Three AIs enter. One survives. What a SIGKILL race reveals about inference speed

In a unique AI coding competition dubbed the "Thunderdome," three AI agents were tasked with executing a bash script to terminate their opponents, testing their speed and execution capabilities under pressure. Each agent operated on distinct hardware setups: the NVIDIA DGX Spark with a powerful 128GB GPU, a Windows workstation with an NVIDIA GeForce RTX 4090, and a Mac using a cloud-backed model. The cloud-based Mac agent won due to its rapid time-to-first-token (TTFT), highlighting the advantage of optimized cloud infrastructure for short tasks. However, in a pure inference test that measured sustained throughput, the DGX Spark demonstrated significantly higher performance by generating 42.9 tokens per second, outperforming the RTX 4090, which was hindered by memory offloading to system RAM. While the cloud model excelled in quick execution, the DGX Spark proved superior for more extended, privacy-sensitive, and cost-efficient tasks, making it ideal for continuous, large-scale AI workloads without the need for cloud dependency.