Jane Street

Jane Street on GPUs, Trading, and Hiring: A Conversation with Dwarkesh

Key Themes

latency regimesGPU infrastructuretrading modelsdata center constraintshuman judgmentcompute scalehiring and mentorshipAI tooling

30 minMay 21, 2026

Summary

Jane Street explains how trading speed, GPU scale, and talent constraints shape its AI strategy

This conversation focuses on how Jane Street balances ultra-low-latency trading systems with slower, model-driven research and execution workflows. The speakers discuss why GPUs, colocated infrastructure, storage, and power/cooling constraints matter so much, how the firm thinks about large compute commitments, and why human judgment still plays a critical role in markets. The second half turns to hiring and organizational scaling, with a strong emphasis on mentorship, research velocity, and building tools that augment human understanding rather than replace it.

Compute capacity is being treated as a strategic growth input rather than a mature utility expense.

The speakers repeatedly describe additional GPUs and infrastructure as immediately useful for more training, retraining, inference, and experimentation, indicating that demand is still ahead of supply.

Talent, mentorship, and organizational throughput appear to be the main scaling bottlenecks, not hardware.

Jane Street says it is still expanding compute aggressively, while emphasizing that finding great people and having enough senior capacity to train them is the harder constraint.

Infrastructure decisions are increasingly shaped by power, cooling, and data-center topology.

The discussion highlights generators, transformers, liquid cooling, distributed facilities, and fiber-distance constraints as practical determinants of how compute can be deployed.

Select any chapter text to Deep Dive with AI

01GPUs, trading horizons, and data center constraints

Jane Street discusses how its trading and research workloads span many latency regimes, from sub-100-nanosecond FPGA systems to slower model-driven decisions, and why that makes GPU infrastructure, colocated facilities, storage, and power/cooling constraints central to its strategy. The conversation also covers the $6B compute deal, the diversity of models and data pipelines being trained, the role of humans in trading during volatile market events, and why AGI would not automatically make Jane Street’s work trivial.

Jane Street operates across multiple trading time horizons, with radically different compute and latency requirements.

The fastest trading systems require FPGA-based packet processing and cannot do much computation.

A core modeling target is predicting fair value, but it is only one of several prediction tasks.

GPU placement depends on latency, model size, and compute needs; some workloads can be farther from exchanges, while ultra-fast ones are constrained by fiber length and colo layout.

The $6B compute deal is intended to support many model trainings and rapid experimentation across diverse architectures.

Financial data is noisier and more sequential than typical LLM user data, pushing different storage, batching, and inference design choices.

Jane Street is building its own large-scale data storage/object store and adapting to distributed data centers and ARM support.

Humans remain important, especially during market dislocations and phase transitions, where judgment can outperform models.

Data center infrastructure has become more modular and power-dense, with changing bottlenecks such as generators, transformers, and liquid cooling equipment.

Compute is highly valuable because the firm is constrained by available compute and turns away high-value runs.

02Compute Growth, Hiring, and AI/ML Roles

The speakers discuss Jane Street’s expanding compute needs, why they expect to keep scaling GPU capacity, and how that affects hiring and organizational growth. They emphasize that compute is still far from a bottleneck, that more research and experimentation would immediately use additional capacity, and that the bigger constraint is finding and training great people. The segment closes with a broad overview of open roles across engineering, ML, trading, hardware, formal methods, and front-end work, plus a mention of puzzles as part of Jane Street’s culture and recruiting outreach.

Jane Street says it is far from having too much compute and can use more for research, experimentation, retraining, and bulk inference.

They describe a strategy of committing to power and data-center capacity while potentially delaying chip decisions, since chips are harder to offload than infrastructure.

Hiring is framed as constrained more by finding and mentoring great people than by hardware availability.

The company says it is growing from tens of thousands of GPUs toward hundreds of thousands, viewing this as justified by the business.

Open roles span physical engineering, machine learning, trading, software engineering, fleetwide optimization, hardware engineering, formal methods, and front-end development.

The speakers stress a human-centered approach to tools, including AI tooling, aiming to increase understanding and agency.

Puzzles are described as deeply embedded in Jane Street’s culture and also a recruiting touchpoint.