In 2025, Jane Street made $39.6 billion with about 3,500 people. JPMorgan, with 317,000, made about the same.

The top four quant firms (Jane Street, Citadel Securities, Hudson River Trading, XTX) pulled in over $65 billion last year. OpenAI, Anthropic, and xAI combined did under $20 billion.

Quant trading is the most successful AI deployment in history.

I think this is going to look obvious in retrospect.

The argument I want to make is sharper than "markets are a good AGI benchmark". It's that the architecture for AGI already exists. It has been running in Long Island and Manhattan since 1988. Renaissance built it. DE Shaw built it. Two Sigma built it. They never called it AGI architecture because they had no reason to. They were too busy compounding.

The first AGI lab is whoever ports that architecture to language models.

What every quant firm actually built

In 2019 my firm ran a stat-arb book on Indian banking stocks. When sector flow imbalanced past a threshold, the basket mean-reverted at 0.5 to 2 day horizons. The strategy worked for sixteen months. Then in March 2020 it inverted. Pandemic retail flow changed the composition of who was trading the basket. What had been mean-reverting became momentum. We caught the regime change in the second week. Most of the firms running similar books didn't catch it for three months.

The difference wasn't the model. The strategy itself was disposable. Every working quant strategy is disposable, on a timeline somewhere between a year and a decade. What mattered was the system around the strategy: the daily review of edge sources, the explicit hypothesis generation about what could break, the willingness to kill a working strategy on weak evidence rather than strong evidence. The strategies came and went. The machine that produced strategies got better.

This is what every quant firm that has ever mattered actually built. Not alpha. The system that generates alpha, retires alpha, generates new alpha, and improves the alpha-generation process itself.

Renaissance has been doing this since 1988. Their reported equity curve compounds at roughly 66% per year before fees, sustained across thirty years and every major market regime. The numbers are not independently audited and Renaissance has not officially confirmed them. They are widely cited in academic finance. UCLA's Brad Cornell concluded in a 2020 paper that the returns "stretched explanation to the limit". To Institutional Investor he was more direct: "like the sun rising in the west".

What does it mean for a fund to compound at 66% for thirty years across regimes its strategies were not trained on? It means the firm is not pattern-matching. It is reasoning. The strategies are different by 1995 than they were in 1988, different by 2008 than 1995, different now than 2008. The firm that produces them has been compounding capability against an environment that fights back.

Renaissance solved a version of the AGI problem. They just solved it for one cognitive primitive (forecasting under uncertainty against adversaries) on one substrate (financial markets) using human researchers as the meta-learner. The architecture is the answer. The substrate is the answer. Only the meta-learner is wrong.

The architecture has a name

Recursive self-improvement. The system that improves the system that produces capability, faster than the capability it produces saturates.

I.J. Good described it in 1965. Yudkowsky formalized it. Aschenbrenner forecast it in Situational Awareness. The CEOs of Anthropic and DeepMind discussed it openly at the World Economic Forum in 2026. Both labs have publicly stated they are spending hundreds of millions on RL environments specifically to enable it. ICLR 2026 hosted a workshop dedicated to it.

The frontier AI labs are racing toward recursive self-improvement using synthetic environments built inside research labs. Coding gyms. Math problem sets. Hand-built RL setups where verifier models grade verifier models. These environments have a fundamental problem: they are constructed by humans who have to decide, in advance, what the system should optimize against. When the system gets good enough, it games the construction. The proxy decouples from reality. Nobody knows when this happens until it does.

Markets do not have this problem. The reward signal is not constructed. It is generated by every other participant pricing reality in real time. The signal cannot be Goodharted at the substrate level (it can be gamed at the margins; that's a different problem) because the participant set adapts faster than any single optimizer can stabilize an exploit. The substrate has been adversarially robust for centuries. It will continue to be adversarially robust regardless of what we train on it.

The frontier AI labs are building the meta-learner. They are missing the substrate. The quant firms have the substrate. They are missing the meta-learner. The first AGI lab is whoever assembles both.

Hassabis already knew

In 2016, according to Sebastian Mallaby's The Infinity Machine, Demis Hassabis assembled a secret hedge fund operation inside DeepMind. Around twenty researchers. High-frequency trading algorithms. An exploratory partnership with BlackRock. Mallaby reports, sourced to a person familiar with the project, that Hassabis's ambition was to compete with Renaissance Technologies.

Google killed it.

Hassabis had spent a decade picking environments and letting agents extract everything they could teach. AlphaGo. AlphaZero. AlphaStar. By 2016 those environments were exhausted. The agents had won at superhuman levels. There was nowhere obvious left to go.

He picked markets. He did this while running the world's leading AI lab. That tells you what he thought came next.

DeepSeek already shipped

In 2024 DeepSeek released an open-weight LLM that broke the AI industry's pricing assumptions and triggered a multi-trillion-dollar selloff in US tech stocks.

DeepSeek's parent company is High-Flyer, a Hangzhou quant fund. High-Flyer's first deep-learning trades went live in 2016. By 2017 most of its trading was AI-driven. They built roughly $14 billion in AUM training models on tick data. In 2023 they pivoted research toward "pursuing AGI" using the same compute stack.

This is the prototype, running in the wild. A quant firm that already knew how to run big GPU clusters, clean adversarial data, and run tight feedback loops at scale. Pointing that capability at next-token prediction instead of next-price prediction.

The institutional machine is converging from both sides. Sam Bankman-Fried, a Jane Street alum, provided some of Anthropic's initial funding. Jane Street has participated in subsequent rounds. Scale AI's founder came from Hudson River Trading. Surge AI's founder came from a Thiel hedge fund. OpenAI's research head spent time as a quant. HRT's AI head came from DeepMind. xAI publicly tested its frontier model on live stock markets in 2025; Musk tweeted about using it to "pay for all those GPUs".

An FT Alphaville piece in January described quants and AI labs as "converging on the same institutional machine: large-scale learning systems attached to balance sheets".

The convergence is real. The endpoint is whoever ports the architecture first.

"Why hasn't Renaissance built AGI?"

The natural objection. If Renaissance already has the architecture, why do they not have AGI?

Because the architecture is necessary, not sufficient. Renaissance's meta-learner is human researchers. Humans are slow. They generalize narrowly. They can only hold so many regimes in attention at once. Renaissance compounds capability against markets at the rate human research teams can compound, which is fast enough to produce the most successful trading firm in history but nowhere near fast enough to produce general intelligence.

The other objection: Renaissance's strategies are narrow. Statistical edge-stacking, not reasoning. This is true and not load-bearing. Narrow strategies are what their meta-learner produces because their meta-learner is optimizing for capital efficiency on liquid markets. A different meta-learner, optimizing for capability gain rather than capital efficiency, would produce different strategies on the same substrate.

What changes when the meta-learner is a frontier language model? Three things.

The meta-learner runs continuously instead of weekly. Decisions that humans make in research meetings happen in milliseconds. Hypothesis generation parallelizes across thousands of streams.

The meta-learner generalizes more broadly. A human researcher who learns something about Indian banking stat-arb has to consciously port that lesson to other domains. A frontier model with the architecture in its weights ports automatically.

The meta-learner improves itself. Renaissance's researchers got better over thirty years through hiring, training, and culture. A model-based meta-learner gets better through gradient updates against the same substrate that produced the alpha. The improvement loop is closed.

This is the missing piece. Renaissance has been running the architecture with a slow meta-learner. The frontier AI labs have been building fast meta-learners against the wrong substrate. The first lab to put a fast meta-learner inside the right substrate is the first AGI lab.

Why frontier models can't do this yet

Current frontier models are trained for chat. The training objective shapes the actor. A model optimized for human approval on chatbot leaderboards is, by construction, the wrong shape for an adversarial environment with delayed feedback and skin in the game.

This is fixable. The model architectures are right. The pretraining is right. The reasoning capabilities are emerging. What's missing is the training scaffold: the environments, the reward signals, the data, the meta-learning loops that turn a chatbot into an alpha-generation system.

Building that scaffold is engineering, not basic research. The +0.42 Sharpe gap between current frontier models and human experts on Indian equity data is not a fundamental limit. It is a capability that hasn't been trained yet. The lab that trains it first is the lab that wins.

What the first AGI lab actually looks like

Picture the institution.

100 humans. 10,000 agents. A vertical equity curve. No public-facing product. No API. Few or no papers. Compute bills measured in nine figures and growing. Payroll measured in eight figures and stable. The company is not raising capital because it does not need outside money. It is closer in shape to Renaissance than to OpenAI.

The agents do most of the work. They generate hypotheses, design experiments, run simulations, deploy strategies, monitor live performance, retire decaying edges, and propose architectural changes to themselves. The humans pick problems, design the meta-environment, audit decisions the agents flag as uncertain, and intervene when something looks wrong.

The agents trade against the world's deepest adversarial substrate. The substrate generates the reward signal. The reward signal funds more compute. More compute trains more capable agents. More capable agents extract more reward. The loop closes.

This is recursive self-improvement, on the cleanest substrate that exists, with an economic flywheel attached. Every other AI lab is funding-bottlenecked by external capital. This lab is funded by the substrate it is training against.

It does not look like an AI lab from the outside. It looks like a quant firm.

Three paths

Three candidates for who builds this.

A frontier AI lab. OpenAI, Anthropic, Google DeepMind, xAI. They have the meta-learner. They lack the substrate, the trading scaffolding, and the corporate freedom to pursue this seriously. Hassabis's 2016 attempt is the cautionary tale. The problem is permission.

An existing quant firm. Renaissance, Citadel, Two Sigma, DE Shaw. They have the substrate and the scaffolding. They lack the meta-learner architecture and the research culture to produce it. They could buy a frontier model off the shelf, but it would be optimized for chat, not for their substrate. The problem is research culture.

A new entity built specifically for this. No models, no scaffolding to start. Structural freedom to build both. No corporate parent to veto. No legacy research culture to fight. The problem is that nobody has done it yet, which means the playbook is unwritten.

I'd bet on the third. The first is the least likely, despite having the most resources, for the same reason Hassabis got stopped in 2016.

The prediction

By 2030, the highest-revenue AI company in the world will not call itself an AI company.

This is already partly true. In 2025, Jane Street alone earned more revenue than every frontier AI lab combined. The top four quant firms together earned more than the entire AI industry. The gap is not closing.

If I am wrong, it will look like this: the most valuable AI company in 2030 is OpenAI or Anthropic or DeepMind, and their primary revenue still comes from licensing models or API access. That is the version of the future I am betting against.

The capacity wall is real. Medallion is capped at $10-15 billion because beyond that, its own size moves the market against itself. The first AGI lab will hit a similar wall. But the wall is higher than it sounds. A fund at Medallion's capacity, compounding at 30% net, produces enough capital to fund any research agenda you want. You don't need to buy the world. You need enough to never need outside money again.

The team that solves general reasoning is the team that can deploy it in the highest-leverage environment available. The highest-leverage environment is markets. The architecture for compounding capability against that environment was invented in 1988 in Long Island and is still running.

Whoever ports it to language models first won't tell you. They will just compound.

When we look back, I suspect the moment AGI arrived will not have been marked by a paper.

It will have been a pattern in the tape. A silent, vertical equity curve.