Researchers Who Left Anthropic Raised $200M in Weeks — Meet Mirendil, the 'AI That Researches AI'

A company with no product and no revenue just raised $200M

Here's the deal: Mirendil, a company founded in early 2026, closed a $200 million seed round on June 25. It was valued at roughly $1 billion, co-led by a16z (Andreessen Horowitz) and Kleiner Perkins with Nvidia participating. At a glance it reads as "another AI startup raise" — but the shock is elsewhere. This company has no product and no revenue. It's only months old, and its founders were working at places like Anthropic just weeks ago.

And it still pulled one of the largest seed rounds in AI history. Seeds are usually a few million to tens of millions; Mirendil took $200 million at the seed stage and became a $1 billion company. The reason is exactly one thing — what this team said it would build: an AI that does AI research itself.

So here's what we'll unpack: who Mirendil is, what it's actually building, why investors bet $200M with no product, and what "AI that researches AI" really means.

The players — 20 researchers who walked out of Anthropic

First, the Mirendil team. The key is the people. The founding crew is about 20 engineers and researchers, and the pedigrees are jaw-dropping: Anthropic, xAI, Google DeepMind, OpenAI — the exact labs building AI right now. Co-founders include former Anthropic researcher Behnam Neyshabur, along with Harsh Mehta, Shayan Salehian, and Tara Rezaei — veterans of large-scale machine learning and optimization research. The company is based in San Francisco.

Next, the money: a16z, Kleiner Perkins, and Nvidia. a16z and Kleiner Perkins are among Silicon Valley's most storied VCs, and Nvidia is the undisputed king of AI compute. Three names like that in a single seed round is itself a signal. Nvidia's participation in particular carries the calculation that "this team will burn enormous amounts of our GPUs" — automating AI R&D demands massive compute.

The third lead isn't a company but an idea: "AI that does AI research." Mirendil wants to build AI that does what human researchers do — design experiments, iterate on engineering problems, make AI systems progressively better — with steadily less human input. Put simply: an "AI researcher who researches AI," written in code.

Tie it together: 20 researchers from the world's top AI labs raised $200M and a $1B valuation — with not a single line of product — on a promise to automate AI research itself. That's the spine.

What's actually confirmed

Words scatter, so here's the table.

Item	Detail
Round closed	June 25, 2026
Amount	$200M (seed)
Valuation	~$1B
Lead investors	a16z, Kleiner Perkins (co-led), Nvidia participating
Founded	Early 2026
Team	~20 (ex-Anthropic, xAI, DeepMind, OpenAI)
HQ	San Francisco
Product / revenue	None
Goal	Frontier models that perform AI R&D
Long-term vision	Let open-source devs & scientists build specialized models (medicine, materials)

Row by row. First, "no product / revenue" and "$200M" sitting in the same table is the whole point. Investors usually bet on product, customers, revenue — Mirendil has none. So what did they buy? Team and vision. A bet that "if anyone can crack this, it's these people." It's a wager on humans that's uniquely possible in AI.

Second, the goal — "models that perform AI R&D" — is loaded. Mirendil isn't building a chatbot or a coding helper; it's building AI that's better at building AI. Designing experiments, driving GPUs directly, looping over research and engineering problems — a coding-agent-style AI research automation platform. If it works, the pace of AI progress itself accelerates: a kind of self-acceleration loop.

Third, the long-term vision is "democratization." Mirendil argues frontier AI work has been "locked inside a few big labs" and wants to let anyone do AI work — open-source developers or scientists building their own specialized models in fields like medicine and materials. The pitch is breaking the big labs' monopoly.

What each side gets

Mirendil's team. First, overwhelming runway — $200M buys more top talent and the compute to train frontier models. Second, time — years of pure research without product or revenue pressure. Third, powerful backers in Nvidia and a16z, locking in compute, capital, and network at once.

The investors. For a16z and Kleiner Perkins, if "AI that researches AI" actually works, it becomes core infrastructure for the next generation of AI. Getting in early at $1B and watching it become a $10B or $100B company would be a legendary return. High-risk, but the kind of bet that, when it hits, carries an entire fund. Nvidia's gain is a different flavor — whether Mirendil succeeds or not, that massive GPU demand flows back as Nvidia revenue.

The surprise winner: the AI talent market. A team walking out of a top lab and raising $200M in weeks signals that "in AI, a few proven researchers can become a unicorn instantly." For top labs that's a flashing brain-drain warning; for researchers it makes the lure of independence even stronger.

Past parallels — wins and losses

These "mega-seeds for star-researcher teams" are a recurring pattern in the current AI boom. The closest parallels are new AI companies founded by ex-top-lab researchers that drew hundreds of millions to billions in valuation right out of the gate. The logic is identical: in AI, elite talent is the moat — grab those people and the rest follows. Mirendil's "$200M with no product" extends that logic.

On the success side, star teams have at times shipped powerful models fast and vindicated the bet. Proven researchers already know what to build and how, so with enough capital they can move quickly. Investors bet pre-product because they've seen this play out.

But the shadow of failure is real too: a star team is no guarantee. Some raise huge rounds and lose direction, set goals too ambitious to realize, or can't close the compute-and-data gap with the giants. "AI that researches AI" is an ultra-hard problem the whole industry is racing on — there's no guarantee a 20-person team cracks it before Google or OpenAI. The $200M is a starting line, not a finish line.

Competitors' counter-plays

The most direct rivals are, ironically, the very labs the team left. Anthropic, OpenAI, and Google DeepMind are already pursuing "use AI to accelerate AI research" internally, with their own researchers using tools that automate model improvement. To them, Mirendil is the challenge of "spinning out what we do in-house and doing it faster." The big labs dwarf Mirendil on compute, data, and headcount, so Mirendil has to win on speed and focus.

Other AI-automation startups will counter too. Plenty of companies aim to automate coding and research with AI agents. Mirendil's differentiator is "specializing in AI R&D itself, not general coding," and rivals can push back that "betting only on AI research is too narrow a market" while chasing broader software development. Who's right is for the market to decide.

The big labs' talent defense matters as well. As the Mirendil story spreads, top labs will have to raise compensation, autonomy, and equity to retain key researchers. Paradoxically, Mirendil's emergence pushes up talent costs industry-wide. For the giants, the counter is simply "retain them with better terms."

So what actually changes

If you're an AI developer or researcher, this is an intriguing signal. If Mirendil's long-term vision — letting open-source developers build their own specialized models — pans out, you might one day get tools to build domain models (medicine, materials) without a big lab. There's no product yet, so for now it's anticipation, but the "AI research automation" trend is worth watching.

If you follow startups or investing, Mirendil sits squarely in the "AI bubble" debate. Whether a $1B valuation with no product or revenue is "a rational bet on the future" or "the peak of froth" splits the room. One side says "talent is everything in AI"; the other says "valuation with no substance." Too early to call — but more deals like this clearly raise both the market's heat and its risk together.

If you watch the whole industry, Mirendil symbolizes the hottest topic of all: recursive self-improvement. If AI starts getting better at building AI, progress could accelerate nonlinearly — an opportunity, and a safety concern. That ex-Anthropic people are doing this is meaningful precisely because the folks who prized safety most are now betting on this direction.

One step further — what "$200M with no product" really means

To read this right, see why investors pour large sums into a company with no product or revenue. The answer: in AI, time is everything. Training models takes enormous upfront investment, and product is years away. So for proven teams, the structure has hardened from "wait for product, then invest" to "put in big money first so they can survive those years." Mirendil's $200M isn't a reward for revenue — it's capital buying future time.

Another easy-to-miss thread is "why AI R&D automation specifically." The biggest bottleneck in AI right now isn't compute or data — it's people. Elite AI researchers are vanishingly few, and their time is the most expensive resource there is. If AI can take over even part of the research process, the single biggest bottleneck loosens. That's exactly what Mirendil targets — automating the rate-limiting step of AI progress. Succeed, and the whole industry's clock speeds up. That's why a16z and Nvidia bet.

Caveats, coldly. First, difficulty: "AI that researches AI" is only partially solved even by the field's best minds working together; no guarantee 20 people beat the giants to it. Second, the compute gap: big labs run hundreds of thousands of GPUs — whether $200M closes that gap is unknown. Third, timing: getting in at $1B amid a bubble debate may later look like foresight or like buying the top. Time answers.

In the end, Mirendil's $200M isn't just a funding headline — it's a coordinate showing where AI money is betting biggest right now. Not a chatbot, not an app, but "AI that builds better AI." Whether that bet becomes the accelerator of AI progress or a symbol of the bubble's peak is something Mirendil's first product, years out, will answer.

🥄 Three Things You're Probably Wondering

— $200M with no product? Isn't this a bubble? You could see it that way. But "putting big money into proven teams to buy them time" has become standard in AI. The room splits between "rational bet" and "froth." Too early to call — the first product will tell.

— Is "AI that researches AI" actually possible? Partly, yes — big labs already do it internally. But fully autonomous, near-hands-off research is still far off and ultra-hard. Whether Mirendil cracks it first, nobody can promise.

— Will this ever help a developer like me? Long-term, possibly. If the vision holds, tools could let open-source developers build their own specialized models. But with no product yet, treat it as "a trend to watch" for now.

Sources

• Investing in Mirendil — Andreessen Horowitz
• Former Anthropic Researchers Launch Mirendil at $1 Billion Valuation With $200M Seed Round — Unite.AI
• Mirendil raises $200M in one of AI's biggest seed rounds — Cryptopolitan
• Ex-Anthropic researchers raise $200M just weeks after quitting to build AI that creates better AI — TechFundingNews
• Mirendil secures $200M seed round led by a16z and Nvidia — CryptoBriefing

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!