← All speakers

Frank Long

Conference affiliation: Coatue

Frank Long is Head of AI at Coatue and a partner on its hedge fund team. He argues that agents could have an even greater impact on investing than on coding. His work examines both sides of that proposition: how agents can improve investment research, and whether their adoption will generate enough computing demand to justify the infrastructure being built for them.

On this page

From product development to AI strategy

Frank worked at Google as a product manager on the Maps Augmented Reality team and the Knowledge Engine team within Search. He subsequently moved to Schmidt Futures, where he held operational responsibilities and served as a technical advisor to Eric Schmidt. At Goldman Sachs, he worked on AI in the Executive Office and became a vice president at the Global Institute, writing about emerging technology, geopolitics, and computing infrastructure. His earlier career spans technology products and their wider consequences. He then joined Coatue’s hedge fund team from Goldman Sachs.

In April 2025, Frank argued that cheaper inference does not guarantee sufficient demand to justify enormous investment in AI infrastructure. As models become more efficient, ordinary chatbot use could require less computing capacity than investors anticipate. His case for agents as a source of demand depended on new uses taking hold: systems that undertake extended tasks, navigate large collections of information, and repeatedly reason through a problem. Infrastructure demand therefore depends on what people adopt, as well as what models can do.

The physical consequences of agents

Frank’s May 2025 analysis of rising compute density traced a tradeoff: placing processors closer together improves performance and cost while concentrating electricity consumption and heat. Specialized electrical equipment and liquid cooling make data-center construction an industrial undertaking, with long lead times and the risk that equipment becomes obsolete.

With George Lee, he developed a proposal for flexible AI electricity demand. Their August 2025 work on demand management argued that some AI computation could pause during periods of grid stress or move between regions. A research agent running for several minutes can tolerate network delays that would damage an interactive web application. That creates room to route work toward available electricity and reduce consumption at peak times. Their proposal makes workload scheduling part of the energy solution alongside building new supply.

Investment research as an engineering problem

At Coatue, Frank organizes AI-assisted investing around selection, size, and timing: what to buy, how much to own, and when to act. His approach to investment workflows starts with the information those decisions require. He argues that anything an investor can see on a screen should be accessible to an agent through an API. That includes external material such as filings and news, together with internal context such as positions, liquidity and risk requirements, and earlier investment reasoning. He credits Coatue’s data science team and its longstanding practice of recording decisions—including trades investors passed on and what they learned afterward—with creating that foundation.

Frank’s engineering approach is to evaluate each step separately and choose the method that performs it best. Deterministic code can check whether a stock’s price-to-earnings ratio is below 15; statistical methods can model relationships in data; a language model can summarize an article. A person may be best placed to meet a management team or judge how new evidence changes an investment view. These methods can also work together: code retrieves the evidence, a model helps interpret it, and an investor assesses its portfolio implications.

Korean-language semiconductor reporting illustrates why the research question determines the workflow. Checking an executive’s reported remark might require translating an article, extracting the quotation, and checking it against its source. Challenging an existing investment view might instead begin with an internal note and search for contradictory reporting. Even gathering news involves choices about relevance, translation, and how to represent the material for a model. Frank treats those choices as variables to test against the quality of each task’s output. His objective is better investment decisions, with automation used where it improves the work.

1 conference talk

Loading more talks and highlights…

References