XTX Markets Built a £16 Billion AI Trading Empire With No Human Traders

When the Sunday Times published its 2026 Rich List in May, Alex Gerko's name appeared at number eight — up from twentieth the year before. His fortune had grown from £8.7 billion to £16 billion in a single year, one of the largest single-year wealth gains on the list. The publication noted that he holds a mathematics doctorate from Moscow State University and founded a quantitative trading firm in 2015. What the Rich List could not easily convey was why this particular company, XTX Markets, had become one of Britain's most profitable private enterprises — or how a workforce of fewer than 300 people running no trading desks had achieved it.

The answer, developed over a decade of building and iterating on deep-learning market-making systems, is that XTX has done something the financial industry spent decades treating as implausible: it has removed human judgment entirely from the trading process and replaced it with an AI system that executes better, faster, and at lower cost than any team of professional traders has managed. XTX now processes more than $250 billion in trading volume every day, while its combined net profit across its UK entities reached £1.71 billion on £3.93 billion in revenue for 2025.

The scale of that result — and what it says about where algorithmic trading is heading — deserves considerably more examination than a wealth ranking provides.

From Moscow Mathematics to Zero Human Traders

Gerko's path to founding XTX runs through a series of institutions that shaped his particular conviction about markets. After completing his undergraduate degree in mechanics and mathematics at Moscow State University, he earned a doctorate in mathematics before joining Deutsche Bank and later GSA Capital, a London-based systematic hedge fund. GSA was an early adopter of quantitative approaches to trading, and it was there that Gerko worked on the statistical models that would eventually underpin XTX.

What he reportedly concluded during those years was not merely that computers could trade faster than humans — that observation had been obvious in high-frequency trading since the early 2000s — but that human traders introduce systematic errors that cannot be corrected through supervision or experience. Emotional responses to drawdowns, anchoring to prior price levels, an inability to process thousands of correlated variables simultaneously: these are not individual failings but structural properties of human cognition that any market-making operation built around people must accommodate.

XTX, incorporated on January 30, 2015 as a spin-off from GSA Capital, was designed from the start around the opposite assumption. There would be no trading desks. No senior trader would decide when to widen or narrow spreads. No risk manager would call for a position to be cut based on instinct. The model would do all of it.

Today, the majority of XTX's roughly 300 employees are mathematicians, cryptographers, machine-learning engineers, and infrastructure architects. Their role is not to trade. It is to write the mathematical models, debug the neural networks, and maintain the compute infrastructure that trades on their behalf.

How AI Market Making Actually Works at XTX's Scale

To understand why XTX's model performs as it does, it helps to be specific about what market making involves at a technical level.

A market maker's function is to continuously post both a bid price (what it will pay to buy an asset) and an offer price (what it will charge to sell it). The spread between these two prices is the market maker's theoretical revenue per transaction. In practice, however, the market maker also accumulates inventory: every time it sells, it becomes short; every time it buys, it holds a position. If the market moves against that inventory — if equities fall while the market maker holds a long position accumulated from customer orders, for instance — the inventory loss can wipe out the spread revenue many times over.

This inventory risk is the central problem of market making. Human traders manage it through a combination of hedging rules, experience-based judgment about when a market is in a trending regime versus a mean-reverting one, and the social information that flows through a trading floor. XTX manages it through statistical models trained on decades of market data and deployed on a GPU cluster that can evaluate thousands of variables simultaneously in microseconds.

The training dataset underlying that system is not a narrow collection of recent prices. According to infrastructure analyses that have examined XTX's disclosed specifications, the firm stores 650 petabytes of high-speed market data — every order book update, every tick, every cross-asset price movement that XTX has encountered or acquired over the course of the firm's operation and beyond. That corpus gives the models statistical grounding that no human trader's experience could replicate. A trader might have fifteen years of professional experience; XTX's system has been exposed to market structure across decades, geographies, and asset classes simultaneously.

The input signals that feed XTX's inference models reportedly go well beyond raw price data. Order-book depth at multiple price levels, the arrival rate of new orders, the short-term autocorrelation structure of price moves, sentiment proxies derived from alternative data, and the transient correlations between loosely related instruments are all candidates for the feature sets that drive deep market-making models at this scale. The challenge is not computing any one of these features but combining thousands of them in real time at a horizon short enough to matter for pricing decisions. GPU parallelism is the enabling technology: operations that would take seconds on CPU can be performed in microseconds when vectorized across thousands of cores simultaneously. XTX's choice to build its own cluster rather than rely on shared cloud infrastructure means it can allocate GPU resources between training and live inference without scheduling conflicts — a non-trivial advantage when the same hardware must serve both model development and production trading.

The inference side of the system operates in real time. When a counterparty requests a price, XTX's AI evaluates the current state of the order book, the trajectory of related assets, recent liquidity conditions, and the firm's current inventory — then produces a bid and offer spread calibrated to reflect all of those inputs. This happens in the sub-millisecond timeframe that electronic trading demands.

The GPU Infrastructure Behind Every Trade

Running this system requires compute infrastructure that most trading firms have never attempted to build. XTX has assembled what infrastructure analysts describe as one of the largest private GPU clusters in quantitative finance: more than 25,000 GPUs, including approximately 10,000 Nvidia A100s and 10,000 V100s, supported by 650 petabytes of high-speed storage. The cluster processes more than one trillion data points per day.

The strategic decision not to rely on cloud computing is significant. Most quantitative trading firms — including Two Sigma, which reportedly could obtain A100 GPUs only about 20 percent of the time at peak cloud demand before migrating its infrastructure — have accepted some dependency on public cloud capacity. For market making, where the profitability of individual trades depends on millisecond execution, cloud provisioning delays are an operational risk that XTX concluded was unacceptable.

In January 2025, XTX announced plans to invest more than €1 billion in a private data center campus in Kajaani, Finland, a city chosen partly for its cold climate (reducing cooling costs) and partly for access to renewable electricity. The first building, spanning roughly 15,000 square meters, was designed to carry up to 22.5 megawatts of IT capacity. In June 2026, XTX contracted with Finnish construction group YIT to begin building a third facility on the same campus.

The decision to own rather than rent infrastructure has compounding strategic advantages. XTX can schedule GPU time for model training and backtesting without competing with other customers. It can tune the network topology between compute nodes for the latency requirements of trading workloads specifically. And it retains the competitive moat that comes from having built, rather than purchased, the systems that run the business.

For context: the infrastructure scale XTX has assembled for trading inference is broadly comparable to the training infrastructure of mid-tier AI research organizations — not the frontier labs' largest runs, but larger than most academic compute clusters. The difference is that XTX's cluster is not used to train general-purpose language models; it is used to run financial market inference in real time, every trading day.

XTX's Position in Quantitative Finance's Competitive Landscape

XTX is not the largest firm in quantitative trading. Jane Street, which posted $16.1 billion in trading revenue and $10.3 billion in profits in the first quarter of 2026 alone, operates at a scale that dwarfs XTX's annual figures. Jane Street employs roughly 3,500 people and retains a hybrid model in which human researchers design the strategies that automated systems execute.

But XTX's model represents something different from Jane Street's. Where Jane Street's scale comes partly from human intellectual capital — the researchers and traders who build new strategies — XTX's profitability is generated almost entirely by the AI system itself, with humans in an engineering and maintenance role rather than an alpha-generation role. XTX's 2025 revenue of £3.93 billion produced from fewer than 300 employees translates to a revenue-per-employee ratio that would be difficult to match in almost any industry.

Citadel Securities, another major electronic market maker, takes a hybrid approach to infrastructure, using both owned co-located hardware for latency-sensitive tasks and cloud computing for less time-critical workloads. This flexibility comes at the cost of the compute scheduling certainty that XTX's fully owned cluster provides.

The structural comparison that illuminates XTX's model most clearly may not be another trading firm at all — it may be DeepSeek's parent company, High-Flyer Quant. High-Flyer, a Chinese quantitative trading firm, assembled roughly 10,000 A100 GPUs for its trading operations, then used surplus capacity and profits to fund the development of the DeepSeek AI models. XTX and High-Flyer are following parallel paths: trading-firm-as-AI-infrastructure-operator, using spread-capture economics to fund compute acquisition, which in turn funds research and further model improvement. The template is the same even if the specific research directions diverge.

The broader market structure context matters here. In the post-MiFID II European regulatory environment, Systematic Internalisers — firms that execute client orders against their own capital rather than routing them to an exchange — have become a significant source of liquidity in cash equities. XTX's dominance in this category reflects both the regulatory shift toward electronic market making and the practical difficulty of competing with a firm that has spent ten years optimizing AI models specifically for this function. Virtu Financial, another major electronic liquidity provider, and IMC, the Dutch market maker, operate in overlapping segments but have not publicly matched XTX's infrastructure investment or its purely AI-driven model for market making specifically.

XTX's market-making position was reaffirmed by the Rosenblatt ELP SI ranking for 2025, which placed XTX first among electronic liquidity providers acting as Systematic Internalisers in UK and EU cash equity markets — a position it has held for five of the past six years. Its equity SI volume for 2025 reached EUR 177 billion, up 50 percent from 2024, with average daily volume exceeding EUR 693 million and a single month (April 2025) in which daily volume surpassed EUR 1 billion for the first time for any electronic liquidity provider.

When Trading Profits Fund Mathematical Research

The uses to which Gerko directs XTX's profits offer an unusual window into his priorities. Beyond building data centers, he has committed substantial resources to advancing the formal mathematical infrastructure that underpins AI reasoning.

In November 2023, XTX launched the AI Mathematical Olympiad Prize — a $10 million competition fund intended to encourage the development of AI systems capable of winning gold at the International Mathematical Olympiad, a competition that requires creative mathematical reasoning of a kind that large language models have historically struggled with. The prize represented an early signal of interest in AI-for-mathematics at a time when the field was still nascent.

By July 2025, XTX had moved further. Gerko personally committed $10 million — split evenly between the Lean Focused Research Organization (Lean FRO) and the Mathlib Initiative — to accelerate development of the Lean proof assistant and the mathematical library it uses. Lean is a formal verification system that allows mathematicians to write proofs that can be mechanically checked for correctness. The Mathlib library, which formalizes a growing body of undergraduate and graduate mathematics in Lean, is a prerequisite for AI systems that aim to reason rigorously about mathematical claims rather than pattern-match from training data.

The rationale is not hard to identify for someone with Gerko's background. The same gap that separates human intuition from mathematical proof — the difference between a plausible argument and a verified one — is the gap that separates current large language models from AI systems that could actually advance mathematics. Gerko's trading system profits from being more precise than human judgment; his philanthropy bets that AI will eventually achieve similar precision in mathematical reasoning.

XTX also serves as the founding donor of the Renaissance Philanthropy AI for Math Fund, a broader initiative supporting AI-for-mathematics research across multiple institutions. Additional projects in this space — including the Systems AI Research institute (SAIR), which has drawn involvement from mathematicians including Terence Tao — have received support within this broader funding ecosystem, though the specific terms of XTX's contributions to SAIR were not publicly confirmed at the time of this article's publication.

👉 Read more: Terence Tao argues AI proof abundance is severing math from human insight

What AI Market Making Cannot Yet Do — and Where the Risks Are

XTX's model, for all its demonstrated profitability, is not without structural risks that go largely unaddressed in accounts focused on its extraordinary returns.

The central technical risk in market making is the regime change problem. Markets spend most of their time in mean-reverting or trending states that a well-trained model can price efficiently. But periodic episodes — flash crashes, liquidity crises, geopolitical shocks — produce one-directional flows that leave market makers holding large, quickly depreciating inventories. Human traders manage this partly by reading sentiment from news and social networks, partly by communicating with other traders, and partly by experience-based pattern recognition that does not rely solely on recent price history. Whether XTX's AI system handles these episodes with comparable resilience is not publicly documented; the firm does not publish model architecture details, failure-mode analyses, or post-incident reports.

The £22.5 million tax ruling that the UK Supreme Court imposed on Gerko — relating to deferred compensation arrangements from his years at GSA Capital before founding XTX — is a personal legal matter unconnected to XTX's trading operations. But the ruling prompted Gerko to begin voluntarily winding up HFFX, a dormant investment partnership that has not invested since 2015. The episode illustrates that regulators and courts are still working through the tax treatment of compensation structures that were common in quant finance a decade ago.

Separately, XTX's Finnish subsidiary filed a lawsuit against Dell Technologies in Irish courts in June 2026, alleging that Dell breached the agreed pricing on 1,680 servers — approximately $549 million in contracted value — by imposing a $62 million to $70 million price increase. The lawsuit illustrates the capital intensity of XTX's infrastructure strategy: a data center campus that requires a nine-figure contract with a single server vendor is a significant operational commitment that brings counterparty and pricing risk alongside its performance advantages.

There is also a longer-term structural question that XTX's model raises for the financial system. A sufficiently dominant AI market maker can, in principle, reduce quoted spreads and improve market liquidity — but it can also concentrate the provision of that liquidity in a single system, creating a single point of failure during the episodes when liquidity is most needed. No regulator has yet specifically addressed the systemic implications of AI-only market makers operating at XTX's scale.

The Architectural Question AI Finance Has Not Yet Answered

XTX Markets has demonstrated that a small team of engineers can build an AI system capable of generating the kind of returns that previously required armies of MBAs, traders, and portfolio managers. The 2025 results — £1.71 billion in net profit from fewer than 300 employees, on $250 billion of daily trading activity — are not a projection or an aspiration. They are audited financial accounts filed with Companies House.

The competitive implications for the broader finance industry are significant. The advantage XTX holds is not primarily algorithmic — many firms can build market-making models. The advantage is infrastructural: a decade of proprietary data, a 650-petabyte training corpus, and a 25,000-GPU compute cluster that no institutional competitor has matched in the specifically owned, latency-optimized configuration that XTX has assembled. Building that moat took ten years and billions of pounds. Replicating it from scratch would require both the capital and the time to generate comparable data.

What the XTX model does not yet resolve is the open problem of AI in market making during genuine crises. The conditions that produced XTX's most profitable years — rising volumes, volatile but ultimately trending markets, consistent structural demand for electronic liquidity — may not persist indefinitely. The same GPU cluster and training corpus that extract billions in normal markets will face a different kind of test in conditions that have no historical analog. Whether the AI systems that made Gerko the eighth-richest person in Britain can navigate those conditions is a question the available evidence does not answer.

The next meaningful milestone to watch is whether XTX's AI architecture — or a derivative of it — can be published as a research contribution rather than kept entirely proprietary. Gerko's philanthropy toward formal mathematics suggests a belief that advances in AI reasoning will eventually benefit from open scientific exchange. Whether that belief extends to the trading architecture itself, or whether the competitive moat remains closed, will do more than any rich list ranking to determine whether XTX is building a durable institution or an extraordinarily profitable first-mover advantage.