Chapter 2: Wall Street's Machines

To understand AI trading, start where the money actually is: the institutions. A handful of firms run the most sophisticated prediction machinery ever built, and their results explain both what AI can do in markets and why you can't buy it.

A note on numbers before we begin. These firms are famously secretive. The figures below come from public reporting and industry analysis, not audited statements the firms published for your benefit. Treat them as directionally true rather than precise. The pattern matters more than any single number.

Renaissance Technologies: the thirty-year head start

Renaissance Technologies, founded by mathematician Jim Simons, runs the most successful investment fund in history: the Medallion fund, available only to employees. Across its lifetime, Medallion has compounded at roughly 66% a year before taxes, about 39% after. Its Sharpe ratios, a measure of return per unit of risk we will unpack in Chapter 13, have ranged from 2.0 to an almost absurd 7.5.

Forget the number for a moment and look at the method. Renaissance has used machine learning for roughly thirty years, back when "AI" was a lab word, not a marketing word. And it made a structural choice most competitors didn't. Instead of letting portfolio managers run competing strategies, Renaissance unifies everything into a single model. One system scores every stock by desirability, then a risk-reduction phase combines the picks into minimum-risk portfolios. Timeframes run from seconds to days. Because it is one model, the system can exploit correlations across assets that a collection of separate strategies would miss, including derivatives mispricings.

The lesson isn't "copy Renaissance." You can't. The fund is closed to outsiders, and the firm manages around $92 billion total. The lesson is what the edge actually consists of: decades of cleaned data, a unified architecture, and talent density. None of that ships in a $99-a-month app.

Citadel: the talent arms race

Citadel manages roughly $76 billion across equities, fixed income, and commodities. In 2026, it did something that tells you where this industry is headed: it expanded a quant team already around 180 people and began recruiting AI researchers directly from Google DeepMind, OpenAI, and Anthropic.

Sit with that. Hedge funds now compete with Silicon Valley for machine-learning talent, at Silicon Valley prices. When the best AI researchers on earth can choose between building chatbots and building trading systems, a meaningful number choose trading systems, because the compensation reflects the value of the edge.

Citadel's Tactical Trading fund returned about 24.7% in the first eight months of 2026, roughly 20% annualized since its 2008 launch. The firm's head of global quantitative strategies has said publicly that AI is reshaping what researchers do rather than eliminating them. The emphasis shifts toward identifying opportunities and directing automated systems. The human doesn't disappear. The human moves up the stack.

Two Sigma and D.E. Shaw: compute as a strategy

Two Sigma manages $60–70 billion and runs more than 100,000 simulations on market data every single day. Its toolkit runs from humble ridge regressions to natural-language processing, ingesting structured and unstructured data alike: prices, economic indicators, social media, satellite imagery, credit-card flows. It even crowdsources signal discovery through competitions, turning the hunt for predictive patterns into a sport with prize money.

D.E. Shaw, another research-heavy giant, points machine learning one level up. Its models continuously analyze the performance of the firm's own strategies and predict which parameter adjustments will improve results, using historical and live data together. The AI isn't just trading. It's managing the trading systems. The firm's Oculus fund, a systematic macro strategy, returned roughly 28% in 2025.

Notice what compounds here. These firms don't just have better models. They have models that improve their models, running on compute budgets that would make a university blush.

Jane Street: the quiet empire

If Renaissance is the legend and Citadel is the talent magnet, Jane Street is the empire nobody talks about. The firm rarely speaks publicly and operates, by most accounts, like a math research lab that happens to trade.

The numbers, reported by Reuters and others: $39.6 billion in net trading revenue in 2025, and a record $16.1 billion in the first quarter of 2026 alone, with profits doubling to $10.3 billion. Roughly 10% of all US equities activity flows through the firm, and nearly a quarter of primary-market activity in US-listed ETFs. About 3,500 employees. Serious medium-frequency strategies, positions held for minutes to days, and, tellingly, equity stakes in Anthropic and CoreWeave. The market maker is buying into the AI labs.

Jane Street is the purest illustration of this book's thesis. The edge is systems, speed, and structure. Not stock tips. Not gut feel. Industrial-scale applied mathematics.

Virtu Financial: the public one

Virtu is useful because it's public, ticker VIRT, so we can actually see the financials. In the first quarter of 2025, market making was 76% of its adjusted net trading income, up nearly 40% year over year. Return on equity ran near 25%. The firm integrates AI and machine learning into predictive analytics, real-time decision-making, and execution optimization, feeding its systems alternative data including real-time news and social media. In the second quarter of 2025, AI-powered strategies drove a 53% quarter-over-quarter surge in market-making trading income.

Virtu shows what "AI trading" looks like as an ordinary business. Not magic. A relentless, measurable grind of better predictions, better execution, and better cost measurement, compounded at scale.

The 2025 scoreboard

For context, here is what a strong year looked like across the quant world in 2025, per industry roundups. Bridgewater's Pure Alpha II gained roughly 33–34%, its best year in the firm's fifty-year history. D.E. Shaw's Oculus added about 28%. Citadel's flagship Wellington fund gained around 10%. Two Sigma's equity strategies were up in the mid-teens. Millennium added about 10.5%. The average hedge fund made about 11%.

Two observations, the way a planner would make them. First, the dispersion is enormous. Manager selection matters more than "AI" as a category. Second, even the best years at the best firms come with drawdowns, lockups, and fees that would make a retail investor flinch. There is no version of this you can buy with a phone app.

What the machines teach us

Look past the mystique and the institutional playbook has four ingredients:

1. Data nobody else has, or the same data cleaned and structured better than anyone else can manage.
2. Compute at industrial scale, hundreds of thousands of simulations a day, models retraining continuously.
3. Talent density, the best applied mathematicians and ML researchers on earth, now hired away from AI labs.
4. Infrastructure, colocation, direct exchange feeds, hardware optimized down to the microsecond.

Every retail "AI trading" product you will ever see is missing at least three of these four. Usually all four. That doesn't make retail tools useless. Chapter 5 will show where they genuinely help. But it sets the frame: you are not buying a smaller version of Renaissance. You are buying something else entirely. The next two chapters explain exactly what the institutions have that you don't, the data edge and the speed game.