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

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

by Bloomberg

52m•July 9, 2026
Listen to Episode

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

Bloomberg’s Odd Lots speaks with Man Group CTO Gary Collier and Head of Data & AI Tushara Fernando about how a major global asset manager is using AI across investing, research, operations, and coding. The episode focuses on the practical realities of deploying generative AI in finance: how it changes research workflows, why data quality matters more than just having the newest model, how agents are being used to generate and validate trading ideas, and why token consumption at the firm has surged 86x since January.

Main Themes

AI in investing is mostly about augmentation, not magic

  • The hosts frame the core question: is AI just an incremental improvement on prior machine-learning and quant tools, or something fundamentally new?
  • Man Group’s view: AI is already improving nearly every role, but mostly by augmenting human decision-making rather than replacing it.
  • For discretionary and fundamental investors, AI helps synthesize huge volumes of unstructured and structured information into actionable insights.
  • For quants, AI is being used to systematize parts of the research process itself.

The biggest value comes from better data handling

  • Man Group emphasizes three main data layers:
    1. Market data — highly structured, massive scale, including tick data and order books.
    2. Alternative/unstructured data — messy sources like podcasts, reports, and other text-heavy inputs.
    3. Institutional knowledge — internal processes, workflows, and firm-specific context.
  • A key insight: frontier models alone are not enough. The real alpha often comes from:
    • tagging and structuring data properly,
    • building a semantic layer across datasets,
    • connecting disparate sources in a way models can reason over.

AI agents are being used in real investment workflows

  • Man Group is already using agentic systems to:
    • read academic papers and data sets,
    • generate hypotheses,
    • write code for signals,
    • run backtests,
    • evaluate outputs,
    • and present ideas for human review.
  • The firm says 15–20 models have already made it through this pipeline and been approved by a human investment committee for live trading.
  • Example given: an AI agent synthesized a podcast discussion from a hyperscaler engineering executive about GPU scarcity and data-center bottlenecks, helping a PM understand implications for the AI trade.

Token Spend, Workflow, and Economics

Token usage is exploding

  • Man Group says token consumption has increased 86 times since January.
  • This growth is not limited to engineering or tech teams; it is spreading to:
    • finance,
    • operations,
    • HR/people teams,
    • and other non-technical functions.
  • The firm sees this as evidence that AI is becoming a general-purpose workplace tool.

The firm is not yet using an internal router to minimize model costs

  • Instead of automatically routing every task to the cheapest suitable model, Man Group is focusing on education and user judgment.
  • Reason: they want employees to understand:
    • which model is best for which task,
    • how context windows are misused,
    • and how to reduce waste.

Users are learning to reduce unnecessary spend

  • Example: coding agents often invoke models for simple actions like Git commands, which can be handled more cheaply outside the LLM loop.
  • Employees are finding and sharing these efficiency improvements back into the platform.
  • The firm treats token budgets as a real operating issue, but not yet as a strict rationing problem.

Safety, Explainability, and Regulation

Risk control is a central bottleneck

  • Man Group says the key challenge is not just compute or data, but organizational readiness:
    • safe deployment,
    • regulated-use constraints,
    • fiduciary responsibility,
    • and avoiding uncontrolled agent behavior.
  • The episode references common AI mishaps like deleting inboxes or photos and notes that such mistakes are unacceptable at enterprise scale.

Explainability still matters

  • Because Man Group is not a high-frequency trading shop, its holding periods are generally days to months.
  • That means trades must still be defensible in plain English.
  • AI-generated ideas are required to produce a clear natural-language rationale before moving to code and execution.
  • Human approval remains essential.

Talent and the Future of Work

AI raises expectations for all hires

  • Man Group says every new hire should “up the bar” on AI familiarity, regardless of role.
  • The ideal employee is:
    • bright,
    • motivated,
    • subject-matter strong,
    • comfortable with automation,
    • and able to think strategically rather than just execute tasks.

Two labor dynamics are happening at once

  • Democratization: junior staff and non-technical employees can do more with AI.
  • Superstar effect: some people become exceptional AI orchestrators and “conductors” of agent workflows.
  • Man Group says it is seeing both, though democratization is more widespread.

Competitive Advantage and Alpha

Data and market access still matter more than the model alone

  • The guests argue that while many firms may soon have access to strong frontier models, not everyone has:
    • proprietary data,
    • broker relationships,
    • execution infrastructure,
    • or decades of trading expertise.
  • Their view: alpha comes from the network of systems, not a single repository or model.

Some strategies may become more accessible

  • AI can lower the cost of entering new markets or asset classes by reducing the upfront labor needed to:
    • understand data,
    • interpret messy pricing structures,
    • and build systematic infrastructure.
  • This could expand the set of markets that can be researched and traded systematically.

Key Takeaways

  • AI adoption in finance is already operational, not experimental.
  • The highest-value use cases are around data synthesis, research acceleration, and code generation.
  • Data structure and internal context matter as much as the model.
  • Human oversight remains necessary for explainability and regulatory reasons.
  • Token spend is rising fast enough to become a major management and budgeting issue.
  • The bigger challenge may be organizational change, not model performance.

Notable Insight

“There isn’t one code repository in Man Group that I can point at and say, that’s where the alpha is.”

That line captures the episode’s central message: in modern asset management, AI advantage is increasingly about combining data, workflows, infrastructure, and human judgment into a system—rather than betting on one magical model or one single signal.