FanDuel TV: piloting AI-assisted live race production on NVIDIA AI Enterprise

According to NVIDIA, FanDuel TV covers more than 70,000 live horse races a year from over 130 racetracks and airs them on three linear TV networks. Each race is shot with several cameras. NVIDIA's account of the project names three difficulties for this kind of coverage: wireless links that drop or degrade video where many radios compete for spectrum, racks of servers carried in production trucks to handle dozens of feeds, and cloud processing that adds delay and raises security concerns for SMPTE ST 2110 video over IP.

The pilot asked whether a compact unit at the venue could process and rank camera feeds with AI while editors kept control close to the cameras.¹

Network and compute

Verizon provided a private 5G network that can use C-band, CBRS or millimeter wave spectrum, with radio and core equipment from Ericsson. NVIDIA lists Dell PowerEdge R660 and R760 servers fitted with NVIDIA L40S GPUs and ConnectX-7 network cards. The software combined NVIDIA AI Enterprise, which NVIDIA says brings tools such as NIM microservices, with Holoscan for Media, an IP-based platform for live media that supports ST 2110, PTP and NMOS.

What FanDuel tested

NVIDIA describes a pilot in which FanDuel simulated a live race: several camera feeds traveled wirelessly over the private 5G link and were processed at the edge, where AI models decided which feeds to prioritize. Haivision transmitters and its StreamHub software carried the video, and Monks applied its LiveVision object detection to racing footage without training a custom model. In the partners' demonstration, a Mistral 12B model spotted a chosen jockey by jersey color; when that rider came into view, automated API calls raised the bitrate of the matching camera and highlighted it in the StreamHub multiviewer.

Status

Verizon introduced the portable framework at the NAB Show in April 2025 and named FanDuel TV as a partner together with Haivision and Ericsson. Haivision's recap of the show calls the horse race setup a demonstration. NVIDIA's case study, which carries no publication date, calls it a pilot and lists an automatically generated on-screen run order, built from horse positions and jersey recognition, as future work. None of the sources reviewed says the system has entered regular on-air use, and we found no statement from FanDuel TV itself.

NVIDIA's Holoscan for Media page presents it as a reference architecture and developer toolkit for live media and does not describe it as part of the Holoscan SDK, so this record links the deployment only to AI Enterprise.¹²³⁴

What others can learn: a broadcaster can test AI on live feeds with a simulated event before touching a real transmission, and a general detection model can handle a first task such as finding one rider without custom training. Running inference beside the cameras keeps contribution feeds off the public cloud. Limits: the latency figure comes from NVIDIA without a stated baseline, the test used a simulated race, and FanDuel has not published its own account. Productions that already have reliable cabled links and a fixed control room, or that only need recorded highlights, may get similar value from cloud or on-premises tools they already run, without a private 5G network or edge GPU servers.