QumulusAI二季度GPU部署超两倍算力扩张受电力容量制约

Senior News Writer,Data Center Knowledge

August 27, 2026

5 Min Read

QumulusAI more than tripled its deployed GPU fleet in the second quarter as it brought contracted computing capacity online and signed $169.7 million in new customer agreements.

The Atlanta-based AI infrastructure provider had 3,088 GPUs deployed at the end of June, up from 952 three months earlier. Revenue more than doubled year over year to $6.7 million, with compute revenue accounting for about 84% of total revenue, compared with 43% a year earlier.

QumulusAI signed 21 new direct AI compute contracts during the quarter, bringing cumulative signed customer contract value to $282.5 million across 40 multiyear agreements. The $169.7 million signed during the quarter represents aggregate expected take-or-pay value.

QumulusAI ended the quarter with 8 MW of HPC capacity under executed lease and colocation agreements, all of it committed to customers, with final GPU deployments underway. The company expects the full 8 MW to generate revenue this year. It has also contracted for up to 3.75 MW in metropolitan Atlanta that it expects to bring into service in the fourth quarter.

QumulusAI’s deployment pace now depends on finding powered data center capacity that can be brought online on customers’ timelines, CEO Michael Maniscalco told Data Center Knowledge.

“It’s really the intersection of power and deployable data center capacity,” Maniscalco said. “The harder part is finding powered capacity that can be ready for service on the timelines our customers need.”

The company is targeting smaller blocks of capacity that can be deployed and expanded as demand grows. Maniscalco said QumulusAI often looks for 1 MW to 3 MW initially, with a path to add capacity at the same site.

“Rather than waiting years for a massive new campus, we can secure smaller pockets of available power and capacity, often 1-3 MW initially, deploy quickly, and then expand from there,” he said.

QumulusAI Targets Smaller Blocks of Powered Capacity

Steven Dickens, CEO and principal analyst at HyperFRAME Research, described QumulusAI as an example of the “sub-hyperscaler build-out” underway across the industry.

“The model they have for more modular data center deployment is behind this, especially when it comes to power activation,” Dickens said.

QumulusAI calls its approach “hyperspeed,” referring to its strategy of deploying AI compute through a distributed network of data center sites. Dickens said execution speed will be a competitive factor for AI infrastructure providers.

“Across the industry, success will be defined by AI providers who can execute, and at speed,” he said.

Crusoe is pursuing a similar modular approach at a larger scale. Its Spark units are prefabricated AI data centers designed to be deployed individually or combined into larger clusters. Crusoe has expanded a deployment with Redwood Materials from four to 24 modular data centers at a site supported by a 12 MW, 63 MWh microgrid.

Crusoe says its Spark units can be deployed in as little as three months, although broader project timelines still depend on site readiness, power and interconnection conditions. Its approach differs from QumulusAI’s in that Crusoe manufactures the modular data center infrastructure itself, while QumulusAI focuses on securing available powered capacity and deploying GPUs into it.

The comparison illustrates two approaches to the same problem: putting AI compute into service without waiting for a conventional large-scale data center project to reach completion.

The neocloud market is scaling alongside demand for dedicated AI compute. CoreWeave reported $2.6 billion in second-quarter revenue and $9.4 billion in capital expenditures, and expects to spend $35 billion to $39 billion this year as it expands data center capacity. Nebius reported $582.3 million in second-quarter revenue, up 454% year over year, while other specialized providers including Crusoe and Lambda are also expanding their AI infrastructure footprints.

Blackwell Lifts Revenue Density, but Raises the Capital Stakes

QumulusAI’s second-quarter results show the financial payoff when capacity reaches customers. Compute revenue reached $5.6 million, up from $1.3 million a year earlier.

The company says its newest Blackwell contracts are generating $18 million to $20 million in annualized revenue per megawatt, compared with just over $16 million per MW across its installed base.

Maniscalco attributed the higher revenue density to the economics of Nvidia Blackwell GPUs, firm pricing for GPU compute and customers’ willingness to pay for capacity that can come online quickly.

“Customers place a premium on capacity that can actually come online quickly,” he said.

QumulusAI purchased 1,632 Nvidia Blackwell B300 GPUs during the quarter to meet customer demand.

The results also show the capital required to support that growth. QumulusAI’s property and equipment increased to $44 million at the end of June from $12.5 million at the end of 2025. Finance right-of-use assets increased to $47.9 million.

QumulusAI reported a $7.7 million operating loss in the second quarter, compared with $2.2 million a year earlier. Depreciation and amortization increased by $5.8 million, primarily because of its expanded HPC infrastructure.

For the first half of 2026, QumulusAI generated $22.3 million in operating cash flow, aided by $30.5 million in deferred revenue as customers paid ahead of delivery.

The higher revenue density of newer GPU deployments can improve the economics of each megawatt, but it does not remove the capital and execution requirements of building the fleet. Customer prepayments can help fund expansion, while QumulusAI still has to finance GPUs, secure data center capacity and maintain enough utilization to turn that infrastructure into returns.

QumulusAI is continuing to pursue additional capacity, with a preference for sites that can be ready for service this year. Its Atlanta agreement provides up to 3.75 MW initially, along with a right of first offer on as much as another 7 MW at the same site.

The company has set a year-end 2026 target of 18 MW, consisting of 8 MW of active HPC power and 10 MW under development. It says it has a line of sight to 2.5 GW by the end of 2027.

The 2.5 GW figure is not operating or contracted capacity. It represents management’s view of development opportunities it may be able to access, making power procurement, site conversion, financing and customer delivery the central tests of the strategy.

Dickens cautioned that one quarter represents only one point in QumulusAI’s trajectory. He said the company’s progress over the past year and its latest results suggest its deployment model is gaining traction.

The larger test will be whether QumulusAI can turn its pipeline into financed, powered and operational capacity quickly enough to sustain customer pricing, utilization and returns.

About the Author

Senior News Writer, Data Center Knowledge

Shane Snider is Senior News Writer at Data Center Knowledge, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. His reporting focuses on the operational, economic, and environmental forces reshaping digital infrastructure, including AI factories, utility constraints, liquid cooling, renewable energy procurement, and next-generation data center architectures. He has won recent Azbee awards for news series and government reporting. Based in Raleigh, North Carolina, Snider covers how hyperscalers, utilities, chipmakers, and infrastructure providers are responding to the rapid rise of AI workloads and global compute demand.

You can reach Shane at [email protected] or on LinkedIn.