One of the World's Largest Hedge Funds on Its 86x Growth in Token Spending

Episode Summary

Executive Summary: Odd Lots interviews Man Group CTO Gary Collier and Head of Data and AI Tushara Fernando on how AI is being used in asset management. They argue the biggest gains come less from frontier models alone and more from structured/proprietary data, semantic layers, workflow design, and safe deployment. AI is already augmenting research, coding, and idea generation across the firm, with token use exploding as agents take on larger tasks.

Main Topics: AI as augmentation in investment workflows (Priority: 5/5): The guests describe AI as a tool that helps researchers, PMs, and traders synthesize large volumes of structured and unstructured information into actionable insights, rather than as a standalone stock-picking oracle. Data quality, tagging, and semantic layers (Priority: 5/5): They emphasize that alpha depends heavily on cleaning, labeling, and connecting datasets, especially unstructured and alternative data, so models can understand context and relationships between fields. Agentic workflows and systematic strategy generation (Priority: 5/5): Man Group is building multi-agent systems that can ideate hypotheses, write code, run backtests, and validate outputs, with some AI-generated strategies already approved by human investment committees. Model choice, cost control, and token budgeting (Priority: 4/5): The conversation covers how the firm manages rapidly rising token usage, how different tasks require different models, and why education rather than automatic routing is the preferred approach so far. Explainability, risk, and regulated deployment (Priority: 5/5): Because Man Group operates in a regulated environment, AI outputs must remain understandable and defensible to humans, regulators, and risk committees, with guardrails to prevent unsafe automation. Organizational change and the future of talent (Priority: 4/5): The guests argue that AI changes hiring and internal workflows: all new hires should be AI-literate, and the best employees will increasingly act as orchestrators of agents rather than hands-on executors. Where durable alpha comes from (Priority: 4/5): They say durable advantage is not just from models, but from the broader ecosystem: proprietary data, trading relationships, infrastructure, research process, and internal knowledge captured in playbooks.

Key Arguments: AI is already embedded across Man Group’s discretionary, quantitative, and operational workflows, with almost every role having an AI-focused tool on screen. For long-only and discretionary investors, AI mainly increases research breadth and speed by summarizing earnings, broker research, podcasts, and alternative data asynchronously. For quant/systematic work, AI is being used end-to-end: idea generation, hypothesis writing, code generation, backtesting, and agentic review before human approval. The highest value often comes not from the frontier model itself but from how the firm structures and labels its own data and connects datasets through a semantic layer. Fine-tuning is useful in some cases, but the current biggest payoff comes from preprocessing, metadata, and clear descriptors that explain what each row or field actually means. AI helps convert messy unstructured information into searchable, comparable knowledge, allowing firms to identify market bottlenecks or investment signals from sources like podcasts. Man Group insists on explainability because its strategies have longer horizons and must be defensible to committees and regulators; it is not pursuing opaque HFT-style systems. Token usage has exploded because agentic workflows can now tackle much larger tasks; the firm is still in an expansion phase rather than a hard rationing phase. Rather than rely on automatic routing, the firm is educating users about model selection and responsible usage, while making budgets transparent at the business-unit level. The likely future of investment labor is a mix of democratization and superstar operators, but the bigger immediate effect at Man Group is broad-based productivity gains across many employees. AI may incrementally expand the opportunity set into harder-to-approach markets, especially where instrument definitions and pricing are messy or verbally negotiated. A durable alpha moat comes from combining AI with years of market access, proprietary data, backtesting capability, and trading infrastructure—not from a single codebase or model.

Data Points: Token consumption growth since January: 86x - Gary Collier says firmwide token usage has risen 86 times since January, across functions including finance, operations, and people teams. AI-generated models approved for trading: 15–20 models - Tushara Fernando says roughly 15 to 20 AI-ideated models have gone through signal construction, validation, and human investment committee review to be traded. Data volume from market ticks: almost a terabyte per day - Man Group ingests every tick from most exchanges, amounting to nearly a terabyte of tick data daily. Time horizon for trading decisions: days to weeks to months - Explains why the firm prioritizes explainability and does not operate like opaque high-frequency trading shops. Agentic task duration improvement: doubling about every 7 months - They cite the METR benchmark, saying the amount of time an agent can work autonomously is doubling roughly every seven months. Human-equivalent task length now possible for agents: about 16 hours - They say current agentic workflows can now handle tasks that would take a human roughly 16 hours. Firm size: 17,000–18,000 people - Used to explain why AI education and adoption have to work across a large, heterogeneous organization.

Pivotal Quotes: "The AI did it." — Gary Collier: He rejects opaque decision-making and says Man Group wants every trade to remain explainable to humans. "It’s this whole network, this whole ecosystem that together I think drives alpha." — Gary Collier: He argues durable returns come from the interaction of data, infrastructure, market access, and AI—not a single model or repository. "We’ve seen people finding quite creative ways to reduce token spend and contributing that back to the platform as a result of it." — Tushara Fernando: Discussing how education and transparent budgeting have led users to improve efficiency rather than just spend more tokens.

Implications: AI in finance is shifting from novelty to infrastructure. Winning firms will pair strong data engineering, governance, and workflow design with model access. The edge is likely to come from proprietary context, not just better models.