As large language models mature, the commercial centre of gravity in the artificial‑intelligence industry is shifting from foundational research to application‑level monetisation. Enterprise software that embeds AI—particularly productivity and collaboration tools—has emerged as the first place where AI produces sustained, measurable revenue. Traditional licensing and seat‑based subscriptions are giving way to hybrid models: a base subscription plus metered “token” charges for model inference, or separately priced AI‑only modules. That change is already unlocking new revenue ceilings for software vendors because usage scales with business activity rather than with headcount alone.
Market research and securities analysts have translated this usage shift into bold forecasts. Gartner estimates that by the end of 2026 some 40% of enterprise applications will integrate task‑specific AI agents, up from under 5% in 2025, and predicts that “agent‑style” AI could contribute as much as $450 billion of enterprise software revenue by 2035. Chinese brokerage notes and banks are echoing the same logic: when model inference becomes a production input, model providers can monetise scarcity through layered pricing and subscriptionised products, while cloud operators increasingly sell raw compute and model firms sell token fuel plus outcomes.
The practical effect is visible in China’s capital markets: valuation frameworks for software and algorithm companies are recalibrating around high gross margins and recurring, usage‑driven cash flow. Software firms that externalise most inference costs as API fees but retain subscription pricing for outcomes can generate durable margins and predictable cash flows, an attractive profile for investors. Domestic analysts argue the country has moved from capability catch‑up to a demand explosion, with Chinese model vendors poised to replicate North America’s “sell‑token” moment as AI permeates engineering and operations rather than merely producing code snippets.
That narrative is reflected in passive investment products that emphasise the compute layer as the long‑term winner. The Tianhong CSI Artificial Intelligence Theme Index Fund (A: 011839; C: 011840) tracks an index skewed toward hard‑technology infrastructure, concentrating weight in semiconductors, communications equipment and software development. As of 13 February 2026 the index’s top sector weights were semiconductors (30.3%), communications equipment (21.0%) and software development (14.7%). The fund’s top holdings map onto the AI stack: optical modules and communications suppliers (19.76%), AI chips and processors (18.77%) and AI servers and systems providers (4.53%).
For investors, the fund offers two share classes aimed at distinct horizons. The A class carries an ongoing operation fee (management plus custody) of 0.6% and is pitched at investors willing to hold for more than a year to capture long‑run compute‑infrastructure dividends. The C class waives subscription fees but levies a daily sales service charge equivalent to about 0.25% annually, making it more suitable for tactical, sub‑year exposures. The product is widely accessible through retail finance platforms and supports systematic investing, which the issuer positions as a convenient way for ordinary investors to gain AI exposure without direct equity selection.
Performance history for the fund shows a volatile but powerful rebound: after steep losses in 2022 (roughly ‑33%), the fund posted gains of around 13% in 2023, 20% in 2024 and an outsized c.65% in 2025. Those past returns underline two points: AI‑linked strategies can amplify returns when industry adoption accelerates, but they are also cyclical and sensitive to technology sentiment. The fund and its index come with standard caveats—tracking error, no guarantee of future returns and the same sector concentration risks that typify thematic products.
The commercial shift to token‑based billing also carries broader market consequences. Token demand that scales with usage can create a durable revenue stream for model vendors, but it also alters bargaining between clouds, model providers and enterprise customers. Clouds may commoditise raw GPU hours while model firms assert pricing power over inference volume and verticalised outcomes. The result will be a new ecosystem of pricing, bundling and vertical partnerships as incumbents and challengers jockey for control of the token value chain.
Regulatory and operational risks complicate the upside. For Chinese players, export controls, cross‑border data governance and geopolitical friction over semiconductor supply chains remain potent constraints. Domestically, rapid price compression in compute or aggressive token pricing could provoke pushback from large enterprise buyers, slowing adoption. Investors should therefore weigh the structural opportunity against execution risks, policy uncertainty and the concentration that comes with thematic indexes.
In short, the metered‑token model is not just a billing tweak: it reframes which companies capture value as AI moves into enterprise workflows. For long‑term investors looking to back the compute base of generative AI, funds such as Tianhong’s CSI Artificial Intelligence Theme provide a pragmatic exposure to the infrastructure winners. But they are a concentrated play on the transition from application hype to capital‑intensive, high‑margin compute and chip ecosystems, and merits careful position sizing within a diversified portfolio.