The rise of the neolabs: when intelligence is just good enough

On record seed rounds, open models, and why most neolabs are not in scope for ellipsis. Each quarter we take one question, from our investors or from our own work, and answer it properly. This quarter: the new AI labs.

In 2015, OpenAI bet an entire company on a single idea: that scaling up neural networks would keep making them smarter, the idea its researchers later formalized as the scaling laws. Google had the researchers, the chips, and the papers, but it was hedged across a dozen research programs and a search business to protect. Focus won. Six years later, Anthropic narrowed the bet further, to reinforcement learning on code sold to enterprises, and is winning that game as we write: it now leads enterprise LLM usage overall, and in code generation it holds twice OpenAI's market share (42% versus 21%).

Today, more than sixty new AI labs each claim to be the third act of this story. They are called neolabs: research companies founded by scientists leaving OpenAI, DeepMind, and Anthropic. Each is focused on a different bet:

• Reinforcement learning: models that learn by trial and error instead of reading the internet. Example: Ineffable Intelligence, the new lab of AlphaGo research lead David Silver.
• World models: AI that learns physics and 3D space from video and sensors, not just text. Examples: World Labs (Fei-Fei Li) and AMI Labs (Yann LeCun).
• Continual learning: models that keep learning on the job instead of being frozen after training. Examples: Core Automation (ex-OpenAI), Adaption Labs (ex-Cohere).
• Formal mathematics: reasoning that can be mechanically proven correct. Examples: Axiom Math, Harmonic.
• Diffusion language models: a faster way to generate text, borrowed from image generation. Example: Inception Labs.
• AI for science: labs that run their own physical experiments to create training data. Example: Periodic Labs.

And a newest cohort is arriving right now, including Mirendil (from Anthropic's reasoning team), Sooth Labs (ex-Meta), and Standard Intelligence (computer use). Radical Ventures' market map, "The Rise of the NeoLab" (June 2026), lists them all.

Funded on a promise

These labs are funded on a promise: that frontier intelligence will keep being paid for. So far it is being paid: Anthropic grew from about $1 billion in revenue run-rate at the end of 2024 to roughly $47 billion today, the steepest ramp in software history. The lesson investors took from OpenAI and Anthropic is that entry price barely mattered: Anthropic's valuation rose from $60 billion to $183 billion within six months of 2025, and even late capital earned venture-scale returns.

As a result, multistage firms and strategic investors now fund neolabs at record prices, and even early-stage funds are abandoning their ownership targets to participate. The largest seed rounds in venture history are now all neolabs: Thinking Machines ($2 billion at a $12 billion valuation, before any product), Humans& ($480 million). Safe Superintelligence is valued at $32 billion with no product, no API, and no published research. These are celebrity outliers, far from our end of the market. But the pricing cascades down to ordinary teams: across all US seed rounds on Carta, the 95th percentile valuation nearly tripled to $200 million in the year to Q2 2026, the fastest rise in a decade. Much of this capital is circular: NVIDIA has joined 32 of the 60-plus neolab financing rounds, more than any venture firm, and much of that money flows back as compute purchases.

Why we passed

Over the past couple of months, we looked at several of these opportunities ourselves and decided not to participate. The teams were exceptional, but the deal dynamics (round size and valuation) do not suit a pre-seed focused fund. More fundamentally, we believe the model layer for general intelligence will commoditize. Five forces point that way:

1. Producing the frontier gets more expensive every year. The cost of training frontier models has grown roughly 2.4x per year since 2016, and the largest training runs are approaching $1 billion. Compute is not the only bill: labs spend an additional 10 to 20% of their GPU budgets on training data, increasingly synthetic data and machine-built training environments, a market that has already made four data vendors billion-dollar businesses. No lab has yet funded a new frontier generation out of its own cash flow; the money comes from capital markets. And money alone is no longer enough: GPU capacity is sold out years ahead, so allocation matters as much as capital. This is why NVIDIA sits on so many cap tables, and why the leading labs are going full-stack, designing their own chips and building their own data centers to escape the queue.
2. Owning yesterday's frontier is worth less every month. The price of a fixed level of intelligence is collapsing: GPT-4-level performance cost $37.50 per million tokens in 2023, and about $0.18 two years later. The endpoint of this curve is free: Meta's Muse Glimmer, a 30-billion-parameter open-weight model released in August 2026 under Apache 2.0, runs always-on agent workflows entirely on a single consumer GPU, with math and coding scores that were frontier-only territory a year ago. Distillation, open-weight models, and competition make this decay permanent, not cyclical.
3. Fewer and fewer paid tasks need the frontier. Most enterprise workloads are already served by good-enough models below the frontier: drafting, summarizing, classifying, routine agent steps. And even inside the premium use case, coding, where model tokens are 70 to 80% of product revenue, buyers are routing work down. Coinbase, for example, defaulted its engineers to the open models GLM 5.2 and Kimi 2.7 at roughly one fifth of frontier pricing, cut its AI spend by about half, and still saw token usage hit record highs. Its CEO's summary: frontier models "for planning, but not for execution, where they can be overkill."
4. Models are becoming interchangeable. Coinbase's savings came not from a one-time migration but from a router that picks a model for every request, by complexity and price. Such routers are now being built everywhere: inside enterprises, inside applications, and as neutral platforms. At the same time, enterprises deliberately keep their context, their data, and their workflows outside the model, so the model stays a plug. A market where every buyer runs a router is a market where every model is a bid. The labs can see this too: they are racing to build coding agents, harnesses, and applications on top of their models, lock-in above the model layer, because they know the model alone will not hold it.
5. Specialization does not save the challengers. The neolabs' best defense is the innovator's dilemma: the incumbents earn their revenue on LLMs and coding, so orthogonal research directions get little of their attention. But the bitter lesson of AI keeps repeating: general models trained at scale beat specialized ones.

There is a counterargument, and it ships every few months: a new frontier release that resets the premium. Fable 5 is the latest, the most expensive model Anthropic has ever listed, with demand rationed at launch all the same, because measured per solved task the expensive model is often the cheap one. But that is exactly the problem. The premium is a property of the release, not of the model.

What we invest in instead

We therefore continue on our "commoditized magic" thesis: we invest in AI unlockers, large frontier applications with a strong "why now", products that were not possible before AI, built behind moats that do not commoditize:

• Data moat: the problem requires unique proprietary data that is hard to generate, synthesize, or scrape, with a clear data flywheel that increases value and stickiness over time.
• Expertise moat: the problem requires unique expertise that is hard to find, define, or simulate.
• Compute moat: the problem is complex and high-value enough to justify expensive, sophisticated intelligence.

The full checklist we apply is on the investment criteria page.