categories—model development, intelligent terminals and components, and application software and technical services.[1][2][3]

That is a meaningful change in emphasis. The established STAR AI index’s August 31 factsheet showed it was 91.9% information technology by weight. Its four largest constituents—Hygon Information Technology, VeriSilicon, Montage Technology, and Cambricon—accounted for 39.45% of the index, making the benchmark heavily dependent on four large information-technology names.[5] None of those four appears in the new application roster.[3]

But an application label is not the same thing as application revenue. The new index methodology asks whether a company is involved in an eligible AI business, then lets market size determine which qualifying names survive. It states no minimum share of sales from AI, no AI-profit threshold, and no required adoption metric.[2] The index is therefore best read as a map of where public-market investors can express China’s shift toward AI use—not as proof that 30 clean, comparable AI application businesses already exist.

The benchmark moves one rung down the stack

The older 950180 index, published in July 2024, was designed to cover companies providing AI’s “basic resources, technology, and application support.” Its August 31, 2026 factsheet shows what that broad mandate became in market terms: the top ten names carried 69.08% of the weight, with its four largest information-technology names ahead of Kingsoft Office in fifth place.[5]

The new 950398 index tilts further downstream. Its September 2026 methodology admits three kinds of business: developers of foundation or industry-specific models; makers of intelligent terminals and components that deploy, adapt, compress, or integrate those models; and developers or technical-service providers applying models to particular business scenarios.[2]

The first constituent list makes that wider meaning of “application” concrete. Kingsoft Office, UCloud, Roborock, Orbbec, EZVIZ, Arashi Vision, China Post Technology, and Haitian Ruisheng all appear in one basket.[3] At the September 11 snapshot, official live data divided the whole index among information technology, industrials, consumer discretionary, and communication services.[7] These names do not share one revenue engine. What joins them is their position at an interface where a model can meet a worker, document, camera, home, warehouse, or industrial process.

This breadth echoes the language of China’s national “AI Plus” policy more closely than the older index’s 91.9% information-technology mix does.[4][5] The State Council’s August 2025 opinion set a goal for adoption of next-generation intelligent terminals and agents to exceed 70% by 2027 and 90% by 2030. It named AI phones and computers, connected vehicles, robots, smart homes, wearables, industrial software, service-sector agents, and public services as deployment surfaces.[4] The new index does not measure progress toward either percentage. It does, however, echo the policy’s application-first emphasis in a capital-market perimeter.

Four screens, then a size ranking

The selection process is more mechanical than the word “premium” may suggest.

First, a security normally needs more than six months of STAR Market trading history. A company ranked among the market’s five largest by average daily total capitalization since listing may enter once it has traded for more than three months. Securities carrying a delisting-risk warning are excluded.[2]

Second, liquidity matters. A candidate’s average daily trading value over the preceding year must rank within the top 90% of the eligible universe. Third, companies with a China Securities Index ESG rating of C or below are removed. Only then does the thematic screen ask whether the business touches one of the three model, terminal, or application categories.[2]

The final selector is not AI revenue, contract growth, user retention, or research quality. It is average daily total market capitalization over the past year. The 30 largest remaining companies enter. Once selected, securities are weighted by free-float-adjusted market value. At a regular adjustment, weight factors are set so no constituent exceeds 10%; that is a reset rule, not a continuous ceiling between adjustments. Constituents are reviewed twice a year, with regular changes taking effect after the second Friday of June and December.[2][7]

This construction has sensible index logic. Seasoning and liquidity make a basket easier to replicate. The ESG floor and delisting screen remove two explicit risk classes. The 10% reset restrains initial domination by any one company. Semiannual review reduces scheduled review frequency—and potential turnover—relative to the older broad index’s quarterly schedule.[2][5]

Yet none of those controls makes the basket purer as an AI earnings stream. A large diversified device maker can outrank a smaller software company whose revenue is more directly tied to model use. A fast-growing AI feature can qualify a company even when its economics remain embedded in a much larger product line. Market capitalization then amplifies what investors already value; it does not independently verify why they value it.

The missing denominator is financial, not technical

Shanghai’s exchange has already published striking evidence of ecosystem scale. In an April review of 2025 annual reports, it said more than 30 STAR Market AI companies collectively produced RMB179.876 billion of revenue, up 25.7%, and RMB13.38 billion of net profit, up 200.8%. The same review said WPS AI services’ average daily token-call volume exceeded 200 billion, which it described as more than twelvefold year-over-year growth.[6]

Those numbers matter, but they should not be pasted onto the new index. The exchange’s review does not identify its “more than 30” companies as the 30 constituents later chosen for 950398, nor does it publish an AI-only revenue bridge for the basket.[3][6] The profit total describes a broader exchange narrative. The WPS token figure describes usage. Neither tells an investor using the index what fraction of constituent sales, gross profit, or cash flow came from AI applications.

This distinction is particularly important at the application layer. Compute suppliers can often point to chips, systems, or capacity sold into a buildout, even if end demand remains uncertain. Application companies face another set of questions: Was AI bundled into an existing subscription or sold separately? Did a robot or camera command a higher price because of model features? Did inference cost rise faster than subscription revenue? Did a pilot become a recurring contract? Did usage deepen retention, or merely create a spectacular token bill?

The methodology does not pretend to answer those questions. It supplies an investable classification, not a common accounting standard.[1][2] That boundary is a strength when it is visible and a risk when the index name is mistaken for due diligence.

Even the history needs a boundary

The new index uses December 30, 2022 as its base date and 1,000 as its base value, although it was not formally launched until September 2026.[1][2] Any series shown before launch is therefore rule-based historical index data, not the realized return of a product that investors could actually have owned under this name. It may help compare market regimes, but it cannot reproduce live fund fees, tracking error, liquidity effects, or the knowledge investors had at the time.

The existing broad index illustrates why that warning matters. As of August 31, it reported a rolling price-to-earnings ratio of 100.61, one-year annualized volatility of 42.44%, and a one-year price return of −5.32%, despite a three-year annualized return of 22.70%.[5] Those figures belong to 950180’s information-technology-heavy composition, not to the new application basket. They show how quickly an “AI index” label can conceal a particular concentration, valuation, and starting point.

China Securities Index’s live data already provides the first post-launch composition check. On September 11, Kingsoft Office carried 12.65% and Roborock 10.67%; together the top five made up 41.97%, and the top ten 64.86%. The sector split was 67.60% information technology, 20.17% industrials, 10.67% consumer discretionary, and 1.56% communication services.[7] That is less technology-heavy than 950180’s 91.9% and slightly less concentrated than its 69.08% top ten, but it is not diffuse. It also demonstrates why the methodology’s 10% language must be read at the adjustment point: market movement can carry live weights above it between resets.[2][5][7]

What would prove the application turn

The index is an early market artifact of a real policy shift. China’s public AI narrative is moving from “can domestic infrastructure train and serve models?” toward “which products and workflows can absorb them?” The omission of the broad benchmark’s four largest information-technology names makes that turn visible.[3][5]

The harder evidence will arrive in company disclosures, not in the index title. Four signals would make the application thesis financially legible:

An AI revenue bridge. Constituents separate AI-native sales, AI-assisted sales, and ordinary legacy revenue instead of grouping them under a strategic narrative.
Unit economics after inference. Companies disclose whether higher model use improves gross profit, renewal, device pricing, or service attachment after compute costs.
Repeated deployment. Paid seats, active devices, recurring contracts, retention, and renewal replace demonstrations and aggregate token calls as the evidence of demand.
Rebalance evidence. At each June and December review, additions and removals come with enough disclosure to show whether their AI business is economically material. The current method requires qualifying involvement, not that materiality test.[2]

If those measures deepen, 950398 could become a useful gauge of China’s passage from AI capacity to AI cash flow. If they do not, the index will still be useful—but as a gauge of market appetite for the application story, not the story’s commercial completion.

The September launch is therefore not evidence that China has solved AI monetization. It is evidence that Shanghai’s index providers now offer a separate place for the market to price the attempt.

cronfeed.work