Hudson River Trading says models from its HAIL laboratory drive a significant fraction of its trading, though it does not quantify that fraction or date when the present level of use began. Its current hiring material gives an unusually concrete map of the stack.
The inference role spans GPU kernels, data streaming, off-the-shelf hardware and in-house ASICs and FPGAs. The pre-training role reaches across kernels, training-data loading, parallelism, networking and fault tolerance. The research role covers featurization, model architecture, training dynamics and how models make trading decisions. These are intended responsibilities, not proof that every named system is already deployed.
HRT also names three intertwined difficulties: low signal-to-noise data; distribution shift, including adaptive adversaries; and ultra-low-latency processing of millions of bursty, irregularly spaced events. That combination makes the research/hardware boundary unusually consequential. A specialized device commits engineering effort to particular computations while the research problem itself keeps moving. HRT advertises a high GPU-to-researcher ratio and a substantial modern-GPU cluster, but supplies no comparison figure.