The Curation Layer Is the Moat in AI-Driven Quant Research

On May 21, NVIDIA published a working system for generating investment signals automatically: three agents that invent ideas, code them into computable signals, and run them through backtests in a closed loop. The best signal the example produced scored 0.0138 on a correlation-style measure of predictive power — an information coefficient, or IC. Most quantitative funds consider 0.02 the minimum for a signal worth trading. Below that line, transaction costs eat the edge.

That is the point, according to people building the next layer. Man Numeric, the systematic investment arm of Man Group, has been running a production version of the same underlying architecture since 2025. Its system, called AlphaGPT, has produced several dozen investment signals approved for live trading, according to coverage by Hedgeweek. The gap between NVIDIA's demonstration and Man Numeric's live book is not the model. It is what happens between the signal generator and the order terminal — a proprietary curation and filtering pipeline that the company does not publish.

"The curation layer is where the work actually happens," said one quant researcher at a large US hedge fund who asked not to be named because their firm's pipeline is not public. "Everyone can build the loop. The question is which signals you throw away."

The curation layer is an unglamorous collection of filters: signal quality screens that reject variants too correlated with existing positions, statistical backtests that stress the signal under different market regimes, risk models that check sector and factor exposure, and human review stages before a signal graduates to paper trading. Man Numeric has not disclosed the exact composition of its pipeline, but the firm has said that AlphaGPT's speed introduces statistical risks — testing numerous variations rapidly increases the probability of patterns that look significant in-sample but vanish out-of-sample, a problem practitioners call p-hacking.

The pattern NVIDIA published is commodity. The pipeline around it is not.

By mid-2025, 73 percent of Y Combinator investment-related startups funded between January 2024 and June 2025 were agentic AI-related, according to CFA Institute Research. The bulk of those startups are not inventing new model architectures. They are wiring together existing components and, increasingly, building the filtering and validation layers that sit between a language model's output and a live trade.

Man Group declined to specify whether AlphaGPT's production pipeline uses the same three-agent architecture NVIDIA published. NVIDIA's technical blog describes a signal-generation agent, a code agent, and an evaluation agent running in a closed loop; Man Numeric has said it uses an agentic system for autonomous signal generation, coding, and backtesting — language consistent with that architecture, but unconfirmed. The gap matters: if the generation layer is commoditizing while the curation layer remains proprietary, the competitive dynamics differ from a world where both are firm-specific.

The infrastructure required to run one of these systems is not trivial. Running the NVIDIA NIM blueprint locally requires a single NVIDIA GPU with 48GB or more of VRAM — an A100 80GB or H100 at current cloud list prices runs roughly $2-3 per GPU-hour depending on provider and spot availability. That puts the compute cost of a small signal discovery operation at thousands of dollars per month before a single signal survives the curation layer and reaches a trading book. Larger funds can absorb that cost as research overhead; smaller quant shops face a compute-to-curation tax that their pipelines may not yet justify.

The institutional threshold that separates a demo from a live signal — an information coefficient above 0.02 — is not arbitrary. It is the point at which transaction costs, slippage, and market impact stop erasing the signal's edge. NVIDIA's published example did not cross that line. Man Numeric's production system apparently does, for some signals, after curation. The difference between the two is the story.

What to watch next is straightforward: whether more traditional quant firms publish their curation pipeline specifications, or whether the filtering layer stays proprietary while the generation layer becomes commoditized. If the latter, the competitive moat in AI-driven quant research has already moved downstream — from the agent that invents signals to the software that decides which ones are worth running.

Man Numeric has approved dozens of AlphaGPT signals for live trading, according to coverage by Hedgeweek. The NVIDIA technical blog that described the underlying three-agent architecture is available on the developer site. The CFA Institute research on agentic AI adoption in finance is published online.