The Missing Fields: What a Record AI Equity Underweight Actually Verifies

Hook

Three facts. That is the complete information content of the item: US equity mutual funds are underweight AI equities; the underweight is the largest on record; hedge funds are positioned on the other side. No magnitude in basis points. No sample count. No benchmark definition. No measurement date. No originating institution.

An auditor's first move when handed a claim with eleven empty fields is not to dispute it. It is to inventory what the claim can support. "Underweight AI" without a declared benchmark is not a measurement. It is a direction. Directions are cheap. They cost nothing to publish and nothing to falsify.

So I did what I do with a half-specified contract: I reconstructed the constraint set that would mechanically produce the observation, then asked whether the same constraint set exists in the systems I actually audit — on-chain allocation vaults, restaking layers, and the AI-agent execution rails that have absorbed more institutional flow than any of them publicly report. It does. Not as analogy. As identical arithmetic.

The number that matters is not the underweight. It is the tracking error budget that generated it — and the same budget now silently governs three classes of on-chain allocation system that have never been stress-tested under redemption pressure.

Context

Start with the legal container, because the container determines the output.

A US equity mutual fund registered under the Investment Company Act of 1940 does not allocate capital the way a family office does. It operates inside a diversification test, a liquidity classification regime, a daily redeemability obligation, and a prospectus that a compliance officer will read literally. Rule 22e-4 requires every position to be classified into one of four liquidity buckets. The 75/5/10 diversification test caps concentrated positions in a defined portion of the portfolio. The prospectus states a strategy. The strategy is legally binding on the portfolio manager in a way that a hedge fund's investment mandate is not.

None of that is commentary. It is the operating constraint set, and it produces allocations that are frequently indistinguishable from views while being nothing of the kind.

Now add the benchmark layer. The overwhelming majority of large US equity mutual funds are benchmarked to a market-capitalization-weighted index. The index determines what "neutral" means. A fund's weight in any sector is measured against that index, not against the economy, not against the opportunity set, and not against the fund manager's conviction.

This is where the AI equity question becomes definitional before it becomes analytical. There is no settled AI sector. The AI equity universe as measured by any index provider is a construction: a set of names admitted by a rules-based screen, weighted by float-adjusted market capitalization, rebalanced on a published schedule. A semiconductor designer, a hyperscaler with a search business, a utility with data center load, and a loss-making agent infrastructure company may all sit inside the same thematic bucket depending on whose construction you use.

Which means the phrase "underweight AI" is underdetermined three times over. Underweight relative to which construction. Measured at which date. Net of which derivative overlay.

I have watched this failure mode from the inside. In 2024, working with a financial institution preparing ETF-related infrastructure, my task was to audit the custody solution's multi-signature implementation. The technical finding was not the interesting part. The interesting part was discovering that the institution's internal compliance interpretation of "cash equivalent" differed from the index provider's classification of the same instruments by two business days. Two days of classification disagreement became a two-day misstatement of the fund's neutrality. Nobody had written it down because nobody owned the definition.

That is the correct lens for the AI underweight story. Before you interpret the number, you have to establish who owns the definition. In the source item, nobody does. It is stated as weather.

Core

The four fields that decide whether the observation means anything

An allocation gap is a measurement. A measurement without units, sample, and timestamp is a rumor. Four fields determine interpretation, and the source item supplies none of them.

Magnitude. Underweight in percentage points of portfolio weight, or as a ratio to the index weight, or as active share contribution? These are not interchangeable. A fund that holds zero AI exposure against a 30% index weight and a fund that holds 26% against 30% produce the same qualitative sentence and differ by two orders of magnitude in risk implication. The first is a structural refusal. The second is noise inside a rebalancing schedule.

Benchmark. Russell 1000, S&P 500, MSCI USA, or a sponsor-constructed thematic index. The choice moves the answer more than any portfolio decision the manager made.

Sample. Twelve funds or twelve hundred. Weighted by AUM or equal-weighted. Including or excluding index-tracking vehicles.

Timestamp. Quarter-end snapshot or period average. A single quarter-end print can be manufactured by rebalancing dates, by a single large redemption, or by a tax-driven distribution cycle.

Now the fifth field, which the source item cannot even gesture at because it does not appear in holdings data: gross versus net exposure. A fund can be underweight in cash equities and overweight in the same risk through total return swaps, index options, or a thematic ETF held as a single line item. Reported holdings and economic exposure have been diverging for a decade. Post-2023 Form 13F amendments brought certain security-based swaps into disclosure, but the mapping is partial and lagged by up to 45 days.

So the honest reading of the source item is this: a directional observation with no measurement attached, published by an outlet whose incentive is the headline rather than the dataset. Verification > Reputation applies to news as much as to code. I am not accusing anyone of fabrication. I am stating that the claim is currently unfalsifiable, and unfalsifiable claims deserve no position sizing.

The underweight may be involuntary, and that changes everything

The consensus reading is that institutions looked at AI and declined. The mechanical reading is that institutions were structurally prevented from participating, and I can show the arithmetic.

Start with the tracking error budget. A large-cap US equity mutual fund typically operates under an ex-ante tracking error target between 150 and 300 basis points annualized. That budget is the fund's scarcest resource. It is spent like capital, and every active weight consumes some of it.

Approximate the standalone contribution of an active sector weight as the product of the active weight and the sector's volatility. A sector with 35% annualized volatility and a 500 basis point overweight consumes roughly 175 basis points of standalone volatility contribution before any correlation benefit. Run that through a portfolio with a realized correlation between the sector and the residual book in the 0.5 to 0.7 range, and the marginal tracking error contribution lands in the neighborhood of 120 to 150 basis points.

That is half the fund's entire error budget, spent on one sector. On a name set that includes multi-trillion-dollar issuers whose idiosyncratic news moves indices by percentage points in a session.

An investment committee does not need a view on artificial intelligence to arrive at the answer. It needs a spreadsheet. The output is capped sector exposure, and the cap is described in the minutes as "valuation discipline." The minutes are honest. The label is wrong.

Add the second constraint: liquidity classification. A newly listed agent-infrastructure company with a thin float and a wide spread sits in a lower liquidity bucket than a mega-cap, and the fund's internal basket limits bite long before the portfolio manager's conviction does. Add the third: the diversification test. Add the fourth: the daily redemption obligation, which is the constraint nobody models until it fires.

I have audited this exact arithmetic in a different container. In three weeks reviewing the initial interest rate model of a major lending protocol, the interesting finding was not a bug. It was that a liquidation threshold which looked conservative on a static basis became non-conservative inside a specific volatility regime, because the model treated a parameter as exogenous that was in fact endogenous to the utilization curve. The proof was mathematical and the impact theoretical, and the senior engineers accepted it because the derivation was reproducible rather than because it was alarming.

The same standard applies here. A sector underweight at an institutional asset manager is usually an output of risk-budget arithmetic, not an input of sentiment. Reading it as sentiment is the single most common error in market commentary, and it is an error of the same category as treating a hardcoded parameter as a market outcome.

Who actually sets the marginal price

If mutual funds are structurally capped, the marginal buyer is somebody else. In AI equities that somebody is the levered, unconstrained, short-duration player: the hedge fund complex, plus the options market, plus the systematic community.

This matters because the identity of the marginal buyer determines what the price is a function of.

A long-only mutual fund's demand is a function of benchmark weight and risk budget. It is slow, it is procyclical in the wrong direction, and it is capped.

A hedge fund's demand is a function of financing terms. Gross exposure is set by prime broker margin, by the cost of total return swaps, by the availability of stock borrow, and by the volatility-adjusted haircut on the collateral in the margin account. None of those variables have anything to do with whether a large language model can plan a multi-step task. They have everything to do with the shape of the yield curve's front end and the balance sheet capacity of four or five dealers.

That is the transmission mechanism. Equity beta in a sector where the marginal buyer is levered is priced by the financing layer, and the financing layer repriced violently in March 2020 and again in the Archegos unwind. Both episodes featured concentration at a single prime broker, synthetic exposure that was under-disclosed in the relevant filings at the time, and a margin call that arrived faster than the underlying thesis could be evaluated.

There is a second-order consequence. The leverage chain makes the AI equity complex a correlated asset, not a diversified one. Every levered holder of the same five names is exposed to the same financing shock. Their entries differ. Their exits converge.

Silence before the breach is not a phrase about fraud. It is a phrase about what filings look like in the quarter before the margin account becomes the price setter.

The on-chain mirror, and why it is worse than the original

Everything above describes a system with disclosure requirements, a prospectus, a regulator, and a 45-day reporting lag.

Now look at the on-chain equivalent.

Automated allocation vaults — the curator model that routes depositor capital across strategies on the basis of a published target — reproduce the mutual fund structure with the disclosure requirements removed. There is a benchmark, usually implicit. There is a tracking error budget, usually unstated. There is a mandate, usually encoded as a set of caps in a configuration contract. And there is a redemption obligation, which in this case is instantaneous and enforceable by a single transaction.

Restaking and liquid restaking tokens reproduce the hedge fund structure. The operator set selection is a mandate. The allocation cap per operator is a tracking error budget. The reward rate is the performance fee. The slashing condition is the margin call — and unlike a prime broker, it executes at block speed with no negotiation, no cure period, and no human in the loop.

In neither case has anyone published the tracking error budget. I have asked. The answer is usually a link to a governance forum post that sets a cap without stating the variance it is intended to control.

Here is the structural comparison, laid out the way I lay out findings for institutional clients:

| Layer | Traditional fund | On-chain allocation system | |---|---|---| | Benchmark | Published index, rules-based | Implicit target, sometimes a config file | | Error budget | Declared, monitored, reported | Undeclared, unmonitored | | Mandate | Prospectus, legally binding | Caps in a config, mutable by governance | | Redemption | Daily, T+1 or T+2, gateable | Per block, atomic, non-gateable | | Disclosure | Quarterly, audited, standardized | On-chain state, unstandardized | | Liquidity classification | Rule 22e-4 buckets, documented | Absent | | Failure mode | Underperformance, outflows | Underperformance, outflows, and a cascade |

The right-hand column is not more transparent in any way that matters. Raw state is not disclosure. A contract that stores a cap does not disclose the reasoning that produced it, the scenario under which it was calibrated, or the identity of the party who can change it.

I want to be precise about the asymmetry, because it is the actual insight. In the traditional structure, a mis-specified mandate produces gradual underperformance. In the on-chain structure, the same mis-specification produces a single block in which the vault's remaining capital is consumed by the largest redeemer. One unchecked loop, one drained vault. The loop does not have to be clever. It has to be first.

The AI-agent execution layer: where the same error becomes fast

Now the part that connects the equity allocation story to code I have actually reviewed.

In 2026 I audited an AI-agent trading platform. The architecture was conventional: a model produced allocation intent off-chain, an execution contract consumed an oracle price feed, and settlement occurred against that price within the same transaction window.

The finding was temporal, not economic. The oracle updated on a fixed cadence. The execution layer finality was shorter than that cadence. In the gap between an oracle update and the next one, the model could observe a price that the settlement layer had not yet consumed, act on it, and settle at the stale price. A slight delay in data updates was sufficient to produce a repeatable, risk-free extraction.

Reduce it to pseudocode, because this is the part that does not belong in prose:

state: oracle.price  (updated every T seconds)
state: agent.intent (computed continuously)

on agent.tick(): p_intent = model.observe(oracle.price) if p_intent != oracle.price: executor.submit(p_intent) # settlement reads oracle.price, still stale # until the next T-boundary ```

The bug is not in the model. The model behaved correctly given the interface. The bug is that the interface presented a cached value as a live one — and the settlement layer trusted it because the value was on-chain and therefore, in the naive reading, true.

That is the same category error as reading an involuntary sector underweight as a view. In both cases, a number is treated as a measurement when it is actually an artifact of the pipeline that produced it.

Code is law, until it isn't. The chain enforces the price it was given, not the price that exists.

I proposed a time-lock: a minimum delay between intent submission and settlement, sized to the oracle cadence. The trade-off is explicit and I want it stated rather than buried. A time-lock imposes a cost on legitimate latency-sensitive strategies in order to remove a cost on the protocol as a whole. It reduces capital efficiency for the honest participant. It does not eliminate the arbitrage; it relocates it to the boundary of the lock, where a slower adversary competes on block positioning rather than on compute.

That is the trade-off structure of every execution-layer mitigation in this space. Nobody has solved it. Several teams have published parameter choices that imply they believe they have.

Where the DA narrative overreaches

The natural next move in this kind of analysis is to argue that agent execution will drive data availability demand, and therefore that the DA layer is the bottleneck to price. I have run the numbers on rollup DA consumption repeatedly and I do not believe the argument holds.

The overwhelming majority of rollups generate data volumes that are a rounding error against the blob capacity available to them. The fee line attributable to DA in the median rollup's cost structure is small, and it is small in a way that has been stable across two capacity expansions. Building the investment case for a DA layer on the thesis that autonomous agents will consume it requires an assumption about agent transaction volume that nobody has documented with a measured series.

There is a sharper version of the point. The data that agents consume is predominantly oracle data — signed price attestations and state proofs — not raw transaction blobs. Oracle data does not care whether a rollup has a dedicated DA layer. It cares about signing cadence, liveness, and the number of independent publishers.

The capital and audit attention being directed at DA infrastructure is being directed at the least constrained part of the stack. Meanwhile the actual constraint — signed-data freshness across heterogeneous execution domains — receives a fraction of the scrutiny. That asymmetry is a market inefficiency, and inefficiencies in audit attention are the ones I am professionally positioned to notice.

Interoperability, value capture, and where the agents will actually run

The same misallocation appears in the messaging layer. Cross-chain messaging for agent execution is technically solved. IBC does it cleanly, with light-client verification that is more rigorous than most of the attested-bridge designs competing with it, and I have said so publicly for years.

Technical elegance has not produced an application ecosystem. Agent liquidity and agent execution keep concentrating on two EVM domains, because that is where the counterparties, the stablecoin float, and the composable collateral are. The messaging layer that routes between them is largely an attestation-based bridge with a five-of-nine signer set, which is a strictly worse security model than IBC and wins anyway.

The reason is value capture, not engineering. A protocol layer that verifies messages correctly but captures none of the economic activity flowing through it has no budget to fund the security work that its correctness depends on. That is a sustainability problem dressed as a technical one, and it is why I remain skeptical of the interoperability token whose chain the agent traffic does not use. Elegance without fee capture is a research artifact. Research artifacts do not fund audits.

What to actually measure

If the source item is under-specified, the corrective is a measurement plan. Here is the one I would run, and the fields I would require before forming a view.

On the traditional side. Form N-PORT filings give monthly portfolio holdings with a 60-day lag for the third month of each quarter, and they include liquidity classification bucket allocations. That is the highest-signal public dataset in this entire question, and almost nobody reads it because it is delivered as a structured XML file rather than a narrative. Pulling three consecutive quarters of N-PORT data across the top two hundred equity funds would answer the magnitude, sample, and timestamp questions directly.

Form 13F gives quarterly long positions across institutional managers above the threshold, with the 45-day lag and the security-based swap amendment. The differential between the mutual fund aggregate and the hedge fund aggregate is the cleanest available proxy for the flow you care about — with the caveat that the 13F population is not the mutual fund population, and the two should never be compared as if they were.

ETF creation and redemption baskets give daily visibility into the passive side, which is the part the source item omits entirely. If passive vehicles are absorbing AI exposure while active vehicles reduce it, the aggregate institutional weight may be flat or rising, and the "record underweight" is a segmentation artifact rather than a market signal. That possibility deserves to be tested before the headline is traded.

On the derivatives side. Futures basis across the relevant index contracts, options skew at the two-and-six-month tenors, and the concentration of open interest at individual strikes. Skew and basis tell you what the marginal levered buyer is paying for optionality and financing. That is the variable that actually sets the price, per the analysis above.

On the on-chain side. Concurrent things, all of them verifiable in real time:

Oracle deviation between the primary feed and at least two independent sources, sampled at the same block. If the deviation is persistent rather than transient, the execution layer is settling against a stale truth.

Vault-level cap utilization, measured as a time series rather than a snapshot. A vault at 100% cap utilization for three consecutive months is not conservatively managed. It is capacity-constrained, and the next large deposit either fails or dilutes the cap.

Redemption clustering. The distribution of withdrawal sizes within each block. A vault whose withdrawals are dominated by a single address per block has an unmodelled concentration risk that no audit report will surface, because audit reports examine the code and this is a property of the depositor set.

Operator set concentration in restaking systems, measured by the Herfindahl index of stake distribution rather than by the number of operators. A hundred operators with one holding 60% of stake is a single point of failure with better marketing.

Contrarian

The consensus reading of the underweight is that institutions are missing the AI cycle and will eventually chase. I think that reading is correct but irrelevant, and the thing it distracts from is more interesting.

Here is the counter-intuitive claim. The AI equity underweight is not the scarce asset. The scarce asset is attestation capacity.

Consider what has to happen for any of this to be safely investable. Somebody has to write the execution contract. Somebody has to write the oracle adapter. Somebody has to sign an audit report that says the settlement path is correct under a defined threat model. Somebody has to accept professional liability for that signature when the model does something the specification did not anticipate.

That supply curve is not elastic. There are perhaps a few hundred people globally with the combination of distributed-systems security depth, capital-markets literacy, and willingness to publish a finding under their own name. Every one of them is currently booked. I know because I am one of them and my calendar is closed through the next two quarters.

The regulatory environment has made this worse, not better, and it has done so in a way that almost nobody has priced. The sanctioning of a privacy protocol's smart contracts established that publishing and operating open-source code can attract liability that previously attached to conduct rather than to artifacts. The downstream effect was not that developers stopped writing code. It was that the insurers, the general counsels, and the senior reviewers at audit firms reassessed the tail risk of putting a signature on an execution-layer review. Several did not reassess it in favor of continuing.

So the real underweight is not in AI equities. It is in the number of qualified parties willing to attest to the correctness of the systems that AI capital will flow through. That underweight is not disclosed in any filing, is not measured by any index provider, and is not visible in the price of anything. It is visible in the delivery times quoted to me when I refer work out.

The second blind spot follows from the first. Everybody models allocation behavior and nobody models the operational constraint set that shapes it. If the attestation supply is fixed and the number of systems requiring attestation doubles, the marginal system ships unreviewed. That is not a prediction about intention. It is a capacity statement.

There is a third-order effect worth naming. Capital that cannot find audited venues does not sit idle. It finds unaudited venues with better yields and a governance process that moves fast because nobody is slowing it down. The mutual fund underweight may, in a roundabout way, be pushing retail-adjacent capital into exactly the venues that have never been reviewed. The headline says institutions are cautious. The plumbing says somebody is absorbing the risk, and it is probably not an institution.

Takeaway

Watch two things over the next two quarters, and neither of them is the underweight number.

First, the N-PORT liquidity classification distribution across the largest equity funds. If the AI-adjacent positions are clustering in the least liquid bucket while the fund's stated redemption profile stays daily, the constraint that produced the underweight is binding harder than the fund's disclosure implies. That is a measurable, falsifiable, and currently unmeasured condition.

Second, the count of publicly attributed audits on agent-execution contracts, net of the count of agent-execution contracts shipped. If the numerator stalls while the denominator compounds, the industry is accumulating unreviewed settlement paths at a rate it cannot inspect, and the next failure will be attributed to a model that behaved exactly as specified.

The underweight is a direction. Directions are cheap. The question worth carrying forward is the one the source item never asked: not whether institutions believe in AI, but whether the systems that will intermediate the belief have anyone left who is willing to sign their name to them. Verification > Reputation. There is currently more demand for verification than there is capacity to supply it, and no index provider has figured out how to price that.