Overview of Inside Hudson River Trading's Blistering Token Burn
This live Odd Lots episode centers on how Hudson River Trading (HRT) is using AI internally, what the firm is learning from the latest frontier models, and why the real constraint in the AI arms race is increasingly power, data-center capacity, and deployment logistics rather than just chips. Guest Ian Dunning, HRT’s head of AI, argues that models are already useful for research acceleration, but still fall short of fully replacing human researchers—at least for now.
What HRT Is Actually Using AI For
Internal research acceleration
HRT is not mainly using AI as a magical trading brain. The biggest use cases are more practical:
- Writing and improving code
- Proposing research ideas
- Monitoring experiments
- Helping quantify and systematize signal discovery
The firm sees models as a force multiplier for its researchers, especially in a high-frequency context where the problems are tightly defined and can be heavily risk-checked.
Models are getting better, but still imperfect
Dunning says each new model release makes it easier to spot a shrinking set of errors. He describes recent Anthropic releases as incremental but meaningful improvements over earlier versions. HRT is still benchmarking models against human researchers and looking for where they fail on quant tasks.
The Big Bottleneck: Power and Deployment, Not Just GPUs
Chips are necessary, but not sufficient
One of the clearest themes is that access to GPUs alone is no longer the full story. HRT may be able to get chips, but the harder part is finding:
- Data-center space
- Enough electricity
- Suitable network/storage infrastructure
- Long-term capacity at a reasonable price
Dunning says firms are effectively scrambling for whatever megawatt-scale capacity they can lock in.
Long-term contracts are now part of AI infrastructure strategy
HRT is negotiating multi-year arrangements for thousands of GPUs at a time. These are not spot purchases; they involve:
- Long-term leases
- Upfront payments or staggered payment schedules
- Counterparty risk concerns
- Power-rights and lease timing issues
He describes a marketplace that has become intensely competitive and still feels immature.
Could Compute Become a Tradable Asset?
A real market for compute is plausible
The conversation touches on the idea of compute futures or other financial instruments tied to AI capacity. HRT could potentially use such products to hedge the risk of waiting too long and paying more later.
That said, Dunning is skeptical about how easily compute can be standardized:
- What exactly counts as “compute”?
- What does physical delivery mean?
- How do you define a benchmark across very different clusters?
His view: the idea is promising, but hard to operationalize cleanly.
AI, Trading, and the Loss of Human Intuition
“Why bother with a story?”
A recurring theme is whether AI-driven trading will increasingly abandon human-friendly explanations. Dunning suggests the industry may be moving into a world where:
- Backtests matter more than intuitive narratives
- Models can exploit patterns humans don’t understand
- Even bizarre signals may be tradeable if they work reliably
The hosts push on the idea that this feels like a loss of control, but Dunning’s response is that humans already accept machine advantage in short-horizon prediction and math-heavy tasks.
Short-term trading is easier to automate than long-term discretionary betting
HRT’s comfort with AI is partly because it operates in a highly controlled, short-horizon environment. That’s very different from long-term discretionary investing, where:
- Positions are held for months
- Risk is intentionally concentrated
- Controls are less mechanical
- Explanability matters more
That makes HFT a more natural place for AI integration.
Talent, Productivity, and “AI Delirium”
Token spend is becoming a real operating expense
HRT now tracks token usage as a genuine cost line. Dunning says many employees are spending roughly:
- $100–$200/day on average
- Some heavy users much more
He sees this as part of a broader productivity shift: if AI meaningfully boosts output, firms that can afford more usage may compound their advantage.
Hiring is changing too
The episode suggests AI is reshaping what employers want:
- Stronger emphasis on clear prompting and communication
- More openness to “theorists” who can generate ideas
- Less expectation that candidates manually implement everything
- Increasing willingness to let AI assist in interviews and workflows
Dunning jokes that “prompt clarity” is becoming a real skill.
Talent competition remains intense, but the landscape is shifting
He notes that big AI labs have become “big tech” in some respects, with changing culture and compensation dynamics. At the same time, the perceived upside for new entrants may be narrowing as the biggest players get larger.
Key Takeaways
- AI at HRT is a research multiplier, not a magical oracle.
- The limiting factor is increasingly power and deployment infrastructure, not simply chip availability.
- Compute may eventually become a financialized asset class, but standardization is hard.
- AI is already changing quant hiring, productivity, and internal workflows.
- The episode captures a strong sense of compounding acceleration—part real, part “AI delirium.”
Notable Insight
The most important scarce resource in the AI race may not be GPUs themselves, but the ability to actually deploy them at scale with power, storage, networking, and long-term contractual certainty.
Bottom Line
This conversation paints a vivid picture of how a major trading firm is adapting to AI: aggressively, pragmatically, and with a healthy mix of excitement and skepticism. HRT sees models as increasingly valuable for research and code generation, but the bigger story is the industrial-scale scramble for compute infrastructure—and the possibility that AI is creating a new class of market structure, talent dynamics, and even future financial instruments.