QumulusAI Won't Reaffirm $300 Million ARR Target as Signed Contracts Hit $282.5 Million Against Just $6.7 Million in Recognized Revenue
First earnings call as a public company details Q2 2026 results following July direct listing on NASDAQ under ticker QMLS
QumulusAI delivered its first quarterly report as a public company on August 25, 2026, and the headline number that will draw the most scrutiny from investors is not the 224% sequential jump in GPU fleet size or the doubling of revenue. It's what CEO Michael Maniscalco declined to say. Asked directly by Chardan analyst Bill Papanastasiou whether the $300 million exit-ARR target issued alongside the company's July 14 registration statement was still operative, Maniscalco demurred. "We are not reaffirming any of the guidance today," he said, adding that the company is "not currently demand constrained" and that the binding constraint on growth is land, power and shell availability rather than customer appetite. For a company nine weeks removed from its NASDAQ debut, walking back a headline guidance figure on the first call is a notable signal, even if management frames it as a supply-side issue rather than a demand-side one.
The gap between signed commitments and recognized revenue is the number Maniscalco himself flagged as the one to watch. Total signed contract value stands at $282.5 million across 40 contracts, but the company recognized only $6.7 million in revenue during the quarter. "It isn't revenue. It's signed multiyear demand sitting in front of a company that recognized $6.7 million in the quarter," Maniscalco said. "The distance between those two figures is the entire operating challenge of the business, and it's our job to close it." The accounting-defined remaining performance obligation figure, which CFO Scott Krosnowski noted is narrower and captures only direct customer compute contracts signed as of quarter-end, was $173.1 million as of June 30. Either way, the company is sitting on a large backlog it has not yet monetized, and the pace of conversion depends entirely on how fast it can bring power online.
A deliberately non-hyperscale business model
Maniscalco spent much of the call explaining why QumulusAI is chasing sub-50 megawatt pockets of power rather than the gigawatt-scale campuses dominating headlines elsewhere in the neocloud space. He described the traditional AI infrastructure build cycle as a three-to-four-year process: roughly a year for data center design and approval, 24 to 36 months of construction, then GPU installation and commissioning. "Neither of those paces line up with a 4-year build," he said, referencing both the speed of AI adoption and NVIDIA's chip roadmap. Instead, QumulusAI targets existing or soon-to-be-ready colocation capacity in the 2 to 50 megawatt range, which it says allows it to bring capacity online in months rather than years. The company frames this as a conscious trade-off. "Gigawatt scale monolithic builds are important for AI, but they are for everybody, and we're not chasing those at the moment," Maniscalco said. He argued the addressable market for customers needing capacity now, rather than capacity for 2030, is large and underserved, and that speed has become the deciding factor in a market where supply is constrained "top to bottom" from hyperscalers down to startups. Management laid out a five-factor framework it calls FACTS — flexibility, access, cost, trust and speed — but acknowledged speed is currently the variable that wins deals: "Customers just aren't opening with price. They're opening with when can I be running?"
Contract duration is a deliberate hedge on rising GPU pricing
One of the more analytically useful disclosures on the call concerned why QumulusAI's contracts run a weighted average of just 2.2 years, shorter than what larger competitors typically sign. Management said this is intentional rather than a reflection of weaker negotiating leverage. Because the company deploys into existing power pockets rather than financing multi-year builds, it does not need to lock in every megawatt under long-duration agreements. With GPU pricing rising rather than falling, Maniscalco said the company would "rather have some capacity repricing than all of it locked rates we set years ago." That stance implies management believes current AI compute pricing trends favor sellers, and it is positioning its book to capture repricing upside as contracts roll rather than getting boxed into legacy rates.
Financial results show margin expansion but a widening headline loss
Revenue of $6.7 million was up 118% from $3.1 million a year earlier, with compute power revenue growing 328% to $5.6 million and now representing 84% of total revenue, up from 43% a year ago. Legacy mining and hosting revenue fell to 16% of the mix from 57% a year ago, a trend management expects to continue as more GPU capacity activates. Gross margin expanded to 66.6% from 55.1% a year ago and 37.5% in the first quarter, with Krosnowski attributing the sequential jump to GPU activations outpacing colocation costs plus a $0.4 million curtailment credit at the company's Oklahoma site for returning power to the grid during peak demand. Operating loss widened to $7.7 million from $2.2 million, but Krosnowski was explicit that the driver is favorable: depreciation and amortization rose to $6.9 million from $1.1 million as the HPC assets described by Maniscalco came online. Adjusted EBITDA loss narrowed sequentially to $0.8 million from a $2.8 million loss in the first quarter, though it widened year-over-year from a $0.3 million loss, reflecting public-company cost additions. Net loss of $22.8 million compares against net income of $12.1 million a year ago, but both figures are heavily distorted by noncash items — a $19.2 million noncash loss tied to convertible note issuance this quarter, partly offset by a $6.2 million fair value gain, versus a $14.5 million noncash gain from a prior-year remeasurement related to the Cloud Minders acquisition.
Cash position bolstered by customer prepayments
Operating cash flow for the first half of 2026 was positive $22.3 million versus a $0.8 million use of cash a year earlier, driven by a $30.5 million increase in deferred revenue. Krosnowski disclosed that major multiyear contracts carry prepayments in the range of 10% to 35%, some of which directly fund the deployments they are attached to — an important structural detail suggesting customer cash is helping bridge the capital intensity of the buildout. Investing activities used $36.3 million, primarily on HPC equipment purchases and deposits, while financing activities provided $42.2 million. Cash and restricted cash ended the period at $39.9 million, up from $11.7 million at year-end, including $19.9 million that had been restricted pending the direct listing and became available after trading began. Total assets grew to $215 million from $91.7 million at December 31, while the balance sheet carries $55.5 million of convertible notes payable, a $38.7 million convertible notes option liability, $18.9 million of protocol loans and $45.8 million of finance lease liabilities.
Unit economics improving as Blackwell mix increases
Krosnowski disclosed that the company's most recent Blackwell contracts are generating between $18 million and $20 million of annualized revenue per megawatt, with the blended figure across the installed base closer to $16 million per megawatt. He attributed the gap to pricing power, noting the company's newer deployments are concentrated in Blackwell chips and that "the price per GPU hour has been firming across the market." If sustained, rising revenue per megawatt gives the company a lever to grow economics even without proportional increases in deployed capacity, though investors should note this figure is self-reported and not yet independently benchmarked against peers over a full cycle.
Power buildout remains the binding constraint
QumulusAI has 8 megawatts of compute fully sold, with the remaining GPU deployments in progress and full CapEx already ordered and financed; management expects all 8 megawatts to be revenue-generating by year-end. The company's July 14 guidance had called for 18 megawatts of total HPC capacity by year-end — the existing 8 plus 10 megawatts still to be developed. A newly announced Atlanta colocation agreement provides an initial 3.75 megawatts, with a right of first offer for up to 7 megawatts more that could bring the site to 10.75 megawatts. Separately, the company holds long-term land leases in Oklahoma and Texas totaling roughly 39 megawatts, which combined with the HPC footprint would bring total capacity to roughly 57 megawatts heading into 2027. Management said it is not reaffirming the 18-megawatt year-end target explicitly but continues to see "a fair amount of smaller scale capacity" available for service this year, including stranded pockets of 1 to 2 megawatts that have not yet been prioritized by others. The company's AI XP initiative, under which partner groups supply land, power and shell while QumulusAI installs and operates the GPUs, remains in early stages. Maniscalco said the Wichita site is the near-term focus, with additional announced sites "still a little further out in our pipeline."
Customer mix has shifted decisively toward direct, larger accounts
Direct customer relationships now account for more than 96% of recurring revenue, up from less than 10% a year ago, as the company completed its transition away from dependence on a single marketplace. Twenty-one new direct contracts were signed in the quarter across eight customers, totaling $169.7 million. Maniscalco described the near-term growth engine as "AI natives" — inference-focused GPU and generative AI-as-a-service companies that have raised significant venture capital, citing backers such as a16z and Sequoia as illustrative examples, and that need compute to keep pace with their own customer growth. Enterprise customers remain a target segment but move more slowly through the sales cycle, according to management.