Inside Hudson River Trading's Blistering Token Burn

Episode Summary

Executive Summary: Odd Lots’ live show with Ian Dunning of Hudson River Trading explored how AI is reshaping quant trading, from model evaluation and research acceleration to the explosive demand for GPUs, power, and data-center capacity. The discussion emphasized that AI is already useful in narrow trading workflows, may increasingly replace parts of human research, and is constrained less by chips than by infrastructure, contracts, and electricity.

Main Topics: AI in quant trading (Priority: 5/5): Ian Dunning described how HRT uses AI mainly to speed up research tasks such as coding, ideation, and experiment monitoring, while noting that the models’ usefulness is improving but still uneven. Model capability and evaluation (Priority: 5/5): The conversation focused on how new model releases are assessed, what kinds of errors remain, and how HRT tries to benchmark models against humans on quant-research-style tasks. Compute scarcity and infrastructure (Priority: 5/5): A major theme was that the real bottleneck is not just GPUs but finding power, data-center sites, and long-term capacity agreements to actually deploy them. The future of market prediction (Priority: 4/5): The speakers debated whether AI and massive compute could push trading toward increasingly opaque, backtested, model-driven strategies that require less human intuition. Build-vs-buy and custom chips (Priority: 4/5): The discussion covered in-house inference chip efforts, vendor lock-in, and the strategic importance of NVIDIA, Google TPUs, and other chip suppliers. Talent, productivity, and AI-assisted work (Priority: 4/5): They discussed how AI changes hiring, interview expectations, and productivity, including the idea that token-rich teams may gain a compounding advantage over token-poor ones.

Key Arguments: AI is already useful at HRT for accelerating research workflows, especially coding, ideation, and monitoring experiments, even if it is not yet fully transformative. The pace of model improvement is fast enough that each major release can feel like a step change, but recent gains are becoming subtler and harder to measure. Trading firms are increasingly investing in AI simultaneously, which could make the advantage from compute scale harder to sustain indefinitely. The main bottleneck in AI deployment is often not the chips themselves but the ability to secure power, sites, networking, and long-term data-center contracts. Long-term compute demand is so uncertain that firms are struggling to forecast GPU needs and are constantly playing catch-up. AI could eventually support more opaque, highly backtested trading strategies that humans may not be able to explain intuitively. Risk management is easier in HFT-style systems with automated controls than in long-horizon discretionary strategies, where delegating to AI would be harder to justify. Custom inference chips are an active area of interest, but training remains dominated by a small number of large vendors with significant moats. AI is changing talent markets by making implementation easier and elevating people who can clearly describe problems and generate ideas. Token access may become a competitive advantage, with richer teams potentially compounding productivity gains over time.

Data Points: Live show attendance: over 300 / 350 people - Odd Lots described the New York City live show as their biggest ever, with attendance cited at roughly 300-plus and then 350 people. Compute spend per employee: $100–$200 per day - Ian said HRT’s average token spend per employee on his team is on this order. Higher token spend range: about $1,000 per day - He noted that some people are spending at this level during bursts of experimentation. GPU request example: 6,000 Blackwell GPUs - Used as an example of the scale and difficulty of sourcing compute capacity on short notice. GPU contract horizon: 3–5 years - Ian described long-term compute contracts as spanning several years rather than being spot purchases. Resource size example: 8,000 GPUs - He gave this as an example of the scale of long-term contracts firms are negotiating.

Pivotal Quotes: "I think we're good. At training models, we have a lot of compute, and people are good at doing the cycle of research, which is required to catch up to the sort of frontier." — Ian Dunning: On whether HRT could build a frontier model or DeepSeek-style competitor. "I feel like we're in that world today. It's sort of post-post-capitalism... markets are just... gambling market[s], including public markets." — Ian Dunning: On whether AI and modern markets are moving toward purely backtested, intuition-light trading. "The chips are available, but not the capacity." — Ian Dunning: On the difference between buying GPUs and actually securing usable data-center infrastructure.

Implications: AI is becoming a strategic input in trading, but scaling it depends on scarce power and data-center capacity. Firms that secure compute and integrate AI into workflows early may gain a durable edge, while talent and process will increasingly center on AI-assisted research.