Reinforcement learning (RL) has recently been explored as an adaptive alternative to hand-designed congestion control, yet most reported evaluations remain confined to simulated or moderate-bandwidth environments. This paper studies the scaling behavior of an Aurora-style Proximal Policy Optimization (PPO) congestion control policy when it is deployed on a real high-bandwidth, high bandwidth-delay product (BDP) wide-area network (WAN) path with a 10 Gbps interface budget. The purpose is not to claim a new state-of-the-art controller, but to identify how a simulator-trained PPO policy behaves when transferred to multi-gigabit operation. We integrate the policy through a user-space PCC shim and analyze transport-level logs, including achieved receive rate, loss dynamics, rate oscillations, and available round-trip time (RTT) indicators. The results show a persistent gap between nominal link capacity and achieved single-flow goodput: the analyzed run reaches a mean receive rate of 49.6 Mb/s and a peak of 438.0 Mb/s, corresponding to 0.50% average and 4.38% peak utilization of the 10 Gbps budget. Additional TCP CUBIC and TCP BBR baselines on the same path reached up to 9.91 Gb/s and 9.72 Gb/s with four flows, confirming that the route itself can sustain multi-gigabit throughput. The rate distribution is heavy-tailed, with short high-rate episodes followed by over-shoot-collapse dynamics, bursty loss, and transient delay inflation. These findings indicate that scale-aware training, more robust reward normalization, and lower-overhead pacing are needed before PPO-driven congestion control can generalize reliably to high-BDP WAN deployments.