The Hang Seng AI Infrastructure Index Is a Rules-Based Product With a Discretionary Heart"

"article": "I. Hook\n\nOn 31 July 2025, Hang Seng Indexes Company published the methodology for the Hang Seng Stock Connect AI Infrastructure Index. The announcement was clean. The structure was orderly. There were no obvious typos in the classification table, and no immediate sign of a bug. That was the anomaly. A product that presents itself as rules-based has declined to name the one rule that actually matters. In my trade, I would call this an oracle. In my work as a DeFi security auditor, the first question I ask about any protocol is: where does the truth come from? For this index, the truth comes from a committee. The committee is the smart contract of this benchmark. Trust is not a variable you can optimize away; it is an unresolved dependency.\n\nII. Context\n\nLet's establish what the index actually is. It tracks Hong Kong-listed stocks that are eligible under the Stock Connect scheme and that satisfy certain parameters related to AI infrastructure. The infrastructure theme is the most literal picks-and-shovels construct ever worn by a conventional benchmark. It does not target AI applications. It targets semiconductor manufacturers, chip equipment vendors, server and storage builders, data-center operators, cloud service providers, network hardware makers, power and cooling suppliers, and the software plumbing that makes large-scale AI possible. The index gives mainland investors, through southbound Stock Connect, a packaged route into the AI buildout from inside the Hong Kong exchange. For global allocators, it is a theme with an official seal.\n\nThe launch date is not neutral. AI infrastructure has become a geopolitical category. Export controls have forced a rethink of the semiconductor supply chain, and Hong Kong is the natural venue for international capital that wants China-linked technology exposure. An official Hang Seng benchmark makes that trade respectable. It is a capital-markets instrument for a policy-driven industrial chain. When the index is quoted on Bloomberg or Reuters, it will become the shorthand for that chain. In a few years, the question will be whether the chain performed the way the index design expected.\n\nIt is a sibling of the Hang Seng Tech Index, but with a completely different temperament. The tech index is wide; this index is deep and cyclical. It tries to isolate the part of the technology stack that currently receives the largest capital expenditure flows. It does not promise innovation. It promises infrastructure. The difference is not just semantic. Infrastructure means ownership of heavy assets, long payback periods, and brutal exposure to the cycle. The index is making a statement that the economic bottleneck of the AI era is not the model. It is the physical world that runs the model. That is a defensible statement. But it is not a diversified one.\n\nIII. Core: The Methodology as Code Review\n\nNow I want to treat the index construction process the way I would treat a Solidity contract. There are three layers in every rule set: selection logic, weighting scheme, and rebalancing schedule. In a smart contract, assumptions are explicit. In an index methodology, they are buried in natural language.\n\nIn 2017, I spent forty hours tracing the Golem network's multi-sig contracts. The gap that mattered was not in the Solidity. It was in the unstated assumption that the multi-sig signers would behave honestly. I published a technical rebuttal that annoyed people who thought 'code is law' was a complete sentence. The lesson followed me into every audit since. The most important code is the code that describes who decides. For this index, that code is the name 'Hang Seng Indexes Company' at the bottom of the methodology. The name is a signature, and a signature is a trust assumption. Trust is not a variable you can optimize away; it is moved into the identity of the signer.\n\nThe selection logic depends on a classification decision. To classify a company as AI infrastructure, someone must decide that its business activity is sufficiently aligned with AI infrastructure. That decision relies on revenue segments. Revenue segments are self-reported by the company, filtered by a data vendor, and then passed to a committee. The final label is not an equation. It is a judgment. In DeFi, I warn every team against relying on a single oracle when the source of truth can be challenged. Here, the source of truth is a committee's interpretation of a fast-moving industry. There is no on-chain proof. There is no challenge window. There is only a document that says the classification may be revised at any time.\n\nThe weighting scheme creates a second problem. If the index is free-float market-capitalization weighted, the largest names will dominate. That is normal for global benchmarks, but AI infrastructure is a sector where the largest names are also the most correlated. Huge semiconductor companies, server integrators, and data-center operators all depend on the same two variables: hyperscaler capex budgets and access to cutting-edge chips. When an index has no factor diversification, market-cap weighting becomes a lever for concentration risk. A cap on the largest constituent can solve single-name dominance, but it introduces turnover, and turnover in a Stock Connect product is not free. It includes stamp duty, FX fees, and the bid-ask spread of waiting for liquidity. Every rule that protects against concentration is paid for with tracking error.\n\nThe rebalancing schedule is the third hidden layer. A typical thematic index reviews constituents every six months. AI infrastructure cycles do not respect semesters. After 2023, the meaning of AI infrastructure changed at least three times. First, memory became a bottleneck, so memory companies joined the theme. Then, power became an investment category because data centers were eating the grid. Then, cooling became a legitimate sub-sector because thermal density changed the physics of the data center. A semi-annual committee cannot keep up with structural repricing. It will always be revising last quarter's definition at next quarter's prices. That lag is not a bug in the index. It is the fee an investor pays for the comfort of an official label.\n\nA buffer rule may reduce turnover by retaining existing constituents unless their rank falls below a threshold. That rule will make the index a little slower, which is a feature. But the same buffer that protects against unnecessary turnover also protects stale classifications. If a company has transformed from AI infrastructure to AI services, and the buffer keeps it in because it has not fallen far enough, the index will hold a ghost. Staleness is the hidden tax on every rules-based product.\n\nThis is also where the backtest fails. The index provider will eventually publish historical numbers, but any backtest of AI infrastructure is an exercise in look-ahead bias. The classification did not exist in 2019. The companies that would qualify today were not always classified as AI infrastructure. Applying today's labels to yesterday's prices will always make the backtest look cleaner than the actual investment experience. I have audited enough algorithmic services to know that a backtest is a narrative before it is a statistic.\n\nIn a bear market, backtests are twice as dangerous. The technology drawdown of 2022 is already in the data. The AI-capital-expenditure drawdown of the next cycle is not. Every index has the same curve: a clean equity line stretching upward, and a footnote that says past performance is not indicative. But in a thematic index, the past performance is not even the same product. It is the same ticker with a different definition.\n\nI have seen this exact design before. In 2021, I audited a decentralized index token. The code was clean. The rebalancing function was sharp. The vulnerability did not live in the function. It lived in the list. The token's governance could swap the entire constituent set without the holder being able to interrogate the decision before it happened. The Hang Seng AI Infrastructure Index has the same shape. The methodology is the code. The advisory committee is the governance. The constituent list is an input, not a source of risk. But in any rules-based system, the list is the smart contract.\n\nThere is another hidden input: correlation. A well-constructed index should provide exposure to uncorrelated betas. AI infrastructure does not. Under the hood, it is a collection of companies whose earnings are driven by the same macro factors: global AI capex, China's semiconductor policy, US export controls, electricity prices, and discount rates. When those factors move, the constituents move together. The blue-chip AI infrastructure company, the memory supplier, and the cooling company fall in the same week. A portfolio that looks like 40 lines on a screen is really a portfolio of two or three risk factors. That is not a flaw in the index rules. It is a flaw in the mental model of diversification.\n\nI need to add a personal note, because this is where my work changes how I read index products. In 2026, I helped build an AI-driven oracle for a decentralized prediction market in Manila. One lesson we learned is that confidence scores are not proof. You can commit to a model cryptographically, but you cannot prove that the model's judgment is honest. The same is true for an index. Hang Seng can commit to a formula. It cannot commit to the speed or the honesty of its taxonomy. That is not an accusation. It is a structural property of the system.\n\nIV. Contrarian: The Shovel Is a Lever\n\nThe conventional argument for this index is that picks-and-shovels is a safer way to play AI than betting on a single application winner. I have spent enough time in commodity and crypto cycles to distrust the shovel. In a gold rush, selling shovels is better than gold mining for a while. Once the rush becomes crowded, the price of shovels rises, capacity expands, and margins collapse. The advantage of the picks-and-shovels approach is a timing advantage. When gold falls, shovel makers do not stop producing. They carry fixed assets, heavy inventories, and expensive debt. The same pattern appears in crypto infrastructure. I audited projects in 2020 that called themselves the infrastructure of DeFi. They minted money in the bull market. In the bear market, they traded like the protocols they served. Infrastructure is not a defensive property. It is a transmission mechanism. When the primary buyer of AI infrastructure cuts capex, the infrastructure seller feels the impact faster and deeper than the AI application provider, because the application provider can cut headcount while the infrastructure provider must carry fixed costs.\n\nThe ETF wave that followed the 2024 US approvals amplified this phenomenon. Institutions asked for a simple expression of the AI view, and index providers responded with taxonomies. The Hang Seng index is part of that wave. But institutional demand does not remove cycle risk. It concentrates it. When the same institutions de-risk simultaneously, the underlying constituents will move as one.\n\nThis is where the bear market lens matters. The index is designed to be bought for years. But the current AI buildout is a short-cycle boom in a long-cycle industry. The index abstracts the cycle away by maintaining a stable label. The label remains stable while the companies underneath stop fitting the category. The index is not a strategy. It is a product wrapper. It invites the investor to forget that classification is a human judgment. The committee may be right for the next three years. That is different from being right forever.\n\nA sensible thesis for this index would require three beliefs: that AI capex is not near the peak; that the index constituents can convert capex into cash flow; and that the committee's classification will not distort the basket. Each of those beliefs is an active bet. The passive label only hides them. There is nothing prudent about owning the most capital-intensive portion of the AI supply chain.\n\nThe most contrarian thing I can say is that the biggest risk is not the stocks in the index. It is the investors who treat the index as a passive default. When the AI capex narrative cracks, an active manager can exit in a day. An index-based investor must wait for a rebalance