NextFin News - Mirendil, a San Francisco artificial-intelligence startup founded by two researchers who left Anthropic after barely a year, is in talks to raise new funding at a $5 billion valuation, according to people familiar with the matter. The discussions come roughly three months after the company closed a $200 million seed round at a $1 billion valuation, a fivefold repricing in a single quarter that underscores how aggressively investors are bidding for startups building self-improving AI.
The size of the round was not disclosed, and the final terms could still change, the people said. But the $5 billion figure, if it holds, would place Mirendil among the most richly valued AI ventures to emerge from stealth in 2026, alongside the startups launched by former OpenAI executives.
The question the round poses is bigger than one startup's valuation. Mirendil is not trying to build a better chatbot. It is trying to sell the engine that the frontier labs keep for themselves: AI systems that do the work of AI researchers — designing experiments, searching for hyperparameters, evaluating models, and running the next training cycle. If that engine works, the moat built by the big labs — thousands of researchers, years of accumulated knowledge, closed-source models — starts to look less permanent. If it does not, Mirendil is a $5 billion bet on a capability that may never leave the labs that invented it.
The Deal: A Fivefold Reprice in One Quarter
Mirendil emerged from stealth on June 24, 2026, announcing a $200 million seed round at a $1 billion post-money valuation. Andreessen Horowitz and Kleiner Perkins co-led the financing, with Nvidia among the participating investors. At the time, it was one of the largest seed rounds ever recorded in artificial intelligence — a nine-figure check for a pre-product company with no shipped model.
Three months later, the company is discussing a valuation that would multiply that price by five. The speed of the repricing is the story. For context, Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, raised $2 billion at a $12 billion valuation in July 2025. Safe Superintelligence, led by former OpenAI chief scientist Ilya Sutskever, has raised roughly $8 billion while marked at a $32 billion valuation. Recursive Superintelligence, another venture built around self-improving AI, came out of stealth in May 2026 with $650 million at a $4.65 billion valuation. Mirendil's proposed $5 billion price would put it in that peer group — a lab with a founding team, a thesis, and no public product, valued as if the product already worked.
The capital is chasing a specific scarcity. Artificial intelligence absorbed roughly half of all global venture funding in 2025, reaching $211 billion, up 85% from $114 billion the prior year, according to Crunchbase data. The AI infrastructure market alone ended 2025 near $337 billion in revenue and is forecast to reach $1.2 trillion by 2030. In that environment, a founding team with pedigrees from Anthropic, Google DeepMind, xAI, and OpenAI — and a thesis that targets the layer beneath every other AI company — commands a premium that would have been unthinkable two funding cycles ago.
Mirendil's founding team is small and senior. The company runs on about 20 researchers and engineers, drawn from Anthropic, xAI, Google DeepMind, and OpenAI. Co-founders Behnam Neyshabur and Harsh Mehta met at Google in 2019, moved to Anthropic in late 2024, and left in December 2025, soon after the launch of Claude Opus 4.5. Neyshabur, now chief executive, has worked on AI for science for more than seven years across Google and Anthropic, where he co-led the Discovery team. Mehta built the first version of Anthropic's internal AI-research platform, initially as a team of one. The founding group also includes Shayan Salehian, a core member of xAI's machine-learning engineering team since its Twitter days, and Tara Rezaei, a 23-year-old MIT graduate and olympiad medalist.
Self-accelerating AI R&D is the most direct path to delivering on AI's broader promise. — Behnam Neyshabur, co-founder and chief executive of Mirendil, in the company's founding announcement
The Thesis: Selling the Engine the Labs Keep for Themselves
The biggest AI labs share a private conviction: the fastest way to build better AI is to point AI at the problem of building AI. They run that loop in-house, and their terms of service stop outsiders from doing the same. Anthropic has said its policy of forbidding customers from using its tools to build competing services is standard among model providers and helps keep frontier AI out of foreign adversaries' hands. As of May, the company said its Claude model wrote more than 80% of the code it merged into its production codebase — a measure of how deeply the loop is already embedded inside the frontier labs.
Mirendil is betting that the lock creates a market. Neyshabur frames the product as AI for AI for science, not merely AI for science. The distinction matters. A drug company does not want a general chatbot; it wants a system that can design and refine a model for drug-target identification without hiring a machine-learning team. The pitch is that work which takes frontier labs months can be compressed into days for an outside organization — a university biology lab, a pharmaceutical company, a materials researcher.
That is a harder thesis to validate than a vertical application, and a larger prize if it holds. Mirendil is not competing for the same budget as a coding agent or a customer-service bot. It is pitching the research-automation layer that companies like Recursive Superintelligence and Periodic Labs would themselves run on. Periodic Labs, another Andreessen Horowitz-backed company, raised $200 million to apply AI to materials-science discovery. Mirendil wants to sell the infrastructure beneath that use case — and beneath every other AI research effort that wants to move faster than its in-house team allows.
To train the systems, the company has secured compute. Mirendil signed a multi-year Google Cloud partnership worth more than $100 million for TPU and GPU capacity, according to reporting on the company's financing. The founders have said a model and product will arrive in the coming months. Until then, the $5 billion valuation is a bet on a team and a mechanism, not on revenue.
Why the Valuation Is a Structural Bet, Not Just a Cyclical Spike
A fivefold repricing in one quarter looks like euphoria. It is — but the euphoria is layered on top of a structural shift in where value accrues in the AI stack.
The structural claim is this: for the past two funding cycles, value in AI has accrued to whoever owns the frontier model. OpenAI, Anthropic, and their peers captured the capital, the talent, and the pricing power because training a frontier model required resources only a handful of companies could assemble. Mirendil's thesis is that self-improving research tooling decouples capability growth from headcount and from closed model access. If an independent vendor can hand a biology lab the same research-automation loop that Anthropic runs internally, the moat shifts from who owns the model to who owns the automation layer. That is a regime change, not a cycle. Regime changes in the stack — from on-premise servers to cloud, from licensed software to SaaS — are where the largest venture outcomes are made.
The cyclical claim is equally real, and it is the risk. The velocity of the repricing — $1 billion to $5 billion in roughly three months, with no shipped product — is a function of how much dry powder is chasing how few pedigree teams. When capital is this abundant and the supply of founder teams with Anthropic and OpenAI on their resumes is this thin, prices reprice first and fundamentals catch up later, if ever. The $5 billion figure is the market pricing the structural possibility at cyclical speed.
The second-order consequence is the one the round's supporters are not saying out loud. If Mirendil's tooling works, it does not just help outsiders — it accelerates every competitor of the big labs at once. Recursive Superintelligence, Thinking Machines Lab, and a dozen smaller model builders would all run faster on the same automation layer. That dilutes the advantage of scale that the incumbents have spent years building. The irony is sharp: the labs' own invention — using AI to build AI — becomes the lever that breaks their lock, but only if they are willing to buy it from an independent company rather than build it themselves.
Andreessen Horowitz put that logic plainly in its investment announcement: Most of the major AI labs have built some version of this platform for internal use, the firm wrote. But they are not necessarily designed for external users, and the labs also do not have a strong economic incentive to release the platforms for their competitors to use. The conclusion: So, we think there really should be an independent company focused on this problem. Mirendil wants to be that company. The $5 billion valuation is the price of the bet that the firm is right.
The Counter-Thesis: The Labs Still Hold the Keys
The strongest case against Mirendil is simple: the frontier labs still hold the keys, and they have every incentive to keep them. They have the best researchers, the largest compute clusters, the proprietary data flywheels, and the revenue to fund in-house automation indefinitely. An independent tooling vendor gets the customers the labs decline to serve — the universities, the mid-market research teams, the organizations that cannot afford a frontier partnership. That is a real business, but it may be a smaller one than the $5 billion valuation implies.
There is also a safety and regulatory wall. Anthropic has pointed to recursive self-improvement — a model rewriting its own code without oversight — as a potential danger, on the theory that it could slip beyond human control. The company disabled access to its most powerful Fable 5 and Mythos 5 models in mid-June after the Trump administration issued an export-control directive citing national security authorities. A startup whose entire product is handing self-improving research capability to outsiders is walking directly into that regulatory scrutiny. If policymakers decide that research-automation tooling is dual-use technology that must stay inside licensed labs, Mirendil's market shrinks overnight.
The founders disagree. They call recursive self-improvement the shortest path to faster science, and a problem that can be supervised rather than avoided. But supervision is easier to promise than to prove, and the burden of proof will fall on the company that is selling the capability, not on the labs that are restricting it.
The specific signal that would prove the structural thesis wrong is concrete: if, by the third quarter of 2027, no independent AI-research-automation vendor — Mirendil or a peer — has shipped a product with a named, paying enterprise or university deployment that demonstrably compresses a frontier-model research cycle by an order of magnitude, then the independent-company thesis fails. The labs' in-house loops win, the tooling layer stays captive, and the $5 billion valuation becomes a marker of how high prices can go before fundamentals arrive.
What to Watch: Three Horizons
Short term (the next 3–6 months): The deal itself. If Mirendil closes at or near $5 billion, the round becomes the new pricing anchor for every pedigree AI lab that raises in late 2026 and 2027. Watch for the round size, the lead investor, and whether Nvidia participates again — a repeat strategic check would signal that compute providers see the tooling layer as essential infrastructure, not a sideshow.
Medium term (6–18 months): The product. Mirendil has said a model and product will arrive in the coming months. The falsifiable test is whether a named customer — a university lab, a pharmaceutical company, a research institute — publicly demonstrates a research cycle compressed from months to days using Mirendil's platform. Without that, the valuation is a resume premium, not a revenue multiple.
Long term (18 months and beyond): The regime question. Does value in the AI stack migrate to the automation layer, or does it stay with the model owners? The base case is a split outcome: independent tooling vendors capture the long tail of research organizations that the big labs will never serve directly, while the frontier labs keep their own loops in-house and continue to set the pace. The upside case is that the tooling layer becomes the new chokepoint — the place where capability growth is bought and sold — and Mirendil, as the first well-capitalized independent vendor, captures a durable franchise. The downside case is that safety regulation and the labs' own vertical integration squeeze independent vendors into a narrow, low-margin niche.
Scenarios, not a single line: base case, Mirendil closes the round and ships a product that finds real customers but does not dislodge the labs' in-house advantage; upside case, the platform becomes the default research-automation layer for non-lab organizations and the valuation proves conservative; downside case, the product never ships at frontier quality, or regulators treat self-improving research tooling as dual-use technology that must stay inside licensed labs, and the $5 billion price becomes a cautionary tale.
The closing judgment: Mirendil's $5 billion valuation is not a bet that self-improving AI will work — the frontier labs have already proven that to themselves. It is a bet that a capability the labs invented for internal use can be productized, sold, and trusted by the organizations they shut out. That is a structural wager wrapped in cyclical pricing, and the next twelve months will show which layer is real.
Explore more exclusive insights at nextfin.ai.
