Silicon Valley's New Mafia: OpenAI and Anthropic Are Mass-Producing Founders

Author: David, TechFlow

It has been a long time since Silicon Valley collectively used the term "Mafia."

The last time was more than twenty years ago. In 2002, eBay bought PayPal for $1.5 billion. A group of young people who had experienced the company's journey from 0 to 1 became financially free overnight and then scattered.

The subsequent stories are well-known. Musk founded Tesla and SpaceX, Peter Thiel founded Palantir, Hoffman founded LinkedIn, Steve Chen and Karim founded YouTube...

Silicon Valley calls them the PayPal Mafia.

"Mafia" is not derogatory; it is a certification, certifying that you come from that winner and have the ability to create another winner.

This term has been quiet for a long time. The conditions required are too harsh: a company successful enough to win, a concentrated distribution of wealth, and a group of people who have seen the world and have not yet been worn down. Google did not spawn a Mafia, nor did Meta. Until recently, it began to be frequently used on another group of people.

Those who left OpenAI and Anthropic.

More than half of 2026 has just passed. The number of people who have left these two leading AI companies to turn around and start new companies is already large enough to list a long roster:

Former OpenAI VP of Research Jerry Tworek founded Core Automation, former Anthropic researcher Behnam Neyshabur and others formed Mirendil. Among the researchers who just left, some are working on verifiable mathematics, some on AI that truly belongs to individuals, and others want to rebuild personal computers at the hardware level...

This path has been walked before. Anthropic itself was founded five years ago by people who left OpenAI. Now valued at $380 billion, it has become the most troublesome opponent for its former employer.

The list is still growing. These leavers are using their respective expertise to prune the branches and leaves of the AI tree.

When the Mafia Starts Claiming Territory Outside Large Models

Let's look at a question first. Why is it that almost no one from this group leaving OpenAI and Anthropic in 2026 is building large models again?

The answer is very realistic: because there is no room left on the main trunk. Training a frontier model costs billions of dollars. OpenAI, Anthropic, and Google are fighting hand-to-hand themselves. For a startup to enter head-on is equivalent to committing suicide.

But the stronger the model, the larger the open space around it.

Today's models are already smart enough, smart enough that the industry bottleneck is no longer "whether it can." Looking at this group of leavers together, you will find that they are actually focusing on "getting work done," and using their expertise to fill places where the large model's tentacles have not yet reached.

For example, can AI actually land on the level of getting work done? Once the work starts, trust becomes an issue. Once trusted, control remains an issue.

Large companies cannot attend to these layers of trouble, and some are not suitable for them to answer themselves. Almost all the companies on this 2026 list are growing on these patches of open land.

The most radical patch of open land is letting AI research AI itself.

Former OpenAI VP of Research Jerry Tworek and several colleagues founded Core Automation, creating an automated research laboratory that lets models read papers, propose hypotheses, and run experiments themselves.

The judgment behind it is quite cold: the bottleneck of AI progress is no longer algorithms, but manpower for research.

And Mirendil, formed by former Anthropic researcher Behnam Neyshabur and others, just raised $200 million. It is another extension of the same logic: a self-accelerating system that lets models participate in improving the models themselves. The human role has retreated from the subject of research to the supervisor.

Once the work starts, new trouble follows: how to confirm that it is done correctly.

Thus, another patch of open land focuses on AI trustworthiness.

In most fields, verifying an answer given by an AI is much more expensive than generating one. Math Inc is targeting this most expensive link. Former OpenAI researcher Jesse Han left to found it, with the direction of turning mathematical proofs into a form that machines can verify line by line.

Mathematics is one of the few fields where right and wrong can be thoroughly tested. Only by running verification through here is it possible to move the same capability into other industries. When AI starts making decisions for people, proving it is right is probably more valuable than letting it do it.

One layer further back is turning intelligence into "daily routine."

Between a model that can answer questions and a model that can do things for you, there is a whole set of dirty work: calling tools, breaking down tasks, remembering context, doing one thing from start to finish... These actually also need the support of auxiliary tools.

Rational and Zavify, founded by former OpenAI and Anthropic employees respectively, are focused on agent workflows, handing over business processes to agents for enterprises;

Coincidentally, former xAI co-founder Igor Babuschkin founded River AI, also wanting to create AI that truly belongs to individuals and is shaped by individuals.

There is also an outlier in this group. Blackstar, built by former OpenAI Codex engineer Daniel Edrisian, is starting directly from hardware, wanting to build a personal computer redesigned for the AI era.

After all the work is done, someone still has to watch over it.

There is a subtle opportunity here. When AI companies announce "our model is very safe" themselves, no one believes it. Players cannot serve as referees. Thus, safety has become a business that can be done independently. Those doing this business are mostly people who have personally researched AI safety in the two laboratories.

To put what they do in plain language, one is finding faults. Syntony, founded by a former Anthropic team, specializes in finding ways to induce AI to make mistakes and trick it into crossing boundaries, testing out the system's flaws before bad actors make a move.

Another is scoring. Resolution, founded by former OpenAI researchers, studies how to confirm that an AI is truly working according to human intent, and also needs to mark the degree of this certainty. Another is setting rules. Guidelight, also from former OpenAI employees, defines what practices count as safe for the entire industry, and then pushes everyone to follow them...

This batch of companies does not touch the upper limit of AI capabilities; they guard the bottom line of AI. The more capable the model, the better their business might be.

And if you ask what the difference is between this group of "AI Mafia" and the PayPal Mafia, the answer is probably dispersed vs. convergent.

The group that left PayPal did everything: payments, social, aerospace, intelligence analysis, because that was an era full of opportunities;

This new group of Mafia members is much more convergent in direction, because there are only these few patches of open land for AI. The group of people who understand models the deepest in the world have voted with their feet on the location of the next bottleneck.

Why Now?

There is a key prop in the PayPal Mafia script: a concentrated liquidity event. eBay's acquisition allowed a group of people to get money at the same time and regain freedom at the same time; only then could the Mafia gain momentum.

In the AI industry of 2026, the prop is in place.

Last autumn, OpenAI arranged a round of secondary share transfers. Employees cashed out a total of $6.6 billion, and the company valuation stood at $500 billion. Bigger events are yet to come. OpenAI and Anthropic are both preparing for IPOs, landing as early as this year. For early employees, the eve of the IPO is the best time to leave. The paper wealth in hand is about to turn into real money. If they leave later, they will be bound by golden handcuffs for a few more years.

The money is in place, and those grabbing people arrived long ago.

Silicon Valley VCs have formed a convention, staking out the exit channels of the two major laboratories. OpenAI's first head of sales, Aliisa Rosenthal, simply changed careers to become an investor, publicly saying she would rely on former colleagues' networks to find projects. Former head of consumer products Peter Deng also joined the venture capital firm Felicis.

Mira Murati mentioned in the previous chapter could raise $2 billion without a product; Mirendil just took in $200 million upon debut... Money is chasing people, to an exaggerated extent.

The risk account for the leavers themselves is actually easy to calculate. The worst result is nothing more than returning to a major tech company to get another million-dollar annual salary.

The Cost of Crowding

Currently, dozens of companies founded by the AI Mafia are crowded on the same few patches of open land. On every path stand opponents who are equally smart, equally well-connected, and equally well-funded.

Another way to say crowded is that most will lose.

The story of the PayPal Mafia is moving because we only remember Tesla and LinkedIn, forgetting the dozens of companies that fell during the same period. This list will likely be the same. Looking back in a few years, perhaps no more than five names will be recognizable.

There is also a more hidden problem. The business of these new companies is either letting AI do the work or watching AI do the work. Customers and prospects are tied to the same premise: model capabilities must continue to improve rapidly. Once this premise slows down, many patches of open land will disappear simultaneously.

But even counting all this, this list is still worth keeping.

The true legacy of the PayPal Mafia is not certain companies, but a principle: when the most important talents of an era start walking out from the same place, following them is usually not wrong.

Twenty years ago, that group defined the second half of the internet. Today, this group is gathering around this AI tree.

Those who have planted the tree know earliest which direction the tree will grow.