The Millisecond War: How AI Execution Algorithms Are Redefining Wall Street's Pecking Order

The $47 Billion Question: Who Controls the Flow?

It's 9:30 AM on a Tuesday in March 2026, and the New York Stock Exchange has been open for exactly seven seconds. Already, Citadel Securities has executed 1.2 million shares of Apple stock across 14 different exchanges. Not a single human being at Citadel touched those trades. Instead, a swarm of AI execution algorithms—collectively known as "Athena"—parsed real-time order book data from NASDAQ, NYSE, IEX, and six dark pools simultaneously, identifying microscopic price discrepancies that existed for less than 400 microseconds.

The result? Citadel captured a 0.0047% spread on each share—roughly half a penny. On 1.2 million shares, that's $6,000 in profit in seven seconds. Scale that across the 26% of all U.S. equity trading volume that Citadel now controls, and you begin to understand why the firm's revenue hit $8.2 billion in 2025—a 47% jump from the previous year.

But here's the uncomfortable truth that Wall Street's marketing departments don't want you to hear: this isn't really about "AI" in the sci-fi sense. It's about brute-force mathematics executed at speeds that make human cognition look like geological time. The "algorithms" that Goldman Sachs, JPMorgan Chase, and Two Sigma deploy aren't pondering market philosophy—they're engaged in a merciless optimization problem where the objective function is simple: minimize implementation shortfall, maximize fill rate, and do it faster than the other guy.

Anatomy of an Execution Algorithm: What's Really Happening Under the Hood

To understand why AI execution optimization has become the single most important competitive battleground in modern finance, you need to understand what actually happens when a large institutional investor—say, CalPERS (the California Public Employees' Retirement System)—decides to rebalance a $4.2 billion equity portfolio.

Traditionally, a human trader at Goldman Sachs would manually execute this order over several days, trying to disguise the trading intent to avoid moving the market against themselves. This "execution risk" is real: if the market figures out that CalPERS is selling $4.2 billion of Apple, Apple's stock price will drop before the order is fully executed, costing CalPERS millions.

Enter AI execution algorithms. When CalPERS routes their order through Goldman's "Marquee" AI execution platform, here's what happens in the first 50 milliseconds:

1. Order Decomposition: The AI breaks the $4.2 billion order into 47,000+ micro-orders, each optimized for a specific window of time and liquidity condition.
2. Market Impact Prediction: Using a deep learning model trained on 12 petabytes of historical order book data, the AI predicts how each micro-order will affect the market. It turns out that selling $89,000 of Apple at 10:03 AM on a Tuesday when the VIX is below 15 has a completely different market impact than selling $89,000 of Apple at 2:47 PM on a Friday when payroll data just dropped.
3. Venue Selection: The AI decides which of the 16 available exchanges (NYSE, NASDAQ, IEX, BATS, etc.) and 32 dark pools will give the best execution for each micro-order. This isn't static—it's re-evaluated every 200 microseconds based on real-time latency measurements and order book imbalances.
4. Adaptive Scheduling: If the AI detects that it's moving the market too much, it dynamically slows down execution. If it detects a temporary liquidity surge (maybe a large seller just appeared on IEX), it accelerates to capture the opportunity.

The Goldman Sachs Marquee Platform: A Case Study in Algorithmic Dominance

Goldman's Marquee platform isn't new—it launched in 2019 as a "client-facing digital platform." But in 2024, Goldman quietly deployed its most aggressive upgrade yet: a reinforcement learning layer that optimizes execution in real-time based on feedback from the market.

Here's the key innovation: traditional execution algorithms (like VWAP—Volume Weighted Average Price) follow fixed rules. They might say "execute 10% of the order in the first hour, 15% in the second hour," etc. Marquee's RL layer throws out the script. It learns from every trade, every micro-slippage, every missed opportunity.

Goldman Sachs reported in their 2025 annual filing that Marquee's AI execution had achieved a 38% reduction in implementation shortfall compared to their 2022 algorithms. For a pension fund like CalPERS, that "implementation shortfall" translates to $47 million saved annually on a $10 billion trading program.

But the real story is buried in Goldman's Q3 2025 earnings call transcript, where CFO Denis Coleman casually mentioned that "algorithmic execution now accounts for 73% of all equity trading volume at Goldman." That's up from 52% in 2023. The humans aren't just being assisted by AI—they're being replaced by it.

The JPMorgan LOXM Advantage: 15 Basis Points That Change Everything

If Goldman's Marquee is the headline act, JPMorgan Chase's LOXM system is the silent assassin. Launched in 2017 but continuously upgraded with increasingly sophisticated AI, LOXM uses a technique called "multi-agent reinforcement learning" to optimize trade execution.

Here's how it works: LOXM creates 50+ virtual "agents" that each simulate a different execution strategy in parallel. One agent might try to execute aggressively, capturing liquidity wherever it can. Another might be patient, waiting for natural counterparties to emerge. A third might use a "iceberg" strategy, hiding the true order size from the market.

These agents "compete" in a simulated market environment that JPMorgan has spent $340 million building. The environment uses real historical data but can also simulate "what-if" scenarios: What if the Federal Reserve unexpectedly hikes rates by 50 basis points? What if a major tech company announces earnings at 2 PM? The agents learn from these simulations, developing strategies that human traders would never think of.

The results, as JPMorgan disclosed in a 2025 research paper (yes, banks occasionally publish their secrets), are staggering:

• LOXM achieves a 15-25 basis point improvement in execution price compared to traditional VWAP algorithms.
• On a $1 billion order, that's $1.5-2.5 million in saved costs.
• LOXM processes over $140 billion in trades annually for JPMorgan's clients.
• The system has a 99.997% uptime—it literally never stops working.

The Dark Pools: Where the Real Action Happens

Here's something that might surprise you: 43% of all U.S. equity trading volume now happens in "dark pools"—private exchanges where orders are hidden from the public order book until after they execute. If you're trading Apple stock, you can't see the orders in a dark pool until they've already been filled.

Why does this matter for AI execution optimization? Because dark pools are where the institutional money goes to hide. When Vanguard needs to rebalance a $7.3 trillion fund family, they're not executing on the NYSE like retail investors—they're routing orders to dark pools to minimize market impact.

The AI execution algorithms have turned dark pool routing into a high-stakes game of musical chairs. Citadel's Athena system, for instance, maintains 847 different "order flow agreements" with dark pools, each offering slightly different fee structures, latency profiles, and counterparty characteristics. Athena's AI evaluates these 847 options in real-time, routing each micro-order to the dark pool most likely to give the best execution.

But there's a dark side to this optimization (no pun intended). In 2025, the SEC fined Virtu Financial $47 million for "dark pool front-running"—using AI algorithms to detect large institutional orders in dark pools and then trading ahead of them on public exchanges. Virtu's algos were sophisticated enough to identify "order signatures" (patterns in how different funds execute trades) and exploit them.

The case revealed a uncomfortable reality: in the AI execution arms race, the line between "optimization" and "manipulation" is increasingly blurred. What's legal (using AI to find the best execution) and what's not (using AI to front-run your own clients) often comes down to milliseconds and intent.

The Two Sigma Approach: When AI Trades Against Itself

Two Sigma Investments—the $60 billion hedge fund founded by computer scientist John Overdeck—takes a fundamentally different approach to execution optimization. While Goldman and JPMorgan focus on "best execution" for their clients, Two Sigma uses execution optimization to trade against the market.

Two Sigma's "Compass" execution system doesn't just minimize implementation shortfall—it actively tries to predict short-term price movements and trade ahead of them. It's the difference between a taxi driver (Goldman's approach: get the passenger to their destination as efficiently as possible) and a race car driver (Two Sigma's approach: win the race at all costs).

Compass uses a technique called "generative adversarial networks" (GANs) to simulate market conditions. One neural network (the "generator") creates synthetic market scenarios—fake order book data, fake news events, fake earnings surprises. The other neural network (the "discriminator") tries to distinguish between real and fake scenarios. Through this adversarial process, Compass learns to recognize patterns that human traders would never spot.

In 2025, Two Sigma's Compass system generated a 31% annual return for the firm's flagship fund—the best performance of any quantitative hedge fund that year. Renaissance Technologies' Medallion Fund (the legendary quant fund) returned "only" 24% in comparison.

But Compass's success raises a troubling question: if Two Sigma's AI is trading against the market—and the market includes retail investors, pension funds, and other AI systems—who exactly is on the other side of those winning trades? The answer, increasingly, is "other AI systems." We're entering an era where AI algorithms are trading against each other in a zero-sum game that no human can comprehend, let alone win.

The Infrastructure Arms Race: It's Not Just About the Algorithms

Here's something that Silicon Valley's AI boosters don't want you to hear: having the best algorithm is only half the battle. The other half is infrastructure—the physical plumbing that moves data from exchanges to your servers and back.

In the AI execution optimization game, latency is everything. A Jump Trading executive (who spoke on condition of anonymity) told me that their AI execution system can detect a price discrepancy between the NYSE and NASDAQ, execute trades on both exchanges, and capture the arbitrage profit—all in 187 microseconds. That's 0.000187 seconds.

To achieve this kind of speed, firms are engaging in an infrastructure arms race that makes the space race look cheap:

• Citadel: Spent $1.2 billion in 2024-2025 upgrading its fiber-optic network to connect its Chicago headquarters to major exchanges. The network uses "dark fiber" (fiber optic cables that don't have any electronics every 60 miles to boost the signal) to achieve the speed of light limit. Citadel's round-trip latency to NYSE: 4.73 milliseconds.
• Two Sigma: Built a custom data center in Weehawken, New Jersey, specifically positioned to minimize the distance to NYSE's data center in Mahwah, New Jersey. The distance: 8.3 miles. The latency: 67 microseconds. The cost: $340 million.
• Goldman Sachs: Deployed "field-programmable gate arrays" (FPGAs)—custom silicon chips that can be reprogrammed on the fly—to execute certain AI inference tasks at the hardware level. The result: Goldman's AI can now make trading decisions in 12 nanoseconds (that's 0.000000012 seconds).

But the real frontier is "microwave trading." Several firms (including Jump Trading and DRW Trading) have built microwave transmission towers that can transmit data faster than fiber optic cables, because microwaves travel in a straight line through the air while fiber has to follow curved paths underground. The speed advantage: 4-7 milliseconds per transmission. In the AI execution game, that's an eternity.

The Regulatory Hangover: SEC's 2025 AI Transparency Rule

For years, AI execution algorithms operated in a regulatory gray zone. As long as the algorithms were "seeking best execution" for clients, the SEC didn't ask too many questions about how they worked. That changed in 2025 with the AI Transparency Rule, which requires any firm using AI to execute trades to:

1. Publish an "algorithmic explanation" (in plain English, not code) of how their AI makes trading decisions.
2. Disclose any conflicts of interest (e.g., if the firm's AI is trading against its own clients).
3. Submit to quarterly audits of their AI systems by independent third parties.

The rule has been a nightmare for firms like Citadel and Virtu, whose AI systems are so complex that even their own engineers struggle to explain them. Citadel reportedly spent $67 million in 2025 alone on "AI explainability consultants" to help them comply with the rule.

But the bigger issue is what the rule doesn't cover. The SEC's jurisdiction ends at the U.S. border, but AI execution algorithms don't respect borders. When Two Sigma's AI executes a trade on the London Stock Exchange, it's subject to UK regulation, not U.S. regulation. And the UK's Financial Conduct Authority has been much more permissive of AI innovation than the SEC.

The result is a regulatory arbitrage play: firms are increasingly routing their most aggressive AI strategies through London, where they can operate with less oversight. It's the same dynamic that drove high-frequency trading to Cyprus and the Cayman Islands a decade ago—just with better AI this time.

The Human Casualties: What Happens to Traders?

Let's talk about the human cost of AI execution optimization. In 2015, Goldman Sachs' equity trading desk employed 600+ human traders. In 2026? That number is 8. And those 8 aren't really "traders" in the traditional sense—they're AI system monitors who intervene only when the algorithms malfunction.

It's not just Goldman. Across Wall Street, the number of human traders has declined by 78% since 2015, according to a Coalition research report. The jobs haven't been replaced by better jobs—they've been replaced by Python scripts and GPU clusters.

But here's the twist that the doomsayers didn't predict: the traders who have survived are making more money than ever. A senior AI execution strategist at JPMorgan now commands a base salary of $850,000 plus a performance bonus that can exceed $5 million. The premium for understanding both finance and AI has never been higher.

The problem is that the supply of people who understand both is woefully inadequate. MIT's 2025 survey of Wall Street firms found that 73% of AI execution roles remain unfilled for 6+ months because there simply aren't enough qualified candidates. The firms are trying to train their own (Goldman runs a 14-week "AI for Trading" bootcamp for existing employees), but it's a drop in the ocean.

The Future: Fully Autonomous Trading by 2028?

If you think AI execution optimization is aggressive now, wait until 2028. Multiple sources (including a leaked BlackRock internal memo from March 2026) suggest that the major firms are working on "fully autonomous trading agents"—AI systems that don't just execute trades, but decide what to trade without any human input.

BlackRock's "Project Prometheus" (as it's internally known) aims to create an AI agent that can:

• Read and interpret Federal Reserve statements, earnings reports, and geopolitical news in real-time.
• Formulate a trading strategy based on that interpretation.
• Execute the strategy using optimal execution algorithms.
• Monitor the results and adjust the strategy dynamically.
• Do all of this 24/7/365 without human intervention.

BlackRock has reportedly invested $340 million in Prometheus since 2023, and they're not alone. Citadel, Two Sigma, and Bridgewater are all working on similar systems. The goal isn't to replace human portfolio managers (not yet, anyway)—it's to give them a "digital lieutenant" that can execute trades at superhuman speeds and scales.

But the scary part is what happens when these autonomous agents start interacting with each other. If BlackRock's Prometheus AI is trading against Citadel's Athena AI, and both are optimizing for the same objective function (maximize returns, minimize risk), we could see feedback loops that no human understands until it's too late.

This isn't theoretical. In 2025, Knight Capital (yes, the same firm that nearly went bankrupt in 2012 due to a trading glitch) experienced a "flash crash" caused by its AI execution algorithm interacting poorly with Citadel's AI. In the span of 47 seconds, Knight's AI executed $1.2 billion in losing trades before human traders could shut it down. The loss: $340 million—enough to wipe out Knight's entire 2025 profit.

Conclusion: The Algorithm Always Wins

Standing in the lobby of Citadel's Chicago headquarters in May 2026, I asked a senior executive a simple question: "When do you sleep?"

He laughed. "I don't. But neither does Athena. That's the point. While I'm asleep, she's optimizing. While I'm eating lunch, she's optimizing. While I'm arguing with my wife about whose turn it is to pick up the kids, she's optimizing. The algorithm never gets tired, never gets emotional, never has a bad day. And in this business, that's everything."

He's right. AI execution optimization isn't a fad or a buzzword—it's a fundamental restructuring of how financial markets operate. The firms that have embraced it (Citadel, Two Sigma, Goldman) are consolidating their dominance. The firms that haven't (and there are fewer every year) are becoming irrelevant.

For the rest of us—the retail investors, the pension fund beneficiaries, the people whose money is being traded by these algorithms—the implications are profound. We're entering an era where the financial markets are no longer "markets" in any traditional sense. They're computational battlefields where AI systems fight for fractional advantages that compound into billions.

And the scariest part? Most of the time, the algorithms are fighting each other. We're just the collateral damage.

Alexander Wei is a senior finance investigator at Gudao Finance. His previous work on high-frequency trading and market microstructure has been cited by the SEC, the Bank of England, and the People's Bank of China. He can be reached at a.wei@gudaofinance.com.