The primary constraint on scaling AI is no longer GPU availability but the physical infrastructure, specifically powered real estate, to house and run them.
The intense, simultaneous AI investment by all major quantitative trading firms will inevitably lead to a 'forcing function' that compresses profit margins over time.
NVIDIA has an exceptionally strong competitive moat in AI training hardware, with Google's TPUs being the only major alternative, which itself introduces vendor lock-in risks.
Large-scale compute must be secured through non-negotiable, long-term contracts for thousands of GPUs, reflecting a severe seller's market.
Advanced AI models are reaching a point of capability where they can meaningfully augment high-level researchers, potentially shifting hiring priorities towards conceptual 'theorists' over pure implementers.
Pre-2024
Dunning states that Hudson River Trading consistently underestimated its long-term GPU requirements, calling this a 'punishing failure in planning'.
Early 2024 (Implied)
Dunning's team found that Anthropic's Opus 4.0 models were not yet capable enough to meaningfully augment their human researchers.
Mid-2024 (Implied)
A capability threshold was crossed with Anthropic's Opus 4.5, which Dunning found could successfully augment researchers, signaling an acceleration in AI development.
2024
Dunning identifies the primary bottleneck for AI compute as having shifted from GPU supply to the availability of powered data center capacity.
Future (2027)
Dunning predicts that NVIDIA's next-generation Rubin GPUs will be sold out for their initial release period in 2027.
▶The Physical Constraints of AI Scale-UpJun 2026
Dunning argues that the primary scarcity in AI infrastructure is no longer the chips themselves, but the physical real estate and power required to operate them. Finding a complete, powered data center solution for thousands of GPUs is now the main challenge, eclipsing the difficulty of acquiring the hardware.
Investors should shift their focus from solely tracking semiconductor supply chains to also analyzing the real estate, power grid, and data center construction sectors as key indicators of AI growth potential and bottlenecks.
▶The AI Arms Race in Quantitative FinanceJun 2026
Dunning details a massive, simultaneous investment in AI by Hudson River Trading and its peers. This involves applying a unified AI trading approach across all global asset classes and developing custom inference chips, but he predicts this intense competition will ultimately lead to margin compression.
The competitive advantage in quantitative trading is becoming less about a single superior algorithm and more about the ability to secure and efficiently deploy vast amounts of compute, suggesting a capital-intensive future where scale is paramount.
▶Navigating a Seller's Market for ComputeJun 2026
According to Dunning, the market for large-scale AI compute is heavily supply-constrained, forcing firms like HRT to be non-selective and commit to long-term (3-5 year) contracts for thousands of GPUs wherever they become available. This dynamic is driven by NVIDIA's market dominance and the sold-out status of even future-generation chips like Rubin.
Firms that can secure long-term compute capacity now, even at high prices, may gain a significant multi-year advantage over competitors who are left to scramble for scarce resources in the spot market.
▶AI-Driven Workforce TransformationJun 2026
Dunning observes that the increasing capability of AI models, such as Anthropic's Claude, is beginning to change the nature of work and hiring at HRT. He notes the firm is considering hiring more 'theorists' who can generate ideas, leaving the implementation to AI, reflecting a shift in the value of human skills.
The demand for talent in highly technical fields may bifurcate, with a premium on high-level conceptual thinkers and a potential devaluation of pure implementation skills that can be automated by advanced AI.