MSCI split the AI supply chain into three parts. The arithmetic underneath it refuses to split so cleanly.
According to reporting on the launch, the index provider carved artificial intelligence into physical infrastructure — chips, data centers, power and cooling — digital infrastructure like cloud and model layers, and an application tier of enterprise software and consumer products. Alongside the product announcement sits a Bain & Company estimate: reaching AI's promised scale by 2031 would require $6 trillion in annual revenue, while existing applications generate roughly $1.2 trillion.
Read those two facts together and something strange happens. The launch is described as a response to investors wanting "more specific exposure." But the number attached to it — the gap between what AI must earn and what it currently earns — is closer to $4.8 trillion. The reporting calls it "hundreds of billions." That is off by roughly a factor of ten.
I have spent enough time auditing token distribution models to know what a rounding error of that size usually means. It means someone is managing a narrative, not a spreadsheet.
Context: why an index is never just an index
For the non-specialist, an index sounds like plumbing — a neutral ruler for measuring a market. It is not. MSCI's revenue comes from licensing fees. The company is paid when asset managers build ETFs and passive products that track its benchmarks. It earns whether AI rises or falls. This is the shovel-seller's business, and it is a very good one: index licensing is high-margin, capital-light, and scales almost infinitely.
So when a firm whose income depends on how many ways a theme can be sliced introduces a new set of slices, that is not a forecast. It is product design. The company's own index head, Jana Haines, framed the launch around risk management — investors seeking "more specific exposure" they can then hedge. The word "hedge" is doing quiet work here. You only need a hedging tool once you are already worried.
The three-segment classification itself is not arbitrary. Physical infrastructure, digital infrastructure, application layer — this mirrors the technology stack and repeats a pattern we watched in the internet cycle, where value capture migrated from telecom equipment to cloud platforms to application software. That migration is real. What is unsettled is where AI currently sits on that curve, and whether the downstream end will ever arrive.
Core: the methodology problem and the migration problem
Here is where my applied-mathematics background makes me uncomfortable. A financial index is a function. It takes a universe of companies, applies a weighting rule, rebalances on a schedule, and outputs a return series. Everything that matters lives inside those parameters — market-cap weighting versus equal weighting, how often you rebalance, whether you add momentum or quality factors. None of that has been disclosed for this series. An undisclosed weighting rule is not a neutral omission; it is a withheld opinion.
From my 2017 work auditing early ERC-20 distributions, I learned that the weighting rule is the ethics. A token that quietly weights toward early whales is a different instrument than one that does not, even if both claim "fair launch." Code is law, but people are purpose — and the people who set the parameters are making a value judgment, not a measurement. The same holds for an AI index. If the physical infrastructure segment is cap-weighted, it concentrates into a handful of semiconductor giants, and the "sub-division" becomes a rebranded concentrated bet. If it is equal-weighted, it drags in thin, illiquid names. The methodology decides whether this is a tool or a marketing device. Until it is published, we cannot tell.
Then there is the gap itself. Bain's $6 trillion is not a revenue forecast. It is a break-even figure — the revenue required to justify the data centers, chips, and power already being built. This is a supply-side back-calculation of demand, and history says such estimates run conservative. In 2000, the fiber build-out assumed demand that arrived years later than projected, and the capital sat idle while it waited. The difference now is speed: cloud providers are committing on the order of $200 billion a year in AI capital expenditure, and that number is still climbing.
The three-tier structure exposes the tension. Physical infrastructure is where the money is actually being made today — that is why it earns its own segment. Digital infrastructure is capital-heavy and unevenly profitable. The application layer is where the revenue must eventually land, and it is precisely where revenue has not arrived. The gap is not a market inefficiency. It is the shape of the whole cycle.
I have seen this movie in DeFi. In 2020, yield farms advertised triple-digit APYs that depended on future emissions, not present cash flow. The model was solvent only as long as new capital kept entering. AI's infrastructure build has a similar reflexivity, just at institutional scale — the returns on compute depend on applications that the compute is supposed to enable. The interest rate curves in Aave and Compound are set by governance parameters rather than live market supply and demand, and the resulting rates often tell you more about the parameter-setters' assumptions than about capital's real cost. Index weighting rules carry the same fingerprint: a human assumption wearing the costume of a market signal.
Blockchain has run this experiment already, at smaller scale. Decentralized protocols bundled their own supply chains into tokens, then into indices, then into structured products. The pattern repeated: the further a product sat from actual cash flow, the more its price depended on narrative. Most DAOs that issued governance tokens learned, painfully, that they had no legal status as entities — when something broke, members discovered they held unlimited personal liability and no tool to hedge it. Community is the new central bank, but only when the community actually controls the reserves. An index cannot manufacture that control.
Contrarian: subdivision is not diversification
The instinct reading this news is that granularity equals safety — that by breaking AI into pieces, investors reduce their risk. The opposite is more likely true. Resilience beats hype every time, and subdivision is not diversification. The more precise the knife, the more it matters where you cut.
Subdivision lets you bet on a single link in the chain. If the application layer never monetizes, an investor who "precisely" allocated there loses more than someone holding a broad AI basket. Granularity sharpens exposure; it does not dilute it. Passive funds tracking these sub-indices will flow into whichever segment carries the best story that quarter, and "index inclusion to inflows to higher valuation to further inclusion" is a feedback loop with no fundamental floor. We should expect ETF issuers to race in, because that is the licensing model's whole point.
And notice who cannot participate in the hedging half of the story. The reporting concedes that index-based hedging and speculation "typically aren't appropriate" for ordinary investors, and it never resolves how a retirement account is supposed to hedge AI exposure at all. So the structure is: institutions get the tools to move within AI, retail holds the concentrated long position and absorbs the residual risk after the sophisticated players have hedged. That is the same asymmetry I watched during the 2022 governance crisis, when core contributors and ordinary users were asked to bear identical downside with wildly different toolkits.
Takeaway
An index is a claim about how the world fits together, and whoever draws the lines holds quiet power over where capital flows. MSCI now draws the lines for AI's supply chain. The question worth carrying forward is not whether the segmentation is clever — it plainly is — but whether the $4.8 trillion it quietly maps will ever be earned, or whether we are simply building a more precise instrument for pricing a demand that has not yet arrived.
Trust, verify. But also, connect. And when a number is off by a factor of ten, verify again.