QumulusAI backlog triples to $515 million on 2,000 Nvidia B300 deal

QumulusAI's contracted backlog nearly tripled to $515.0 million in one quarter. The GPU cloud provider, which listed on Nasdaq in July, said on Nov. 10 that remaining performance obligations rose to $515.0 million at the end of September from $173.1 million three months earlier.

The jump came from $240.9 million of new three-year take-or-pay agreements covering more than 2,000 Nvidia Blackwell B300 GPUs in two dedicated clusters. Cumulative signed customer contract value reached $528.7 million. Take-or-pay terms put the volume risk on the customers, not on QumulusAI. That is the promise. The caveat: three-year commitments from a young customer base remain untested through a pricing downturn.

How fast is revenue actually growing?

Third-quarter revenue was $14.8 million, up from $6.7 million in the second quarter and $3.1 million a year earlier. Compute revenue was $14.1 million, or about 96% of the total. The remainder came from crypto mining and mining hosting, a business the company says is shrinking. Gross margin, calculated by MLQ.ai as revenue less cost of revenue excluding depreciation, reached 72%, up from 67% in the second quarter and 47% a year ago.

Adjusted EBITDA came in at $3.7 million, the first positive quarter in the company's history. The company did not give guidance in the release. It said it would do so on its conference call.

What is hiding inside the operating loss?

The operating loss widened to $41.6 million from $7.7 million in the second quarter. The driver was not operations. General and administrative expense included $26.4 million of common stock issued for professional services tied to the direct listing. Depreciation and amortization rose to $14.9 million.

The net loss of $14.4 million was cushioned by $32.3 million of non-cash gains on the fair value of a convertible note, a related option and warrants. Strip those gains out and the picture darkens considerably. Compare the net loss of $14.4 million against $22.8 million in the prior quarter and $9.5 million a year earlier, and the trajectory is uneven at best.

Who is paying for the GPU build-out?

The expansion is financed mostly through customer prepayments and leases rather than equity. At Sept. 30, QumulusAI carried $79.0 million of deferred revenue, $122.5 million of finance lease liabilities, $32.5 million of loans through the USD.AI protocol, roughly $61 million of notes payable and a $32.0 million convertible note. Against that stack sat $21.9 million of cash and $9.2 million of restricted cash.

Interest expense rose to $5.4 million from $0.5 million a year earlier. That tenfold increase deserves attention. Lenders have pulled back from GPU-backed debt elsewhere in the sector, as MLQ.ai reported in its coverage of CoreWeave's 16% bonds and a $40 billion SpaceX chip lease. QumulusAI's leverage model depends on prepayments continuing to arrive on schedule.

What capacity is actually energized?

Active capacity reached 8 megawatts. Another 1.2 megawatts is contracted in Denver for the first quarter, and 0.5 megawatt in Columbia, South Carolina, for December. These are firm commitments with dates attached, not roadmap statements.

"We sell the capacity before we energize it, we deploy into pockets of existing power, and we bring it online in months, not years," Chief Executive Michael Maniscalco said in the release. The claim is consistent with the backlog math. The company offered no evidence on how long the sell-before-build model holds once larger clusters, and larger customers, enter the picture.

Larger rival CoreWeave also reported third-quarter results on Nov. 10, giving investors a same-day comparison across the GPU cloud sector. QumulusAI has shown it can sign contracts and cross into positive adjusted EBITDA at 8 megawatts; whether it can service $515 million of obligations on $21.9 million of cash and rising interest costs is the question its next quarter must answer.