CoreWeave announced on 20 August a multi-year agreement with Hudson River Trading, under which the quantitative trading firm will run its AI research and model-development workloads on CoreWeave's platform. The release's subheading describes it as a "multi-billion dollar deal." That phrase is the entirety of the financial disclosure.
What is specified
The technical detail is concrete where the commercial detail is not. HRT gets clusters of NVIDIA Vera Rubin NVL72 and NVIDIA HGX B200 GPU systems, high-throughput Direct Connect networking, and Spectrum-X Ethernet. CoreWeave positions the win against its own benchmarking record — MLPerf results, Platinum ranking in SemiAnalysis ClusterMAX 1.0 and 2.0, and a first-place inference ranking from Artificial Analysis.
What the common framing gets wrong
Two things, in opposite directions. Some coverage reports that CoreWeave declined to quantify the deal — but the company's own release does put an adjective on it, "multi-billion dollar," in the subheading. What is missing is not the characterisation but the figure: no contract value, no term in years, no committed capacity, no minimum. "Multi-billion" spans $2bn and $9bn equally well, and over an undisclosed number of years it does not convert into an annual revenue figure at all. The second error runs the other way: Vera Rubin is not deployed silicon. Describing HRT as "powered by Vera Rubin" states a future configuration in the present tense, with no delivery date in the release.
Why a trading firm
The customer type is the interesting part. Quantitative trading firms have run large-scale machine learning for well over a decade on their own hardware, for latency and secrecy reasons. One of them contracting out research training to a neocloud is a signal about the economics of owning frontier-class clusters — even for firms that can comfortably afford to.