AI cloud infrastructure provider QumulusAI has secured more than $124 million in customer commitments through new three-year subscription agreements with open-access AI cloud platform Hyperbolic and another undisclosed artificial intelligence inference provider.
The contracts include nearly $21.9 million in upfront customer commitments and will support the deployment of 1,280 Nvidia Blackwell GPUs across dedicated AI infrastructure clusters.
Blackwell-Powered AI Infrastructure
Under the agreements, QumulusAI will deploy 160 bare-metal servers supplied by Lenovo and Supermicro, equipped with Nvidia B300 and B200 Blackwell GPUs.
The systems will be interconnected using Cisco Nexus networking technology to create high-throughput, low-latency AI clusters designed for production-scale inference workloads.
Unlike traditional hardware sales, the contracts are structured as GPU-as-a-Service subscriptions, providing recurring revenue for QumulusAI while allowing customers to access AI computing resources through predictable operating expenses.
Focus on Inference Rather Than Training
QumulusAI is positioning itself around what it describes as an inference-first infrastructure strategy.
While many AI infrastructure deployments have historically been optimized for model training, the company believes the next phase of AI growth will be driven by large-scale inference applications, including generative AI services, autonomous agents, deep-research platforms and AI coding assistants.
According to the company, traditional AI infrastructure often relies on generic architectures that include excess CPU resources, memory and storage capacity, increasing costs without improving inference performance.
QumulusAI said it has redesigned its infrastructure to align CPU, memory and storage resources more closely with actual inference workload requirements.
Targeting Lower AI Operating Costs
The company estimates that its optimized architecture can reduce AI inference costs by approximately 20% compared with conventional infrastructure configurations.
Rather than maximizing hardware specifications across every component, the design focuses on improving utilization rates and reducing underused resources surrounding GPU clusters.
Chief Executive Officer Mike Maniscalco said the industry is moving beyond the initial phase of GPU scarcity.
“AI infrastructure can no longer be built using one-size-fits-all designs.”
“Inference workloads have very different performance and economic requirements than model training environments.”
According to Maniscalco, infrastructure providers must increasingly focus on efficiency, utilization and economics as enterprises scale AI services into production environments.
New Infrastructure Model Emerges
The company argues that AI inference is becoming a distinct infrastructure category with different requirements from model training environments.
Training systems are typically designed for intensive bursts of computation and large-scale data movement, while inference platforms prioritize predictable latency, continuous utilization and cost efficiency over extended periods.
QumulusAI’s approach combines multi-year GPU subscriptions, distributed infrastructure deployments and workload-specific optimization to create what it describes as an inference-focused computing fabric.
The company believes future AI infrastructure purchasing decisions will increasingly be based on metrics such as utilization rates and cost per inference rather than total GPU counts alone.
Expansion Continues
The latest agreements build on QumulusAI’s broader infrastructure expansion strategy.
Earlier this year, the company secured a $45 million convertible note facility to support GPU purchases and data center development.
QumulusAI also received approval for a modular data center project in Denton, Texas, which is expected to provide approximately 20 MW of capacity.
The company currently works with colocation providers across several US markets, including Atlanta, Kansas City, Philadelphia, Denver and Brooklyn.
As AI adoption expands and inference workloads become a larger share of overall computing demand, QumulusAI is betting that infrastructure efficiency and workload optimization will become key competitive advantages in the rapidly evolving AI cloud market.
Source: DataMagz