Firm rents four Nvidia H200s to test '80x cheaper' DeepSeek claim (www.tomshardware.com)

The Call Center Doctors recently tested Nvidia's H200 AI GPUs to run DeepSeek V4.1 Flash, claiming it could be "80x cheaper" than using Anthropic's Claude Opus 5.5 for similar work. The consultancy found that operating costs with DeepSeek's API ranged from $3,500 to $7,000 for a month, significantly lower than the $140,000 it would cost using Claude. However, despite the promising claims, the tests revealed inefficiencies, with the server primarily engaged in rereading stale text rather than generating new tokens, leading to suboptimal handling of workloads. This testing underscores the cautious optimism surrounding DeepSeek's capabilities. While its API pricing presents a cost-effective alternative to Claude, the performance during real-world usage raises questions about its practical viability. The observed limitations in generating new codes and the costs involved in renting GPU resources highlight the complexity of transitioning to AI models, where raw computational power doesn't always translate to performance efficiency. As the AI community examines DeepSeek's potential, these findings provide critical insights into operational challenges and cost-effectiveness in leveraging advanced AI models.