QumulusAI announced more than $124 million in customer subscriptions under three-year GPU-as-a-service agreements with Hyperbolic and another unnamed AI inference platform. The contracts include nearly $21.9 million in upfront customer commitments and support deployments totaling 1,280 NVIDIA Blackwell GPUs. The company said the agreements establish long-term recurring revenue while validating demand for infrastructure optimized specifically for AI inference workloads.
The deployments will utilize 160 Lenovo and Supermicro bare-metal servers equipped with NVIDIA B300 and B200 Blackwell GPUs. Cisco Nexus One networking will provide the cluster fabric connecting the systems. According to QumulusAI, the infrastructure has been engineered specifically for large-scale inference rather than model training, with CPU cores, memory, and storage configured to match the requirements of production inference workloads.
QumulusAI said its workload-specific architecture can reduce AI inference costs by approximately 20% compared to standard reference designs. One of the customers, Hyperbolic, operates a cloud platform that provides GPU infrastructure for AI startups, research organizations, and enterprises running training, fine-tuning, and inference workloads. The deployments are intended to support open-source AI applications including deep-research agents, automated coding systems, and other asynchronous AI services that require high-throughput, low-latency infrastructure.
• More than $124 million in contracted customer subscriptions over three-year terms
• Nearly $21.9 million in combined upfront customer commitments
• Deployment of 1,280 NVIDIA Blackwell GPUs across 160 servers
• Systems built using NVIDIA B300 and B200 GPUs from Lenovo and Supermicro platforms
• Cisco Nexus One selected as the AI cluster networking fabric
• Infrastructure optimized specifically for AI inference rather than model training
• QumulusAI claims approximately 20% lower inference costs versus standard reference architectures
• One disclosed customer is Hyperbolic, an AI cloud platform serving startups, researchers, and enterprises
“AI infrastructure can no longer be built using one-size-fits-all designs,” said Mike Maniscalco, CEO of QumulusAI. “Inference workloads have very different performance and economic requirements than model training environments. By tuning infrastructure to the workload itself, we can improve utilization, reduce costs, and accelerate deployment timelines for customers operating at production scale.”
| Company Profile: QumulusAI | |
|---|---|
| Company | QumulusAI |
| Headquarters | The Biltmore Innovation Center, Tech Square (Georgia Tech) — Atlanta, Georgia, USA |
| Leadership | Mike Maniscalco (CEO) |
| Business Focus | Vertically integrated, distributed AI cloud infrastructure & GPU-as-a-Service (GPUaaS) |
| Core Strategy | Inference-first, demand-led infrastructure optimized to reduce open-source LLM inference costs by ~20% |
| Latest Backing | $500M non-recourse financing facility via USD.AI (secured late 2025) |
| Milestone Deal | $124M+ in 3-year AI infrastructure subscriptions (including $21.9M in upfront commitments) |
| Hardware Scale | 1,280 NVIDIA Blackwell GPUs provisioned across 160 bare-metal clusters |
| Hardware Specs | NVIDIA B200 SXM (192GB vRAM) and NVIDIA B300 SXM6 (288GB vRAM) configurations |
| Server Partners | Supermicro (B200 deployment) and Lenovo (B300 deployment) |
| Networking Platform | Cisco Nexus One (Cluster fabric management) |
| Key Customers | Hyperbolic and other major open-source inference platforms targeting agentic AI and automated coding |
🌐 Analysis
While legacy hyperscalers typically deploy standardized, over-provisioned clusters meant to handle massive foundation model training, QumulusAI strips out the virtualization hypervisors and rightsizes the supporting hardware—specifically tuning CPU core counts, local NVMe storage, and system memory to match the exact mathematical bounds of large-scale open-source LLM inference. By eliminating hypervisor overhead and matching 8-GPU Blackwell baseboards (NVIDIA B200 and B300 SXM configurations) with optimized host processors like the Intel Xeon 6767P and Platinum 6960P series, the company achieves an estimated 20% reduction in production inference costs. Furthermore, their “hyperdistributed” physical deployment strategy focuses on building smaller, modular footprints near municipal tech hubs rather than chasing gigawatt-scale rural data center campuses.
The selection of Cisco Nexus One as the cluster fabric also highlights the expanding role of Ethernet-based AI networking architectures. As AI deployments scale, infrastructure providers are seeking ways to balance GPU performance, networking efficiency, and deployment costs. QumulusAI’s focus on workload-specific configurations aligns with a broader industry movement toward specialized AI infrastructure stacks optimized for particular workloads rather than relying exclusively on standard reference architectures.