Model of Models — v7.3 final delivery

Journey

Good morning.

I am filled with emotions. Two years ago, we were as excited as the new cohort: innocent, but ambitious.

Last year, on this very same day, we sat where you are sitting now, at the pre-defense of the cohort before us: nervous, and a little lost.

And today is quite sad. It is our last moment to sit here together, my dear classmates, after the toughest academic challenge of our lives.

My first thanks go to our professors, because you took us on a real research journey. And honestly, I still don't know how you did research in your time, with only books, but…

No ChatGPT. No Claude.

In these two years, not only did our research journey break through, but so did AI — the fastest change in human history.

AI took only two years to reach forty percent of working-age Americans. The internet took five. The personal computer, twelve.

And the capital markets followed. In 2025, OpenAI raised forty billion dollars in a single round. Just one year later, it broke its own record: one hundred and twenty-two billion.

And Anthropic is even preparing to go public, at a value of two trillion dollars or more.

Sounds crazy, huh? The largest listing ever… for just a simple chatbox.

And most of us use that box for free.

That is what caught our attention. What is really behind that box? Why invest so much in something given away? And how can such a business survive?

It infers what you mean

At first, AI looks like a smarter search engine, with almost the same screen.

In a search engine, you type a few keywords in, and you get a list of links out.

But AI understands what you mean, in your own words, and gives you exactly what you want, in one answer.

In AI, this better matching is called inference.

It calls on other firms

And AI can do more than answer. It can act on the result.

Ask for a birthday postcard for your mom. It understands what you want, calls its own image generator to draw the card, and even an outside print-and-ship service to mail it.

This assembling of different services, inside and outside the firm, is called tool calling.

It reasons, and it learns

Another common switch is Think, for extended thinking.

Switch it on, and AI checks much more context, step by step, before it answers.

And it learns from use. What it remembers carries into your next conversation, and your ratings and corrections can help train the next version, for everyone.

This advanced way of managing knowledge is called reasoning.

And above these three functions sits one more switch, the most basic one.

The foundation model

The AI model. It is what we hear about most, and where the money goes.

Its name follows a standard the whole industry uses: the vendor, the family, the tier, the version.

In academic words, it is a foundation model: trained on broad data, and adapted to many different uses.

It is trained once, at a huge fixed cost, close to a billion dollars today. Then it is wrapped into products: a chatbox, a coding agent, a voice assistant.

So the entry is very expensive. In the largest catalogue, a hundred and thirty-one organizations offer models. But only forty-nine train their own.

All the others build on someone else's model. Few can afford to enter, and those who do compete hard.

Few enter, and each stands apart

Among these forty-nine vendors, twenty-five are American, fifteen Chinese, five European, and four from elsewhere: a race on the scale of the world economy. And each one tries to take a position of its own.

OpenAI keeps its best model closed. It is famous for its image model, but it has recently left the video race. Yet it gives away open models, and even its coding agent, Codex, is open source. And it opened the GPT Store, where outside developers build on its models and can earn from them.

Google also keeps its best model closed, but gives away many light, free open models. Its coding agent for developers was open source, then closed. Inside its own family, Gmail and the rest, its free use is the most generous. But it connects very few outside apps, and gives outside developers no way to earn from it.

Anthropic keeps everything closed, even its coding agent, and its models make only text: no audio, no video. Yet developers welcome Claude more than any other: more than sixty thousand software projects build on its library. One reason: Anthropic created open standards that let outside tools plug into Claude, such as MCP and skills.

The Chinese and French vendors open everything they offer, even their best model. Yet open does not mean adopted.

And some have already stopped offering their own models: nine of the forty-nine, Apple among them.

The big issue

We have now seen what is behind the box. Two of our questions remain: why invest so much in something given away? And how can such a business survive?

Together, they make a big issue: something difficult that deserves attention, and that matters in practice.

Running an AI model business is both challenging and crucial.

Not only because the model costs a fortune to build, is often given away, and earns nothing until others get involved,

but also because the money at stake is huge, and when users and third parties do not come, the business stops.

The managerial question

Every day, we hear investors and managers of AI vendors worry about their decisions: they raise huge sums, and still lose money.

They face many decisions at once: open the model or keep it closed, charge or give access away, build every layer or rent some, and more. Each choice changes how others build on the model, and how long the business lasts.

So this becomes a real managerial question: how do these decisions, taken together, keep an AI model business alive?

The research question

In academic words, after a comprehensive review of the literature, the decisions are strategic conditions, and keeping the business alive is business model viability.

Because the decisions are taken together, I ask about combinations. And because I want to know what is enough, I ask which are sufficient.

So my research question is: which combinations of strategic conditions are sufficient for business model viability among AI foundation model vendors?

Model of Models

So this is an organizational study in strategy:

a configurational analysis of viable business models for AI foundation models.

In short, as the title of the monograph says: Model of Models.

A hypothesis: AI as a Platform

We already know what a foundation model is. But why do its vendors spend so much, and still offer it for free? In real life, only a platform does that.

PEMAK: three functions of a platform

There is a well-known analytical framework for platform business models, built through an economic lens.

Matching brings two sides together, like a dating service.

Assembling puts the work of many suppliers into one offer, like a telecom provider that sells phone and TV together.

Knowledge management turns what users give into a shared stock, like Wikipedia.

Some do one or two. Some, like Amazon, do all three.

Platform Economics: Matching, Assembling, Knowledge management. In short: PEMAK.

Six trade-offs

Each function has two trade-offs: you gain on one side, and lose on the other.

Matching: the market goes from competition to monopoly; and the platform goes from integrated, like a generalist, to unbundled, like specialists.

Assembling: the offer goes from narrow to wide; and payment goes from a fee to free. When it is free, users pay with attention or data. The authors call it pollution.

Knowledge management: organization goes from hierarchical to spontaneous among peers; and rights go from closed to open.

Familiar platforms, then AI in the triangle

For example: an e-commerce site matches buyers and sellers, a search engine a query and pages, a social network users and content.

Each one assembles: catalogue, payment, delivery; links, maps, ads; many producers in one feed.

Each one manages knowledge: ratings, clicks, engagement.

And each one picks a side. Search engines and social networks are free: you pay with attention and data. An e-commerce site takes a fee from sellers. Wikipedia is spontaneous and open.

Now, recall the three functions of AI we saw at the start: inference, tool calling and reasoning.

Inference matches questions with answers. Tool calling assembles images, code and outside apps. Reasoning manages knowledge: it learns from your conversations. A perfect fit with the triangle.

Now take OpenAI, and its decisions on the trade-offs.

ChatGPT heads toward monopoly, with its AGI strategy. OpenAI focuses on AI, not on other computing infrastructure. One multimodal model handles text and images. ChatGPT is mostly free to use. OpenAI manages its model weights in-house, and it keeps its best model closed.

Two years ago, people still saw ChatGPT as a productivity tool, a digital service, or an AI product. Today, more and more call it a platform.

Brousseau and Pénard's conclusion: there is no optimal model, only the best practical compromise.

And as they also asked for more investigation, almost twenty years ago, I am now taking up the call.

Six choices, one configuration

To see it at once, we make a simple geometric move: the six axes turn around one centre, and each takes the name of a condition.

A choice on the named side sits on the outer ring; the other side sits near the centre. Join OpenAI's six dots, and a shape appears: its surface and its position are the configuration of its six choices.

Converting "managing knowledge" into a more AI-native word, "reasoning", PEMAK becomes the MAR radar. The same shape will stand for every vendor.

Viability: self-sustainable, not profitable

PEMAK also says that a business model is viable when it is self-sustainable.

But viable does not mean profitable: a similar study showed that many French online banks were losing money, and yet some of them were still alive.

So a business model is viable as long as it has not died.

Why give it away: complementors

If profit is not required, why do these vendors give so much away, even their models for download, after spending so much to build them?

Open or closed, they all care about the builders.

Because a foundation model is never made for the end user. It is a base: it becomes useful only when others build on it: an app, an agent, a new model.

Platform ecosystem research calls these builders complementors. Each one may bring its own users, and other users come on their own.

Attracting them is called complementor engagement. Its target is ecosystem generativity.

Generativity sounds like generative AI, and rightly so: generative AI is all about generativity.

Generativity: the triangle becomes a wheel

On the triangle, generativity is the three functions feeding one another.

Assembling plugs other firms' tools into the model: more complementors, more uses.

Matching brings users to those uses: more uses, more users.

And knowledge management turns that use into a better model, so more firms want to connect.

The triangle becomes a wheel. Generativity is the wheel turning, round after round.

Model of Models: the turning wheel and its halo

Put together, these concepts give our framework, Model of Models.

Complementor engagement is the wheel. Viability is whether the wheel keeps turning: if it stops, the business model dies. And the faster it turns, the stronger the business grows.

Generativity is the size of the halo: the more complementors and users, the larger it grows.

And the configuration, the MAR radar, sets how the wheel turns.

MoM = PEMAK × time-space: many wheels, one world, over time

PEMAK is endogenous and static: one platform, at one moment. MoM is PEMAK extended in time-space: it covers many platforms, and follows them over time.

First, space. Step back: there are many wheels, in one world.

Each vendor turns its own wheel at the centre of its own halo, and every one reaches into the same pool of complementors and users.

And the size of a business does not set the size of its halo: a small vendor can hold a large ecosystem.

Then, time. A configuration is only a position at one date.

Slide the date, and everything moves: the shapes change, businesses and halos grow or shrink, and each wheel turns at its own speed.

GEN VIA MAR

Put together, our Model of Models says that viability is the generativity that comes from the MAR configuration.

In short, MoM: GEN VIA MAR.

Our MoM, validated with QCA

The beauty of our MoM is that it can also be validated empirically, with a relatively new but fast-growing method: Qualitative Comparative Analysis, or QCA.

Since its introduction, more than a thousand papers in business and management have used it, including the award-winning EDBA paper by Rocchi, from our previous cohort.

Like qualitative research, it reads each case in depth; like quantitative research, it compares many cases at once.

Why QCA fits: conjunction, equifinality, asymmetry

It fits both the reality of this industry and the design of our framework.

No single choice works alone: the choices work together, as a whole, which is what trade-offs literally mean. QCA calls this conjunction, and treats each case as one whole configuration.

Managers' different decisions can lead to the same outcome: QCA calls this equifinality.

And a death is not simply the mirror of a survival: QCA calls this asymmetry.

Six conditions, in crisp sets

QCA fits PEMAK naturally: each of its trade-offs becomes a condition, as if the method were made for it.

QCA writes them in a simple logic: for example, open and not paid is enough for viable. In theory, six conditions give two to the power of six, sixty-four possible configurations.

This suits our question better than a regression: it handles asymmetric causes, and it gives a readable computation instead of an uncontrollable equation.

So each condition is a membership, one or zero, as in csQCA, the most common version of QCA.

Five propositions, as set relations

QCA propositions play the role that hypotheses play in quantitative methods.

The first is about necessity: every viable vendor should keep a free tier.

Three are about sufficiency: an adjacent monopoly, a community-driven model, or a user-first model should each be enough to stay viable.

The fifth is about asymmetry: no configuration should be enough for death.

And the thresholds for both necessity and sufficiency are fixed in advance, as common standards.

The whole population: a census

QCA also asks for the whole population: not only the survivors, but also the ones that failed.

No register existed, so I built my own primary dataset from the largest public model catalogue, from its first archived day to the cut-off.

Forty-nine of its organisations are vendors that build their own models.

Four are too recent to judge, and one is a boundary case, which leaves forty-four.

Several dated documents, one code

Each condition has its own written rule. Open, for example, means weights anyone can download, under a licence that allows commercial use.

And no code rests on one document. Mistral's open code, for example, rests on several kinds of public documents, and they all say the same thing: open equals one.

Across the census, the codes cite more than sixteen hundred dated documents of many kinds, and the record keeps growing.

The main thresholds were also fixed in advance, following common academic and industry standards, and tested in seventy-one variants.

Binary, blind, matrix, truth table, minimisation, in R

From there, the computation is one line.

Every cell was coded twice, the second time blind to the first, and the two agreed on almost every cell.

The result is one matrix of ones and zeros, with the outcome, viable or not, coded by its own ordered rule. It is public.

R then turns the matrix into a truth table: identical rows group into the configurations actually observed.

Then it minimises the table: when two viable rows differ in one condition only, that condition drops out. What remains are the solutions.

All of this runs in R, step by step, as you see in the code. That is the whole design, on one page.

Necessity, then three solutions

The results are still preliminary, but they already show what the framework expected.

First, one condition is necessary: nearly every viable vendor keeps a free tier.

Then R gives three solutions, from the most detailed to the simplest, and all three fit the data equally well.

From the propositions to three strategies

Where do these solutions come from? Most of their paths answer one of the five propositions.

The free tier is necessary, as expected; the monopoly, community-driven and user-first paths all lead to viability; and no configuration leads to death.

Read as ways of keeping the wheel turning, these paths become three strategies: own the whole stack, put the users first, or let the community drive.

Beside them, niches survive too. No single model wins.

The hypothesis and what we observe

This chain can be read in two ways.

The first is the hypothesis: a vendor lives on what it charges for access.

But the data show that open, free vendors without a monopoly survive too.

So what we observe is this: complementor engagement — generativity — not income, keeps an AI model vendor alive.

Seeing the results: from Venn diagrams to the UpSet plot

QCA usually shows its results with tables of configurations and Venn diagrams. With this many conditions, they are not intuitive enough.

Visualization is key to sensemaking for practitioners, according to L. Garreau et al. So I reinvent the UpSet plot, popular in business.

It shows the whole map at once: every configuration, and the vendors behind it.

And on the left stands out the one condition nearly every viable vendor shares: a free tier.

44 vendors as MAR radars

Our MoM can show the results even better: here are all the vendors as MAR radars, grouped by shape.

Viable in burgundy, ended in grey. And some ended vendors have living twins.

What this work offers

This work is among the first to treat AI as a platform, with a business model framework that can be seen and measured.

For scholars, I conceptualised MoM as PEMAK extended in time and space: generativity via the MAR configuration.

And I validated it empirically with QCA, on the whole industry.

For practitioners, business strategies can now rest on a rigorous framework, with configurations that are easy to see and to update.

And I hope it helps policymakers too: they can now see a vivid map of the whole industry.

Limits, next steps, and thanks

But the industry is still young, and the data are still new, so the validation is not yet complete.

Next come fuzzy sets and longitudinal analysis, and more measures: the size of each ecosystem and of each business, and how fast each wheel turns.

This draft has been discussed at academic events with experts in strategy, economics and methods, and with many young PhD students in the field, around the world.

The answer

To conclude, back to the research question.

Across a whole industry, many different business models stay viable, and each of them does so through generativity.

AI works as a platform, but not one where the winner takes all.

Keeping the generativity is what keeps an AI model business alive.

Thank you.

One more thing

Before we finish, one more thing.

Everything is open, and kept up to date, on the MoM site.

You will find all the data and results, for every vendor.

And you can find me on GitHub.

Thank you.