Mirendil is a San Francisco-based AI research startup that closed a $200 million seed round on June 25, 2026 at a roughly $1 billion valuation — one of the largest seed financings ever recorded for an AI company. The round was co-led by Andreessen Horowitz (a16z) and Kleiner Perkins, with participation from NVIDIA. Mirendil was founded by Behnam Neyshabur and Harsh Mehta, both former Google and Anthropic researchers. The team includes alumni from Anthropic, OpenAI, Google DeepMind, and xAI. Mirendil's mission is unusually narrow: build AI systems specifically designed to accelerate AI research itself — not consumer chatbots, not enterprise productivity software, not coding agents. The goal is AI that helps human researchers iterate faster on better AI, with eventual self-improvement loops on specialized domains like biology, chemistry, medicine, materials science, and robotics.

Why is Mirendil's $200M seed valuation so unusual?

Two reasons. (1) Size — $200M seed at $1B valuation is among the largest seed rounds in any sector, ever. The previous reference points are Mistral's 2023 seed ($113M at ~$260M) and SSI's 2024 seed ($1B at $5B). Mirendil sits in that rarified tier. (2) Stage and revenue — Mirendil has no public product, no public revenue, no public benchmarks, and 20 employees. The valuation is entirely a bet on founder pedigree and thesis. Both lead investors (a16z and Kleiner) have raised giant 2026 AI funds and are deploying aggressively; Mirendil is a high-conviction bet at the top of those funds' typical seed range. The implicit message: in the 2026 AI super-cycle, founder quality and thesis specificity command pre-product unicorn valuations.

Who is Behnam Neyshabur?

Behnam Neyshabur is Mirendil's co-founder and one of the most-cited researchers in modern deep learning theory. He spent years at Google Research and then Anthropic working on generalization theory, optimization, and the foundations of why neural networks work. He is known for influential papers on implicit bias of gradient descent, generalization bounds for deep networks, and the empirical study of why over-parameterized models generalize. His co-founder Harsh Mehta worked with him at Google and Anthropic. Their starting team of 20 researchers and engineers is drawn from the four leading frontier labs (Anthropic, OpenAI, Google DeepMind, xAI), which is itself a notable recruiting signal in the 2026 talent market — particularly given the same week saw two senior departures from Google DeepMind (Noam Shazeer to OpenAI, John Jumper to Anthropic).

How is Mirendil different from other AI labs?

Mirendil's positioning is recursive: AI for AI research. Anthropic and OpenAI build frontier models for general consumer and enterprise use. xAI focuses on truth-seeking and grok-style products. Google DeepMind ranges from products to fundamental research. Mirendil's narrow target is the AI research community itself — researchers, ML engineers, scientific institutions building specialized models. The closest comparisons are Sakana AI (Japan-based, evolutionary AI research methods), Goodfire (mechanistic interpretability), and parts of DeepMind's Gemini-research team. Mirendil's bet: as frontier models become commodities, the bottleneck shifts to research velocity, and tools that accelerate research itself become the highest-leverage layer in the AI stack. Whether this works depends on whether AI-assisted research yields visible, paper-citation-level acceleration within 24 months.

Quick Answer

Mirendil $200M Seed: AI Building AI (June 2026)

Published: June 26, 2026

Mirendil $200M Seed: AI Building AI (June 2026)

Mirendil closed a $200 million seed round at a roughly $1 billion valuation on June 25, 2026 — one of the largest seed financings ever recorded for an AI company. Andreessen Horowitz and Kleiner Perkins co-led, with NVIDIA participating. The founding team (Behnam Neyshabur and Harsh Mehta) and 20 researchers and engineers are drawn from Anthropic, OpenAI, Google DeepMind, and xAI. The thesis is unusually narrow: build AI systems specifically designed to accelerate AI research itself, not consumer chatbots or enterprise software.

Last verified: June 26, 2026.

TL;DR

$200M seed at $1B valuation — among the largest seed rounds in any sector, ever
Co-leads: Andreessen Horowitz (a16z) and Kleiner Perkins; NVIDIA participated
Founders: Behnam Neyshabur (deep learning theory) and Harsh Mehta (ex-Google, ex-Anthropic)
Team: 20 researchers from Anthropic, OpenAI, Google DeepMind, xAI
Mission: AI to accelerate AI research itself — biology, chemistry, materials, robotics
Context: same week saw Noam Shazeer leave Google DeepMind for OpenAI, John Jumper leave for Anthropic; senior-AI-talent mobility is at all-time highs

The thesis

Mirendil’s pitch is that the bottleneck in modern AI is no longer scale or compute — it’s research velocity. The argument:

Pre-training is becoming commodity. Multiple labs (OpenAI, Anthropic, Google DeepMind, xAI, Meta, Mistral, Qwen, DeepSeek) now produce roughly comparable frontier models. Marginal returns on raw scale are shrinking.
Post-training is where capability gain happens. The largest delta between models in 2025-2026 has come from RL fine-tuning, agent-specific training, and domain-specialized models — not from raw pre-training.
Research velocity determines who wins post-training. A lab that runs 10x more experiments per researcher discovers better post-training recipes faster.
AI-assisted research compounds. If AI can help researchers design experiments, analyze results, generate hypotheses, and write code, then research velocity itself becomes the input to capability gain.
The highest-leverage layer is therefore tools for AI researchers. This is what Mirendil is building.

The recursive nature is the point: AI that builds AI. The risk: this thesis assumes specific bottlenecks that may not be the actual bottlenecks. If pre-training scale matters more than the team believes, the thesis weakens.

Why this valuation works

A $200M seed at $1B is the kind of round that requires multiple things to be true simultaneously.

Founder credibility

Behnam Neyshabur is one of the most-cited researchers in deep learning theory. His work on generalization, implicit regularization, and why over-parameterized networks work has been foundational for the past decade. Harsh Mehta has the operator and product side. The team is 20 people drawn from four frontier labs.

In the 2026 talent market — where senior researchers regularly command multi-hundred-million-dollar packages to switch labs — a 20-person team with this pedigree is a >$1B asset just on hiring cost replacement.

Investor dynamics

a16z and Kleiner both raised giant new AI funds in 2026. Kleiner closed $3.5B across two funds in June 2026; a16z has a similar-scale AI vehicle. Both firms need to deploy capital into category-defining bets, and the supply of those bets is constrained.

NVIDIA’s participation is the second-order signal: NVIDIA has been investing strategically across the AI stack, and a check in Mirendil hedges against the possibility that AI-for-AI-research becomes the next dominant compute consumer.

Comparable rounds

Company	Round	Amount	Valuation	Date
Safe Superintelligence (SSI)	Seed	$1B	$5B	Sept 2024
Mistral	Seed	$113M	~$260M	June 2023
Mirendil	Seed	$200M	~$1B	June 2026
Reflection AI	Seed	$130M	$555M	June 2024
World Labs (Fei-Fei Li)	Seed	$230M	$1B	Sept 2024

Mirendil sits comfortably in the “AI research with credentialed founders, pre-product unicorn” tier. The pattern is established; the question is execution.

What Mirendil will actually build

The public description is general. Inferring from the team’s prior work and the thesis:

Likely product directions

Experiment management for ML research. Tools that help researchers run, track, and reason about thousands of training experiments. Mirendil could be the successor to Weights & Biases, but AI-native.
Automated hypothesis generation. AI that proposes ablation studies, hyperparameter sweeps, architecture changes, and reads the literature to suggest related work.
Domain-specialized model builders. AI that helps scientists build specialized models for biology, chemistry, robotics — by handling data preparation, model selection, fine-tuning, evaluation.
Research code agents. Long-horizon agents that can implement papers, run experiments, generate plots, and write the methods section of new papers.
AI co-author / AI principal investigator. The endpoint: AI systems that can independently run research programs, with human oversight on direction and review.

Unlikely product directions

Consumer chatbot — explicitly ruled out
Generic enterprise productivity — outside the thesis
Frontier model competitor — they would compete with their own investor network (NVIDIA, plus a16z and Kleiner have other frontier-model bets)

Why now

The June 2026 timing is significant on three axes.

1. Senior talent mobility is at all-time highs

The same week as Mirendil’s announcement (June 18-19, 2026), Noam Shazeer left Google DeepMind for OpenAI and John Jumper left Google DeepMind for Anthropic. Meta has hired multiple senior OpenAI and Google researchers in 2025-2026. Senior AI talent is the most mobile asset class in tech right now, and Mirendil’s team of 20 from four labs is the most direct expression of that pattern as a company-formation event.

2. The capex-to-revenue gap is widening

In June 2026, AI stocks sold off on concerns that hyperscaler capex (hundreds of billions in 2026) is outpacing measurable revenue from AI. Forrester reported that 25% of planned enterprise AI spending is being pushed to 2027. PwC found 56% of CEOs see no AI revenue or cost benefits yet. The market is asking harder questions about whether more compute = more capability. Mirendil’s thesis (research velocity > raw scale) aligns with this skepticism.

3. Tool layer is undermonetized

Foundation models, applications, and infrastructure are well-monetized layers. The research-tools layer (W&B, Hugging Face, MLflow, Comet) has consolidated and is comparatively underdeveloped relative to the size of the AI research workforce. Mirendil bets that this layer is the next category of large outcomes.

What could go wrong

Thesis is wrong. If pre-training scale matters more than research velocity, Mirendil’s tools accelerate the wrong thing. Frontier labs continue to win on raw capex.
Frontier labs build it internally. Anthropic, OpenAI, Google DeepMind, and xAI all have strong incentives to build their own research-acceleration tools, and they have the data and compute to do it better than an external company. Mirendil needs to ship something the labs can’t or won’t build.
The 20 founding researchers leave. With senior AI talent so mobile, the team itself is the product, and any one of them could leave for a $50M-$200M package elsewhere. Retention is the operational risk.
Valuation creates execution pressure. A $1B seed compresses the runway-to-justify timeline. The next round will need to justify $3B-$5B+ to mark up. That requires demonstrable research velocity wins within 18 months.
The category never materializes. AI-for-AI-research could remain a workflow improvement rather than a standalone category. In that case, Mirendil’s TAM is smaller than the valuation requires.

How to think about Mirendil as a builder

For most AI builders, Mirendil is not directly relevant in 2026:

It will not have a public product for some time
If/when it does, the first product will likely target frontier labs, well-funded scientific institutions, and large enterprise research teams
General AI app builders will not interact with Mirendil tools

For researchers and AI/ML engineers: watch for Mirendil’s first product announcement (likely 2027) and any open-source releases.

For investors and operators: Mirendil is the cleanest expression of “AI research tools” as a category bet. Its trajectory will signal whether the category is real.

Bottom line

Mirendil is a maximum-conviction bet on a specific thesis: that research velocity is the new bottleneck in AI, and tools to accelerate AI research itself are the highest-leverage layer in the stack. The founders, team, valuation, and investor list all reflect that conviction. The execution risk is real and the thesis is contestable. The next 18-24 months — what Mirendil ships, who uses it, and whether observable research-velocity gains result — will determine whether this $1B seed is remembered as visionary or as the peak of the 2026 AI funding cycle.