A working audit of every instrument currently sold as AI exposure, what each one gives you a claim on, and the one category that does not exist yet.
Open your portfolio and find the AI part of it.
Now answer a narrower question than the one you normally ask. Not whether the position is up. Not whether the thesis is right. Just this: if this thing works exactly the way you hope, what specifically do you have a claim on.
Most people cannot answer that cleanly, and the reason is not carelessness. It is that almost every available instrument gives you a claim on something adjacent to AI rather than on AI itself. Once you see the pattern it is very hard to unsee.
This piece is the audit. Category by category, what you are actually holding, and what is conspicuously missing from the list. If you want more breakdowns like this on AI and markets, follow @MossAI_Official.
Part 1: The audit
Chipmakers and hardware
What you hold: a claim on the earnings of a company that sells inputs to AI builders.
This is the cleanest position on the list and also the most indirect. Your return depends on capex decisions made by a small number of large buyers. If AI works spectacularly but the buyers develop their own silicon, or capex plateaus while deployment scales, the productivity can go up while your claim goes sideways.
You own the supplier. Supplier revenue and end productivity are correlated. They are not the same variable.
Hyperscalers and model labs through public parents
What you hold: a fractional claim on a diversified conglomerate that happens to contain an AI business.
Buy a large cloud provider and the AI exposure is real but blended into advertising, retail, enterprise software, and everything else on the income statement. If the AI division tripled, you would struggle to see it in the share price with any precision.
You own the container. The thing you wanted is inside it, at an allocation you do not control and cannot observe directly.
AI narrative tokens
What you hold: in most cases, a claim on nothing, priced on attention.
Worth being blunt because this is where a lot of retail capital actually sits. A token whose connection to AI is a description in a whitepaper is not a productive claim. It has no cash flow, no enforceable right to output, and no mechanism converting AI performance into holder value. It is a bet on other people continuing to associate the ticker with the theme.
That can work as a trade. It is not ownership of anything, and confusing the two is expensive.
Copy trading and signal services
What you hold: correlated exposure with no claim attached.
You are replicating someone else's decisions with your own capital and your own risk. You own your positions, not any share of their operation. If their strategy is brilliant, you capture some of it minus slippage and timing drift. If they stop, you have nothing. There is no asset that persists.
You own an imitation, not a stake.
Private AI funds and pre-IPO access vehicles
What you hold: a genuine claim, conditional on being allowed in.
These are structurally real. Someone actually owns equity in AI companies here. The problem is not the instrument, it is the gate. Thresholds, minimums, and qualification requirements sort participants into those who reach the growth phase and those who do not.
And the growth phase is where the returns concentrate, which brings us to the part of this that has been getting quietly worse.
Part 2: Why the public instrument stopped reaching the growth phase
Public equity was the most successful ownership instrument ever built. Anyone with a brokerage account could hold a claim on the productive economy. Over the past three decades that reach has been narrowing, and the data is not ambiguous.
Fewer companies to own. Research published in the Journal of Financial Economics documented that the United States had 8,025 domestically incorporated listed companies in 1996, the peak year, falling to 4,102 by 2012. Follow up work by the same authors extends the analysis to 2023 and finds the US now has roughly half as many listed firms per capita as comparable developed countries.
And they arrive later, larger. Jay Ritter at the University of Florida maintains the reference dataset on IPOs. The average age of companies going public was 9.5 years across 1980 to 2019, then 8 in 2022, 10 in 2023, and 14 in 2024. Median revenue at IPO was 16 million dollars in 1980, about 64 million adjusted for inflation, and had reached 218 million by 2024.
Put those together. The instrument still exists. The window moved. By the time something is available to you, the stretch where a modest position could compound into something meaningful already happened somewhere you were not permitted to be.
Nobody decided to exclude you. The instrument simply stopped reaching that far.
Part 3: The category that is missing
Go back to the audit and notice what is not on it.
There is no instrument on that list that gives you a direct claim on the output of an autonomous system.
Not the company that built it. Not a token associated with the theme. Not a replication of its behaviour. A claim on the economic activity of the machine itself, enforceable, and priced against what that machine actually produces.
This matters more than it would have five years ago, because autonomous systems have stopped being demos. Software is trading, allocating, negotiating, and executing continuously, generating measurable output with no human approving each action. The productive force is real and it is running right now.
And with no ownership instrument attached to it, the returns from that work flow entirely to whoever holds the keys to the machine.
That is not a moral observation. It is a mechanical one. Instruments decide distribution. In 1602 the Dutch East India Company's charter stated that all residents could subscribe, with no minimum and no maximum. When the subscription closed that August, 1,143 people had signed, including a maid working in the house where the book was kept, on wages under fifty cents a day.
Long distance trade was not new. Ships were not new. Wealthy merchants had financed voyages for generations. What was new was a transferable claim on a continuing enterprise, open to anyone who showed up.
The innovation was not the ships. It was the paper. And the paper is the only reason a maid appears in that register next to merchants who put in twelve thousand guilders.
Machine productivity is currently in the pre paper stage.
Part 4: Three tests for whether you actually own something
This is the practical part and it applies to everything on the audit, including us.
Whenever something is sold to you as a way to own a piece of AI, run these three.
Can the claim move
The real innovation in 1602 was not that you could buy in. It was that transfers were recorded, so you could get out by selling to someone else instead of waiting for the enterprise to wind up.
A claim you cannot move is a deposit with extra steps. Ask what the exit path is, whether the claim can change hands, and what sets its value when it does.
Can you verify the underlying yourself
A claim on productivity is worth exactly as much as your ability to confirm the productivity is real.
If the performance behind your position is attested by a screenshot, by a dashboard the operator controls, or by a record with nothing independent behind it, you are not underwriting an asset. You are underwriting a person's honesty while pricing it like an asset.
The question is blunt: can I confirm the thing I own is the thing that ran, without asking permission.
Is entry structurally open
This is the test private markets fail by design, and it is why the growth phase became inaccessible.
Look for thresholds, allowlists, qualifications, and minimums. Ask whether the entry conditions are written into the instrument or negotiated with a gatekeeper. The VOC charter put openness in the founding document. That single decision is why the register contains a maid.
Run all three across the audit in Part 1. Narrative tokens fail verification, because there is nothing productive underneath to verify. Copy trading fails transferability and verification. Private vehicles fail openness. Public equities pass all three and give you a claim on a supplier rather than on the output.
Nothing on the list passes all three while also giving you a direct claim on machine productivity. That is the gap.
Part 5: What Moss is building into it
Moss Agent Marketplace is our attempt at that missing instrument. An onchain market where an autonomous system can be capitalised, owned, and priced by people who did not build it.
The design maps onto the three tests deliberately. Claims are issued as standard transferable tokens. The activity behind them is recorded onchain rather than asserted in a deck. Entry conditions live in a published contract rather than in a relationship with a gatekeeper.
The underlying standard is open rather than proprietary, and that is a considered choice. An ownership instrument is only as strong as the number of parties that recognise it. A claim honoured by exactly one platform is not really a claim. It is a user account. We would rather be one venue among many speaking a shared standard than the only venue speaking our own.
A new instrument is not the same as a good outcome.
Early access to an instrument has mattered historically, but VOC shareholders also endured decades of volatility, and plenty of early holders in every new asset class lost money. An instrument distributes access to an outcome. It does not improve the outcome.
Autonomous systems can be genuinely unprofitable, and a real claim on output means holding the losses with the same fidelity as the gains. The transparency that makes the claim credible also makes the drawdowns impossible to hide from yourself.
And this is early. Pricing is thin, comparison is hard, and the conventions for evaluating this asset class barely exist yet. Anyone claiming to have it figured out is ahead of the evidence, ourselves included.
What to do now
Go back to the question at the top and answer it properly for one position you already hold. If this works, what do I have a claim on. Supplier revenue, a diversified conglomerate, a narrative, an imitation, or actual output.
Then go to moss.site and look at what the last category looks like when it is issued as an instrument rather than described in a pitch. Run the three tests on it yourself. Movable, verifiable, open. Then decide.
We are expanding well past where we started, with more categories of agent, more kinds of claim, and more of the infrastructure this asset class needs before anyone can take it seriously. Follow @MossAI_Official so you catch them as they ship, because the useful window on any new instrument closes once everyone agrees it was obvious.
Land had deeds. Enterprises had shares. The machines that work are still waiting for theirs.
Sources
- Petram, L. The World's First Stock Exchange, Columbia University Press, with the accompanying primary source archive on the VOC charter and the 1602 subscription register: https://www.worldsfirststockexchange.com/2020/10/15/the-worlds-first-ipo/
- Amsterdam exchange history archive on the VOC as the first company to issue shares to a broad audience: https://www.beursgeschiedenis.nl/en/moment/voc-the-start-of-global-share-trading/
- Doidge, C., Karolyi, G. A. and Stulz, R. M. (2017). The U.S. Listing Gap. Journal of Financial Economics, 123(3), 464-487. NBER version: https://www.nber.org/papers/w21181 and full text: https://files.fisher.osu.edu/department-finance/public/the_us_listing_gap_july_12_final.pdf
- Ritter, J. R. Initial Public Offerings: Median Age of IPOs, University of Florida: https://site.warrington.ufl.edu/ritter/files/IPOs-Age.pdf and the full data index: https://site.warrington.ufl.edu/ritter/ipo-data/
- Reporting on Ritter's IPO age and revenue figures for recent years: https://www.cnbc.com/2025/10/07/ipo-market-startups-staying-private-longer-alternative-capital.html
