Silicon Valley has started using the word "mafia" again, and this time it is being attached to people leaving OpenAI and Anthropic.
In a commentary published by TechFlowPost, author David argues that the term has reappeared because the two leading AI labs are now producing founders at scale. Through just over half of 2026, he writes, enough former employees have left OpenAI and Anthropic to start companies that the list is already long.
A startup wave is forming around the edges of large models
The article frames the trend against the better-known PayPal story. In 2002, eBay acquired PayPal for $1.5 billion, and a group of employees who had seen the company go from zero to one left with money, experience, and freedom. They later went on to build or help build Tesla, SpaceX, Palantir, LinkedIn, and YouTube, among others. Silicon Valley called them the PayPal Mafia.
Here, "mafia" is not used as an insult. The piece describes it as a kind of credential, proof that someone came out of a winning company and might be able to build another one. David says the word had been largely dormant because the conditions are hard to reproduce: a company that wins big enough, a concentrated liquidity event, and a cluster of people who have seen scale but have not been flattened by it.
Now the label is being applied to alumni of OpenAI and Anthropic. But unlike the PayPal cohort, this group is not scattering in every direction or rushing to build another frontier foundation model. The reason, as the article puts it, is straightforward. Training a cutting-edge model can cost tens of billions of dollars, and OpenAI, Anthropic, and Google are already fighting for the core layer. For a startup, walking directly into that contest is close to suicidal.
What has expanded instead is the space around the model. As model capability rises, the bottleneck is no longer just whether the system can do something. The harder questions now sit one layer outward: can AI actually get work done, can the result be trusted, and can it be controlled? According to the article, many of the startups founded by former employees are gathering in exactly those open spaces that large labs either cannot address quickly or are not well positioned to answer themselves.
Automated research and self-improving systems are drawing founders
The most aggressive bet in that open ground is to let AI research AI.
Former OpenAI research vice president Jerry Tworek and several colleagues launched Core Automation. The company is building an automated research lab where models read papers, generate hypotheses, and run experiments themselves. The judgment behind the company, the article says, is blunt: the next bottleneck in AI progress is not only algorithms, but the supply of human researchers.
Mirendil is presented as a closely related extension of that same idea. The startup was formed by former Anthropic researcher Behnam Neyshabur and others, and TechFlowPost says it has just raised $200 million. Its focus is on self-accelerating systems that let models participate in improving the models themselves. In that setup, the human role shifts away from being the main researcher and toward supervising the loop.
Once AI starts doing more of the work, another problem shows up fast: how do you know it got the answer right? That is where the article places the next large opening, trust and verification.
In many fields, it can be more expensive to verify an AI-generated answer than to produce one in the first place. Math Inc is targeting that expensive layer. Founded by former OpenAI researcher Jesse Han, the company is trying to turn mathematical proofs into a format that machines can verify line by line. The article notes that mathematics is one of the few domains where correctness can be tested all the way through. If verification can be made to work there, the same approach may later be carried into other industries. Once AI begins making decisions on behalf of people, proving that the answer is correct may be worth more than generating the answer.
Founders are also pushing into agent workflows, personal AI, and hardware
The report then moves to a more practical layer: how to turn model intelligence into systems that can actually handle daily tasks.
A model that can answer questions is not the same as a model that can carry a task from start to finish. Between those two sits a lot of dirty work: calling tools, breaking down tasks, remembering context, and seeing a job through. The article says these functions need their own support systems.
Rational and Zavify, founded by former employees from OpenAI and Anthropic respectively, are working on agent workflows that allow enterprises to hand business processes to AI agents. River AI, launched by former xAI co-founder Igor Babuschkin, is trying to build what the article describes as AI that truly belongs to the individual and is shaped by that individual.
One company stands out for approaching the problem from the device layer instead. Former OpenAI Codex engineer Daniel Edrisian founded Blackstar, aiming to build a personal computer redesigned for the AI era.
Safety, red-teaming, and scoring are becoming standalone businesses
If AI is going to do more work, someone has to watch it. The article sees a subtle opening here. When an AI company says its own model is safe, few people take that at face value. The athlete cannot also be the referee. That makes safety a business that can stand on its own, and many of the people building it are the same ones who previously worked on AI safety inside the labs.
TechFlowPost groups their work into several buckets.
- Syntony, founded by a former Anthropic team, focuses on trying to make AI fail. It probes the system, pushes it toward mistakes, and tests whether it can be tricked into crossing boundaries before bad actors try the same thing.
- Resolution, founded by former OpenAI researchers, studies how to determine whether an AI is truly following human intent and how to attach a measurable confidence level to that judgment.
- Guidelight, also started by former OpenAI employees, is working on defining what counts as safe practice across the industry and pushing others to follow those rules.
These companies are not extending the upper bound of model capability, the article says. They are trying to protect the floor. If models become more capable, their businesses could become more valuable.
This group is narrower than the PayPal Mafia
David draws a contrast between the old PayPal generation and the current AI founder class with one simple pairing: dispersion versus concentration.
After PayPal, founders spread into payments, social platforms, aerospace, and intelligence software because opportunity was broad. This AI cohort is moving in a more converged way because the open spaces around large models are limited in number. In the article’s framing, some of the people with the deepest understanding of AI models are effectively voting with their feet on where the next bottlenecks are likely to appear.
Liquidity and venture money helped trigger the exodus
The PayPal Mafia had a defining ingredient: one concentrated liquidity event. eBay’s acquisition gave a group of employees cash and freedom at the same time. The article argues that AI now has its own version of that setup.
Last autumn, OpenAI arranged a secondary share sale that allowed employees to cash out a combined $6.6 billion, bringing the company’s valuation to $500 billion, according to the report. A bigger event could still be ahead. TechFlowPost says both OpenAI and Anthropic are preparing for listings, with the earliest possible timing this year. For early employees, that creates a narrow window: leave before listing, when paper wealth is close to becoming real cash, or stay and remain tied up for years under golden handcuffs.
Venture capital has already moved in. The article says Silicon Valley firms are now effectively waiting around the exit channels of the two labs. OpenAI’s first sales head, Aliisa Rosenthal, has become an investor and has said publicly that she plans to use her former colleagues’ network to source deals. Former consumer product lead Peter Deng has also joined venture firm Felicis.
The financing examples in the piece are meant to show how aggressively capital is chasing talent. Mira Murati, the article says, raised $2 billion without a product, while Mirendil pulled in $200 million as soon as it emerged. Money is following people, and it is doing so at an extreme level.
From the founders’ side, the downside case is unusually easy to calculate. The worst outcome, as the article puts it, is simply going back to a large company and taking another seven-figure annual compensation package.
Crowding is real, and most of these startups may still lose
The piece does not end as a simple celebration of an AI founder boom.
It points out that dozens of companies created by this new "AI mafia" are clustering in the same few pockets of opportunity. On each path, the competitors are similarly smart, similarly connected, and similarly well funded. Crowding has another name: most players lose.
The PayPal story is inspiring partly because people remember Tesla and LinkedIn and forget the dozens of companies from the same era that did not make it. David suggests this list may look the same in a few years. Of all the names being discussed now, perhaps fewer than five will still be widely recognized later.
There is also a quieter risk underneath many of these business models. Whether a startup is building systems that let AI do work or systems that monitor AI doing work, both the customer demand and the market outlook depend on the same assumption: model capability must keep advancing quickly. If that pace slows, many of today’s open spaces could shrink or vanish at the same time.
Even with those caveats, the article argues that the list is worth watching closely. The real legacy of the PayPal Mafia was not just a handful of famous companies. It was the broader lesson that when the most important talent of an era starts leaving the same place, following where they go is often a useful guide to what comes next. Twenty years ago, that cohort helped define the second half of the internet era. Today, this one is gathering under the AI tree, and the people who planted it may still be the first to see where it will grow.