QumulusAI Raises $124M, Bets on GPU Efficiency

QumulusAI has secured over $124 million in three-year customer subscriptions with Hyperbolic and another AI inference platform, covering 1,280 Nvidia Blackwell GPUs across 160 bare-metal servers. Nearly $21.9 million was committed upfront, giving the neocloud provider immediate working capital. The company is shifting AI infrastructure focus from GPU scarcity to efficiency, claiming its inference-first architecture cuts costs by roughly 20% by rightsizing CPU, memory and storage around actual workload behavior. QumulusAI argues inference now demands purpose-built infrastructure distinct from training environments, measured by cost per query rather than peak performance.