Source files: Hedge-fund token-burn raw ledger; HRT promoted note; title-blind topical ledger Source status: HRT local transcripts verified against the public episode page and a second transcript mirror; Man Group local audio/transcript and promoted note ingested 2026-08-16 and checked against the canonical publisher page; X evidence publicly indexed but not API-retrieved. Confidence: MEDIUM for the HRT directional range; LOW-MEDIUM for any cross-firm generalization.
The best current public hedge-fund signal is HRT. Iain Dunning, HRT’s Head of AI, described average token use on his team at roughly $100–$200 per employee per day, with bursty users around $1,000 per day. If the average range held across 250 workdays, the arithmetic is about $25,000–$50,000 per active team user per year. The $1,000 figure annualizes to $250,000 only as a deliberately unrealistic “every day is a burst day” stress case.
Man Group supplies a different signal: its public July 2026 discussion described token consumption growing 86×, but did not disclose the starting point or an absolute spend figure. The same discussion indicates the firm has not yet adopted a salary-to-token ratio and expects the relevant unit to shift from people to agents and workflows.
So the defensible answer for Thursday is:
There is no public hedge-fund industry benchmark yet. One HRT team has disclosed a very high directional range—roughly $25,000–$50,000 per active research user per work-year at the reported average—while Man Group is measuring explosive growth without yet reducing it to salary ratios. The right next KPI is cost per verified research workflow, not tokens per employee.
Why “per employee” is the wrong final denominator
Employee-level burn is useful for capacity planning and adoption segmentation. It does not answer whether the spend is productive. Two researchers can have identical token bills while one is running high-value, reusable research agents and the other is repeatedly expanding context, calling tools inefficiently, or asking a model to redo work.
For a hedge fund, token consumption should be reported in layers:
| KPI | What it tells the firm | Example denominator |
|---|---|---|
| Cost per licensed or active employee | Adoption, access, and capacity planning | User-day or month |
| Cost per research workflow | Economic cost of a bounded process | One data investigation, code review, or document synthesis |
| Cost per verified artifact | Whether spend produces reusable work | Accepted feature, cleaned dataset, or completed backtest |
| Cost per accepted hypothesis/signal | Research throughput with quality gate | Hypothesis that survives pre-registered testing |
| Cost per attributable investment outcome | Portfolio economics | Only after point-in-time, out-of-sample attribution and controls |
The last metric should be delayed, not improvised. A fund should not claim that tokens created alpha because a researcher using an agent produced a profitable idea. The required comparison includes the research clock, point-in-time data boundary, out-of-sample period, model/version, human involvement, and multiplicity correction.
What McKinsey will probably say about it
McKinsey’s July 2026 AI-economics material points in the same direction. Its likely recommendation is to create an AI FinOps control plane that inventories use cases, attributes cost, routes work across models, instruments context/tool/retry behavior, and reports cost per task or outcome. For a hedge fund, that becomes:
- Give each research workflow an owner and an explicit budget.
- Separate model inference from retrieval, context, tool calls, retries, and human verification.
- Route extraction and routine coding to smaller/cheaper models; reserve expensive reasoning for steps where it changes the research result.
- Require a quality gate before an agent can run unattended, even for a bounded task.
- Review the workflow portfolio quarterly: cost, reuse, latency, failure modes, verified artifacts, and investment-process relevance.
That is why “tokens per employee” is likely to appear as a diagnostic, not as the headline business case.
How to interpret the HRT and Man Group signals together
HRT gives an absolute scale signal from a named practitioner. It is unusually useful because it allows arithmetic annualization, but it is still one team’s estimate and not a vendor bill or audited budget.
The newly captured June 2026 follow-up also adds the infrastructure denominator: the discussion links the token bill to model-training and research workflows while describing GPU lot sizes, power, and available data-center capacity as practical constraints. The episode page describes those topics, and the archived timestamped transcript contains the underlying recording text. This does not turn a practitioner estimate into an audited firm budget; it does show why token cost should be joined to GPU, power, latency, and workflow-throughput measures.
Man Group gives a growth signal and a more important organizational signal: token spend is increasingly attached to agentic workflows rather than directly to people. An 86× increase could reflect more users, longer contexts, more agents, more model calls, a low starting base, or all of these. Without the base and denominator, it cannot be compared with HRT’s dollars per day.
The public X discussion adds color but no stronger benchmark. A startup anecdote of $300,000 in annual Claude spend for a $150,000 engineer and vendor-adjacent claims about $250,000 of annual tokens per engineer are useful stress cases. They should not be presented as hedge-fund norms. The more actionable X signal is the repeated warning that context and tool schemas can dominate the bill, which supports McKinsey’s call for workflow instrumentation.
Questions for McKinsey
- “Do you have a finance or hedge-fund benchmark for cost per active researcher-day, or only broader enterprise AI-spend data?”
- “Would you budget tokens per employee, per workflow owner, or per agent portfolio?”
- “What is your preferred definition of a verified research artifact?”
- “How do you prevent a productivity dashboard from rewarding token volume rather than useful, reusable output?”
- “What controls do you recommend before connecting an agent to data production, backtesting, or an investment decision?”
Evidence Boundaries and provenance
The HRT signal comes from a local transcript and promoted research note: Inside Hudson River Trading’s Blistering Token Burn. The Man Group signal is from the dated canonical publisher episode, local audio/transcript capture, and the promoted Man Group note. The local publisher transcript contains transcription errors and does not disclose an audited spend baseline. X material was publicly indexed but not retrieved through the X API in this run, so it is commentary-level evidence. None of these sources establishes fund-wide spend, performance attribution, or a salary/token ratio.
Raw ledger: hedge-fund-token-burn-per-employee-2026-raw.md Related queue: finance AI people broad ingestion queue Related audit: hedge-fund AI capability audit