Global Flow Strategy · By Exponential Technology · November 21, 2026
Anthropic's debut restocked the AI funds. Ten private books woke up holding $7.5 billion or more at the offer price. After a win like that, venture money goes back to work within 12 to 18 months. That clock started Thursday, and there is one place it has an obvious reason to go: the layer that turns raw market data into something a model can actually use.
The next check goes to the layer that makes AI useful to people who manage money: classified flow, positioning, and forecasts structured so a model can actually reason with them. Exponential is building that layer, and we are opening a Series A to build it at scale.
What this note covers, section by section:
1. The Marks, and What They Actually Mean — the numbers, and what a stake is worth when it becomes a check.
2. What the First Session Showed — who actually bought the debut, read from the tape.
3. Why the Money Moves — how a liquidity event turns paper into spending power, and on what schedule.
4. The Problem the Models Cannot Solve — why a language model still knows almost nothing about what happened in the market this morning.
5. The Translation Layer — what it would take to fix that, and who is building it.
6. What We Can and Cannot Prove — where these numbers come from and what they do not say.

1. The Marks, and What They Actually Mean
Start with the top of the table, because two funds carried half the paper.
Anthropic sold its shares to the public at $83 each on Wednesday. Anyone who owned stock before that moment suddenly owned stock with a public price on it, which is what people mean when they say a stake got marked. Lightspeed's book was worth about $28 billion at that price, against roughly $3.5 billion it had put in. Iconiq's was worth about $26 billion, against $6 billion invested.
Both of those positions worked out to just under two percent of the company, which is the number the Wall Street Journal reported from its own sources on November 19. Two percent of a very large company is a very large amount of money, and it is concentrated in two hands.
Behind them sits a second group. Menlo Ventures held about $22 billion, built on the largest single check the firm had ever written, a $500 million lead on Anthropic's Series D round. A lead investor in a venture round negotiates the terms and usually sets the price, which is why leading carries weight beyond the money. Coatue held about $16 billion across three separate rounds. GIC, Singapore's sovereign wealth fund, held about $15 billion and co-led the February round alongside Coatue. Eric Schmidt held about $15 billion from a position he took in 2021, when the company was valued around $461 million.
The rest of the table tells you the most about how this worked. Spark Capital turned $75 million into roughly $11 billion, better than ten times its money, and it is the only venture fund with a board seat named in the filing. Sound Ventures held about $11 billion, General Catalyst about $10 billion, Bessemer about $7.5 billion on something near nine times its investment.
Outside the venture table sit the strategics, which is the word for companies that invest for business reasons rather than financial ones. Amazon held close to 15% and Google close to 14%, both largely through a $26 billion private placement several of them funded together. Those positions were not built to be sold. They were built to make sure the compute and the distribution stayed close.
What would make this section wrong is a mark that does not survive the lock-up. Forbes put Iconiq at $31.8 billion on the first day's close rather than the offer price, which tells you the number moves with the stock. Either way the shape of the table holds, and either way these are published estimates rather than disclosed allocations.
2. What the First Session Showed
The marks tell you what the stock is worth. The flow tells you who actually bought it, and the answer is not the people you would expect.
Exponential's Equity Flow classifies every trade on the US consolidated tape as institutional or retail, so the first session can be read rather than guessed at. Anthropic traded $86.4 billion across 5.3 million trades on its debut. Institutions sold $47.73 billion and bought back $35.63 billion, which works out to $12.10 billion more sold than bought. Retail did the opposite, buying $1.57 billion against selling $1.47 billion, for a net $109 million in.

Read that plainly. The big accounts were net sellers into the listing by more than $12 billion, and smaller accounts were the buyers on the other side. That is a familiar pattern at a large debut: early holders who are now allowed to sell take some off the table, and the demand that meets them comes from people buying a first position rather than adding to a stake. It is also why a first-day pop and a first-day flow can point in opposite directions without either being wrong.
One session is one session, and the Friday data had not landed when this was written, so treat this as a first read rather than a trend. But it is a real read rather than a story, and it sets the baseline for everything that follows: the question from here is whether institutions keep selling into strength, or whether the selling was a one-day clearing event.
3. Why the Money Moves
A fund that has just been paid does not sit on the money. It puts it back to work, and the historical window for that is 12 to 18 months.
Here is why. A venture fund raises money from its own investors and has a fixed period to deploy it, usually five years or so. Cash sitting idle is a problem, because the fund earns its fee on capital committed rather than capital parked. When one of its companies goes public, two things happen at once: the fund's track record improves, which makes the next fundraise easier, and the partners suddenly have a very visible reason to be writing new checks.
That money is often called dry powder — committed capital that has been raised but not yet spent. A mark like this one does not create new dry powder by itself, because the shares are usually locked up for months after a listing. What it does is reset the conversation. It makes the next fundraise easier, it makes the partners more confident, and it makes limited partners more willing to commit again.
Most of this table has already shown its hand. Lightspeed filed to buy more shares in the offering itself, which is an unusual thing for a fund sitting on a $28 billion position to do unless it wants more exposure. Iconiq ran crossover orders. A crossover fund is one that invests in both private and public companies, and the point of running crossover orders is to keep buying after the listing rather than cashing out.
The rest simply have room. Ten books at $7.5 billion or more, most of them in early, most of them still holding. That is an enormous amount of realized and near-realized capacity pointed at one industry, and it needs somewhere to go.
What would make this wrong is a market that stops cooperating. A long stretch of risk-off sentiment freezes new commitments regardless of how good anyone's marks look, and a lock-up that drags longer than expected keeps the paper locked. Neither is in the tape right now.
4. The Problem the Models Cannot Solve
Every one of those funds is now under pressure to explain what comes after the model. The honest answer is that the models are not yet useful enough to institutions, for a reason that has nothing to do with intelligence.
A large language model is exceptionally good at language. It can read a filing, summarize an earnings call, write code and hold a conversation. What it cannot do, on its own, is tell you who sold 40 million shares of a stock between 10:15 and 10:30 this morning, what kind of account they were, what they paid, and what the person on the other side was doing.
That is not a shortcoming of the model. It is a missing input. The model has no way to observe the market directly. It reads text. Markets do not produce text; they produce a torrent of individual trades, each with a price, a size, a timestamp and two counterparties, and the meaning of those trades is not written anywhere obvious.
Two things have to exist before a model can help an institutional investor make a decision. The first is an ontology — a structured map of what the things in the market are and how they relate. What is a market maker, what is a pension fund, what does it mean when one sells to the other, which of those relationships matter and which are noise. Without that map, a model reading trade data is reading a phone book with no names in it.
The second is the micro-level facts themselves, at the granularity where decisions actually get made. Not that the market fell 2%, but that a specific population of accounts sold a specific quantity into a specific kind of buyer over a specific window. That is the difference between knowing the symptom vs knowing the cause.
What would make this wrong is a model that gets there without the data — one that learns market structure well enough from public prices to reason about it reliably. Five years of effort suggests it cannot, because the information simply is not in the public record.
5. The Translation Layer
This is the category Exponential is building in, and the way to describe it is a universal translation layer for all markets.
The way to think about it is a layer that sits on top of the raw feeds. Underneath is the firehose — every trade on the US consolidated tape, every futures position reported to the government, every economic release as it lands. No model can drink from that directly, and the feeds themselves do not explain what any of it means. What has been missing is a common layer that turns those feeds into something a machine can reason with, and that layer does three jobs. It classifies, which is the job Equity Flow does on the stock tape and Futures Flow does on positioning — deciding whether a given trade came from an institution or a retail account, or whether a futures book belongs to a fund or a producer, and in which direction. It organizes, placing each observation into a structure that says what it is and what it connects to. And it delivers, handing the model a clean, dated, checkable fact rather than a pile of numbers.
The difference this makes is easy to see in a single question. Ask a model why a stock fell yesterday and it will give you a plausible story assembled from headlines. Ask it with Exponential's Equity Flow data underneath and it can tell you that institutions sold $2.47 billion of that stock in the session after earnings while retail bought $404 million of it, which is a different answer and a more useful one. The first is a narrative. The second is a fact you can act on or argue with.
The same is true of macro numbers. When the government reports inflation, the headline number and the pieces underneath it tell different stories, and the pieces are what determine what happens next. Exponential's Macro Economic Forecasts publishes numbers for those pieces eight days before the official release, component by component, with a documented accuracy record that runs about 50% better than consensus. That record is what makes it possible to say, ahead of time, that a number will be driven by one component rather than another — which is precisely the kind of judgment a model needs to borrow.
The point is what this becomes once it exists. Real-time flow analytics stop being one vendor's dataset and start being a common substrate — a shared layer that any model can sit on top of, the way an operating system sits under applications. The funds above are looking for the next layer of the same trade, and the next layer is not another model. It is the layer that carries the information and the relationships those models need before they can be useful to anyone managing real money. That category has a small number of serious participants and a very large amount of unmet demand.
What would make this wrong is consolidation. If the large platforms build or buy this layer first, the independent window closes. That is a real risk and it is why timing matters more here than in most categories.
6. What We Can and Cannot Prove
Everything above the line comes from published reporting, and the footing matters.
Four outlets, one table. Bloomberg published its tally of the biggest winners on November 19. The Wall Street Journal carried the top two stakes and the just-under-two-percent figure. The Information reported on early investors' returns, including Spark's better-than-ten-times outcome. Forbes marked Iconiq at the first day's close. Nothing contradicts across them, which is why we are comfortable using the numbers.
What is not disclosed matters just as much. Allocations were never made public. Unnamed sources filled the gaps in the reporting. None of it may account for dilution, which happens when a company issues more shares and everyone's percentage shrinks. The exact per-fund share counts sit in the prospectus, in a table called principal and selling stockholders, and that is where anyone who needs precision should go.
What is disclosed is the structure of the offering itself. The company sold 956.4 million shares at $83. Existing holders — founders and people who sold in the April tender — sold 23.6 million more. The founders' block came to $1.75 billion, and the April tender cleared $4.66 billion at $25.91 a share, which tells you how far the valuation moved in seven months.
7. Bottom Line
We think the lasting consequence of this listing is a restocked deployment cycle, not a repriced balance sheet.
The marks will move with the stock and most of them will be forgotten within a year. The clock will not. Twelve to 18 months from Thursday, a meaningful share of this table has to put money to work, and the most defensible place to put it is the layer that makes AI useful to the institutions that will pay for it. That layer is data, structure and classification, and it is being built now.
Here is what would change our mind, with dates. A risk-off stretch through the first quarter of 2027 freezes commitments regardless of marks. Lock-up expiries turn holders into sellers and pull capital back out. Consolidation in the data layer removes the independent option before the money arrives. Any of the three would reset the argument.
Longer term we are constructive on the data and translation layer through the 2027 deployment cycle. This is not a view on any single fund's marks, and it is not a recommendation about any security.
One practical consequence. Some of the funds in this table already hold positions in the infrastructure that AI needs to be useful, and the ones that do not are actively looking for it. If we had to guess, you may find them on the cap table for the Series A for Exponential providing it with the resources to build out the substrate layer described above at scale. If you manage capital and this thesis is the one you have been circling, that is the conversation to have. Reach out to sales@exponential-tech.ai.
We will revisit this when the fourth-quarter deployment data prints.
The trade ticket is stamped. The clock is running.
This material is provided to institutional investors for informational purposes only. It is not investment advice, an offer, or a solicitation. Third-party stake values are published estimates as of the dates shown and may not reflect dilution. © Exponential Technology Inc.






Comments