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

Man Group, one of the world's largest hedge funds, has seen its token consumption increase 86 times since January, a figure the firm did not anticipate and one that spans finance, operations, and people teams rather than being confined to technology departments. The primary driver is the scalability of agentic workflows: the METR benchmark shows that the duration of tasks an agent can autonomously complete is doubling roughly every seven months, with agents now capable of completing work that would take a human 16 hours. Man Group's head of data and AI, Tushara Fernando, whose title was updated from head of data and machine learning specifically to reflect the addition of generative AI, draws a distinction between traditional ML focused on prediction and generative AI focused on enabling people to create things.

Approximately 15 to 20 trading signals ideated entirely by AI have passed through the firm's full signal construction and validation process and been approved by a human investment committee to trade client assets. The agentic quant research system, built over more than a year and a half, automates idea formation, code construction, backtesting, and output evaluation by reviewing academic papers and labeled datasets, reasoning about economic hypotheses, and using further agents to evaluate prior agent outputs. Critically, AI agents write an investment hypothesis in plain English before writing any code, preserving the firm's requirement that all trading decisions remain explainable to satisfy its obligations as a regulated business with fiduciary duty.

AI tools are now visible across every role at the firm including discretionary investing. A portfolio manager covering the AI trade used an AI agent to transcribe and synthesize a podcast from a hyperscaler head of engineering, identifying GPU scarcity and data center bottlenecks as investment signals. The engineering head had noted that data centers are becoming scarce and increasingly difficult to find at the scale needed to train large models, pointing to future demand for better inter-data-center networking. Shared workflows such as backtesting and reading investment reports are encapsulated in AI playbooks stored in a knowledge platform accessible across the trading floor, though discretionary fundamental investors guard roughly 10 percent of their processes as a personal source of alpha and sometimes withhold investment particulars from that shared system.

On model selection, the firm finds that for quant research tasks the quality of structured and tagged underlying data matters more than using the latest frontier model, while for coding tasks the latest frontier models are preferred. This aligns with a Bridgewater paper describing the fine-tuning of an open source version of Qwen on proprietary data to identify newsworthy financial information, where combining the open source model with proprietary data outperformed the most frontier US models for that specific task. Man Group ingests nearly one terabyte of tick data per day and categorizes its data into structured market data, alternative and unstructured data, and institutional knowledge, with the institutional knowledge layer described as quite new.

Man Group modeled expected use cases toward the end of last year to set token consumption budgets, then federated those budgets to individual business units rather than managing them centrally. The firm chose not to build an automated query routing classifier and instead relied on employee education to guide model selection across its roughly 1,700 to 1,800 staff. Budget transparency led employees to find creative ways to reduce token spend, including intercepting tool calls outside the agentic loop to avoid coding agents unnecessarily processing entire command outputs as tokens. The firm has not yet built an internal router to optimize token spend, indicating it remains in an experimentation phase rather than a cost-control phase.

Man Group identifies organizational change management as the primary bottleneck to AI deployment, not idea generation or technical capability. The firm now requires every new hire regardless of role to demonstrate familiarity with AI and expects that as the cost and time to execute tasks falls rapidly, the focus will shift toward planning what to build and how workflows connect across teams. Formal conversations about ratios of token spend to employee salaries have not yet begun, though the firm expects to apply its existing principle of directing resources to where they produce the best economic outcomes. Joe Wiesenthal cautioned that the 86x growth in token usage still needs to show up in a concrete number such as expense reduction or revenue generation to be validated, and raised the unresolved question of whether rainmakers will implicitly upload their knowledge to AI systems or revert to protecting it manually.