Qumulus’ $124M Deal Highlights the Next Hurdle in AI Infrastructure. 3 Minute Read. As workloads shift from training to production inference, idle resources can be the most expensive issue, according to Getty.
Since the year 2024, companies offering AI infrastructure have been striving to accommodate a greater number of GPUs on their premises. They are confronted with a new challenge: finding ways to keep them occupied. According to QumulusAI, it has secured over $124 million in three-year AI infrastructure deals related to Nvidia Blackwell implementations, such as a partnership with AI cloud service provider Hyperbolic.
The agreements focus on inference tasks, where the economic factors typically rely more on the effective utilization of GPU resources rather than acquiring them. “Ensuring the largest and most adaptable clusters were the main concern,” QumulusAI CEO Mike Maniscalco explained to Data Center Knowledge. Today, an increasing number of customers are keen on deploying models in large-scale production while still desiring the adaptability to perform minor training or adjustments on the same infrastructure.
The alteration in consumer behavior extends far beyond AI cloud providers.
