Inside Hudson River Trading's Blistering Token Burn

June 5, 2026

31 min episode · 2 min read

Ian Dunning

Episode

31 min

Read time

2 min

Topics

Investing

AI-Generated Summary

Published Jun 5, 2026

Key Takeaways

✓GPU Compute Scarcity: Securing 6,000 Blackwell GPUs with a full data center solution for Q4 2026 delivery is effectively unavailable at reasonable prices. Next-generation Rubin GPUs for 2027 are already heavily pre-sold. Organizations needing compute must enter procurement queues now or face costly delays and perpetual catch-up cycles.
✓AI Trading Signal Interpretability: HRT's models operate effectively on short time frames without human-interpretable logic — similar to how neural networks predict one-minute price moves with no explainable rationale. The practical threshold for trusting uninterpretable signals may extend to longer time frames as model performance compounds across thousands of instruments simultaneously.
✓Data Center Counterparty Risk Management: Long-term compute contracts — typically 3–5 years, sometimes requiring 50% upfront payment — carry meaningful credit risk on both sides. HRT evaluates NeoCloud providers by reviewing their CDS spreads and considers purchasing CDS insurance at roughly 10 cents per GPU-hour equivalent to hedge provider insolvency risk.
✓Token Spend as Productivity Proxy: HRT teams average roughly $100–$200 per person daily in AI token costs, with high-usage individuals reaching $1,000 per day during intensive research bursts. Organizations should monitor whether high spenders are discovering novel workflows rather than defaulting to restricting usage, as token access may compound productivity advantages over time.
✓Hiring Shift Toward Theorists: HRT is moving toward open-book, AI-assisted interviews and reconsidering the traditional quant archetype. As AI handles implementation work, the premium shifts to candidates who generate novel hypotheses and communicate requirements precisely — what Dunning describes as clear, unambiguous problem description skills rising sharply in hiring value.

What It Covers

Ian Dunning, Head of AI at Hudson River Trading, discusses how HRT applies AI to high-frequency trading across global markets, the escalating GPU compute shortage, counterparty risks in data center contracts, talent acquisition shifts, and whether AI-driven trading signals can operate without human-interpretable logic.

Key Questions Answered

•GPU Compute Scarcity: Securing 6,000 Blackwell GPUs with a full data center solution for Q4 2026 delivery is effectively unavailable at reasonable prices. Next-generation Rubin GPUs for 2027 are already heavily pre-sold. Organizations needing compute must enter procurement queues now or face costly delays and perpetual catch-up cycles.
•AI Trading Signal Interpretability: HRT's models operate effectively on short time frames without human-interpretable logic — similar to how neural networks predict one-minute price moves with no explainable rationale. The practical threshold for trusting uninterpretable signals may extend to longer time frames as model performance compounds across thousands of instruments simultaneously.
•Data Center Counterparty Risk Management: Long-term compute contracts — typically 3–5 years, sometimes requiring 50% upfront payment — carry meaningful credit risk on both sides. HRT evaluates NeoCloud providers by reviewing their CDS spreads and considers purchasing CDS insurance at roughly 10 cents per GPU-hour equivalent to hedge provider insolvency risk.
•Token Spend as Productivity Proxy: HRT teams average roughly $100–$200 per person daily in AI token costs, with high-usage individuals reaching $1,000 per day during intensive research bursts. Organizations should monitor whether high spenders are discovering novel workflows rather than defaulting to restricting usage, as token access may compound productivity advantages over time.
•Hiring Shift Toward Theorists: HRT is moving toward open-book, AI-assisted interviews and reconsidering the traditional quant archetype. As AI handles implementation work, the premium shifts to candidates who generate novel hypotheses and communicate requirements precisely — what Dunning describes as clear, unambiguous problem description skills rising sharply in hiring value.

Notable Moment

Dunning describes a moment when HRT's model spontaneously clustered meme stocks, crypto-adjacent equities, and certain Wall Street Bets favorites together in high-dimensional space — with no fundamental basis — yet the grouping proved coherent under analysis, illustrating how model logic can outpace human explanation entirely.

Know someone who'd find this useful?