startups

The $18.8 Billion AI Brain Drain: Why Top Researchers Are Leaving Big Tech to Build the Next Generation of AI

• 12 min read

The $18.8 Billion AI Brain Drain: Why Top Researchers Are Leaving Big Tech to Build the Next Generation of AI

In January 2026, Bloomberg coined a new term for the venture capital phenomenon sweeping artificial intelligence: the "coconut round." It was a tongue-in-cheek escalation of the traditional "seed round" — because the funding amounts had grown so large that calling them seeds seemed absurd.

The name stuck because the phenomenon demanded recognition. In 2026 alone, venture capitalists have funneled $18.8 billion into AI startups founded since the start of 2025, according to Dealroom data. That's on track to surpass the $27.9 billion invested in companies launched since the start of 2024. And the most striking pattern in this funding frenzy isn't the amount — it's who's receiving it.

The recipients are the architects of modern AI, the researchers who built the systems now powering billions of user interactions. David Silver, who led AlphaZero at DeepMind. Yann LeCun, Meta's former chief AI scientist and Turing Award winner. Tim Rocktäschel, DeepMind's former principal scientist. Anna Goldie and Azalia Mirhoseini, who contributed to Anthropic's most advanced work. These aren't entrepreneurs with pitch decks and prototypes. They're scientists with decades of accumulated expertise who have decided that the next breakthrough won't happen inside a Big Tech lab.

The Numbers Behind the Exodus

The scale of this migration is unprecedented in technology history. Consider the funding announcements from the past three months alone:

Ineffable Intelligence: David Silver's months-old startup raised $1.1 billion at a $5.1 billion valuation — a record seed round led by Sequoia Capital and Lightspeed Venture Partners, with participation from Index Ventures, Google, Nvidia, and the UK's Sovereign AI fund. The company had barely launched its website.

AMI Labs: Co-founded by Yann LeCun after his departure from Meta, raised $1.03 billion at a $3.5 billion pre-money valuation to build "world models" — AI systems that can learn from continuous real-world data.

Recursive Superintelligence: Co-founded by DeepMind's former principal scientist Tim Rocktäschel, reportedly raised $500 million with enough demand to stretch to $1 billion. The company is pursuing self-improving AI systems.

Ricursive Intelligence: Founded by former Anthropic and DeepMind researchers Anna Goldie and Azalia Mirhoseini, raised $335 million across two rounds in December and January to build AI tools for semiconductor design.

Periodic Labs: Founded by former OpenAI and DeepMind staff, raised $300 million in September to develop autonomous research laboratories.

Humans&: Launched by former Anthropic and xAI employees, raised $480 million in January for reinforcement learning-based AI systems.

The pattern is unmistakable. The people who built the current generation of AI at Google, Meta, OpenAI, Anthropic, and xAI are now building the next generation elsewhere. And investors are betting billions that their elsewhere will produce something better.

Why Now? Three Forces Driving the Exodus

This isn't simply a story of disgruntled employees seeking independence. Three structural forces are converging to make this the most consequential talent migration in AI's history.

1. The Narrowing Focus of Big Tech Labs

The race for AI dominance has created a paradox: success requires focus, but focus eliminates exploration. As Elise Stern, managing director at French VC Eurazeo (which backed AMI Labs), explained to CNBC: "When you're in a race, you narrow focus. That creates a vacuum. Entire areas of research, like new architectures, agents, interpretability and vertical models, are being deprioritised, not because they don't matter, but because they don't win the immediate race."

Consider what's happening inside the major labs. OpenAI has systematically eliminated "side quests" — cutting Sora video generation, delaying ChatGPT's planned erotica features, and restructuring its science department — to focus on enterprise and coding revenue. Google's DeepMind has reorganized around Gemini deployment. Meta has concentrated on integrating AI into its advertising and content recommendation systems.

This narrowing creates two problems. First, it leaves genuinely important research directions underfunded and understaffed. Second, it frustrates the researchers who were drawn to these labs precisely to pursue ambitious, exploratory science. As HV Capital partner Alexander Joël-Carbonell told CNBC: "Inside the large foundational labs, the pressure to deliver benchmark performance and maintain rapid release cycles leaves limited room for genuinely exploratory research, particularly outside the dominant LLM paradigm."

2. The Rising Doubt About Scaling Laws

A growing number of AI researchers are questioning whether simply making large language models bigger will be enough to reach the next level of AI capability. The scaling hypothesis — that increasing model size, data, and compute will automatically produce more capable systems — has driven billions in investment. But the returns may be diminishing.

AMI Labs articulated this concern explicitly: "AI has made major progress in content generation, but still struggles with grounding, causality, and reliable behavior in real-world settings. As AI moves beyond screens into industry, robotics, healthcare and other physical environments, those limitations become increasingly important."

This skepticism is fueling interest in alternative approaches. Ineffable Intelligence is betting on reinforcement learning without human data — a technique where AI learns purely through trial and error, similar to how AlphaZero mastered chess and Go without studying human games. AMI Labs is pursuing "world models" that understand physical causality. Ricursive Intelligence is applying AI to chip design, a domain where raw language model capability is less important than specialized reasoning.

The researchers leaving Big Tech aren't just starting companies. They're starting companies built on fundamentally different technical assumptions than the ones driving their former employers.

3. The Economic Opportunity of Independence

The financial incentives for leaving have never been stronger. The "coconut rounds" are delivering valuations that would have been unthinkable for pre-revenue startups even two years ago. Ineffable Intelligence reached a $5.1 billion valuation before shipping a product. AMI Labs hit $3.5 billion pre-money. These aren't outliers; they're becoming the standard for AI labs founded by recognized researchers.

For the founders, the upside is enormous. For Big Tech, the cost of retention is rising proportionally. Google, Meta, and OpenAI can offer competitive compensation, but they can't offer the ownership and autonomy of a startup. And they increasingly can't offer the intellectual freedom that attracted top researchers in the first place.

The London Effect: Geography as Competitive Advantage

An underappreciated dimension of this trend is geographic clustering. London is emerging as the unexpected epicenter of AI startup formation, driven largely by DeepMind's gravitational pull.

Ineffable Intelligence is based in London. Recursive Superintelligence is incorporated in the UK. Jeff Bezos's AI lab, Project Prometheus, is reportedly negotiating for office space near Google's AI hub in King's Cross. The UK government's Sovereign AI fund is actively co-investing in these companies, providing both capital and political legitimacy.

This clustering isn't accidental. DeepMind's continued presence in London after its 2014 acquisition by Google created a dense network of talent, expertise, and institutional knowledge. Former DeepMind employees know each other, trust each other, and have experience working together on ambitious projects. When they leave to start companies, they don't start alone — they start with teams that have already proven they can ship breakthrough systems.

The London ecosystem also benefits from proximity to European regulatory frameworks. As the EU AI Act takes effect and the UK develops its own AI governance approach, European-based labs may have advantages in navigating compliance requirements that their American counterparts lack.

What Are They Actually Building?

The diversity of approaches among these startups reveals something important about the current moment in AI: the field is fragmenting. There is no longer consensus about what the next breakthrough looks like or how to achieve it.

Reinforcement Learning Without Human Data

Ineffable Intelligence, founded by David Silver, is the purest expression of this alternative vision. Silver was the architect of AlphaZero, the system that taught itself to play chess, shogi, and Go at superhuman levels without any training on human games. At DeepMind, he spent over a decade developing reinforcement learning systems that learn from experience rather than imitation.

Ineffable's stated goal is to create a "superlearner" capable of "discovering knowledge and skills without relying on human data." The company's website makes a striking claim: "If successful, this will represent a scientific breakthrough of comparable magnitude to Darwin: where his law explained all Life, our law will explain and build all Intelligence."

This is audacious even by AI startup standards. But Silver has credibility. He didn't just work on AlphaZero; he led the team that produced it. And his investors — Sequoia, Lightspeed, Nvidia, Google — are betting that his track record justifies the ambition.

World Models and Physical Understanding

AMI Labs, co-founded by Yann LeCun, is pursuing a different critique of current AI. LeCun has long argued that large language models, despite their impressive capabilities, lack genuine understanding of the physical world. They can generate plausible text about physics, but they don't have intuitive models of how objects behave, how causality operates, or how to plan actions in complex environments.

AMI's "world models" approach aims to build AI systems that develop these intuitions through continuous interaction with real-world data. The potential applications span robotics, autonomous systems, scientific discovery, and any domain where AI needs to operate effectively in physical environments rather than just generate text about them.

Specialized Vertical Applications

Ricursive Intelligence represents a third strategy: applying advanced AI to specific, high-value domains where deep expertise matters more than general capability. The company is building AI tools for chip design, leveraging the founders' experience with Google's AlphaChip project.

As co-founder Anna Goldie explained to CNBC, this vertical focus serves a business purpose as well as a technical one: "For chipmakers to trust us with their most valuable IP, we have to be Switzerland, and that wouldn't be possible if we were at Google." Independence from Big Tech becomes a competitive advantage when selling to enterprises that view those same tech giants as potential competitors.

The Investor Logic: Why Bet Billions on Pre-Product Companies?

The venture capital enthusiasm for these startups might seem irrational. These companies have no products, no revenue, and no proven business models. Yet they're raising more money at higher valuations than most late-stage software companies.

The logic becomes clearer when you understand what investors are actually buying. They're not buying products; they're buying optionality on the future of AI itself.

As Eurazeo's Stern explained: "Founders who have worked at frontier labs have unique insight. They know what works at scale, and they know exactly what is being left on the table internally. That's where the opportunity lies."

Investors are making a portfolio bet. If large language model scaling continues to deliver improvements, the Big Tech incumbents will likely win. But if the next breakthrough requires a fundamentally different approach — reinforcement learning without human data, world models, specialized vertical applications, or something entirely unexpected — the startups founded by researchers who recognized this first will be enormously valuable.

The $18.8 billion invested in 2026 isn't a bet on any single startup. It's a hedge against the possibility that the current AI paradigm has limits that its creators are best positioned to see.

Implications for the AI Ecosystem

This talent migration will reshape the AI industry in several predictable ways:

Accelerated Research Diversification: With hundreds of millions in funding, these startups can pursue research directions that Big Tech has deprioritized. Even if most fail, the survivors may produce genuine breakthroughs that advance the field.

Rising Acquisition Prices: Big Tech companies that want to reclaim this talent will need to pay premium prices. Google invested in Ineffable Intelligence despite it being founded by a former DeepMind researcher — a sign that the old taboos against funding poached talent are breaking down.

Increasing Competition for Compute: These well-funded startups will join the already fierce competition for GPUs and AI training infrastructure. Nvidia's participation in multiple rounds suggests the chipmaker sees these startups as important future customers.

Regulatory Scrutiny: As these companies grow, they'll face the same regulatory pressures as their Big Tech counterparts. The EU AI Act, potential US AI legislation, and emerging global frameworks will apply equally to startups and incumbents.

A Changing Career Calculus for Researchers: The coconut round phenomenon creates a new career path for AI researchers. Rather than choosing between academia and Big Tech, they can now choose startup founding as a legitimate third option — with funding available at unprecedented levels.

The Risk of Hype

It's important to maintain perspective. Not all of these startups will succeed. Many will fail to deliver on their ambitious technical goals. Some will struggle to find product-market fit even if their research is sound. The $5.1 billion valuation for Ineffable Intelligence, a company with no product and no revenue, reflects optimism rather than proven value.

But the trend itself is real and significant. The people who built the current AI revolution are voting with their feet, and they're voting for something different than what their former employers are building. Whether that something turns out to be better, worse, or merely different remains to be seen.

What is clear is that the AI industry is entering a new phase — one characterized by fragmentation, diversification, and intense competition between fundamentally different visions of how intelligence can be built. The coconut rounds are planting seeds that will determine the shape of AI for the next decade.

And David Silver, for his part, has stated that any money he makes from Ineffable will go to "high-impact charities that save as many lives as possible." In an industry often criticized for its profit motives, that's a reminder that at least some of the people building the future still remember why it matters.

--