Inside Hudson River Trading's Blistering Token Burn

Hudson River Trading's AI-driven trading operation is racing toward a future where models predict markets with near-magical precision—yet the real bottleneck isn't compute, but access to power and data center capacity. Ian Dunning, head of AI at the firm, reveals that while they could theoretically launch a deepseek-style LLM, the capital and infrastructure demands make it impractical. Instead, they're locked in a high-stakes race to secure GPU clusters, often accepting suboptimal sites and long-term contracts with hyperscalers. What’s more unsettling: the firm is already seeing AI agents outperform human researchers at ideating trading signals—though they’re still only at 'intern level'—and the most powerful models now make decisions based on patterns no human can interpret, like linking meme stocks through hyper-dimensional clustering. The result? A self-reinforcing cycle where the wealthy get faster, more productive, and more profitable—simply by spending more on tokens. This isn’t just AI adoption; it’s a new economic regime where speed, scale, and capital are the only moats left. The episode also exposes the hidden fragility beneath the AI boom: data centers are scarce, contracts are complex, and even the most powerful firms are playing catch-up. Dunning admits he’s failed to forecast demand for GPUs, and now every new model release forces a scramble.