QumulusAI, Inc.

Hello, everyone. Thank you for joining us and welcome to the Cumulus AI second quarter 2026 earnings call. At this time, all participants are in listen-only mode. After today's prepared remarks, we will host a question and answer session. If you would like to ask a question, please press star 1 to raise your hand. To withdraw your question, please press star 1 again. Today's call is being recorded. I will now hand the conference over to Zachary Graeve, Head of Investor Relations. Zachary, please go ahead.

Good afternoon and welcome to Cumulus AI's second quarter 2026 earnings call. I'm Zachary Graeve, Head of Investor Relations, and with me today I have Mike Maniscalco, our Chief Executive Officer, and Scott Krosnowski, our Chief Financial Officer. Thank you. I'd like to remind everyone that certain statements made on this call may constitute forward-looking statements. These statements include those about our strategy, capacity and deployment plans, customer agreements, and future performance. Please be advised that these forward-looking statements are covered under the safe harbor provisions of the Private Securities Litigation Reform Act of 1995 and that Cumulus AI avails itself of the protections of the safe harbor for these statements. forward-looking statements are based on management's current expectations and assumptions as of today and are subject to risks and uncertainties that could cause actual results to differ materially from those stated or implied those risks are described in our risk factor section of the registration statement on form s1 as amended file number three three three two nine two five one four filed with the Securities and Exchange Commission Those factors may be updated in our subsequent filings, so please read these reports and future filings that cumulusAI will make with the SEC. CumulusAI disclaims any obligation to update or revise any forward-looking statement except as required by law. Please note that we may refer to non-gap measures in today's call. Any non-gap measure will be paired with the most directly comparable gap measure, and reconciliations are included in today's release. Now with that, I'll turn the call over to Mike.

Thanks, Zachary, and thanks to all of you for joining us. This is our first call as a public company, so I'll start the way I did on the NASDAQ stage in July with an expression of gratitude for our customers, employees, shareholders, board members, and partners. A number of them committed to this company well before there was much evidence to go on. We're thankful for that and excited about the future, which is the reason the rest of this call will focus on executing on the opportunity ahead. Let me tell you how I got here. Thank you so much for joining us. If software was going to become this easy to build, then software wasn't the constraint anymore. Whatever sat underneath it was. And what sits underneath the AI tools are the models and the compute to power them. The heavy AI infrastructure, power, data centers, and GPUs were the enabler of every good thing I was watching happen. So less than eight months after I left the infrastructure space, I came back. And I came back through Atlanta. I was in town visiting Georgia Tech, my alma mater, where I met the Cumulus AI team. Cumulus AI's own story runs on a parallel track. Two companies, both founded in 2019. One built data center assets and power. The other built blockchain managed services. In December of 2022, they merged into what is now cumulusAI. And in April of last year, we acquired the CloudMinders, which brought GPU as a service as a foundation. Two halves of an answer put together. Thank you for watching. Last week, we announced a co-location agreement in Metro Atlanta, our home market, for up to initial 3.75 megawatts. That agreement also carries a right of first offer for up to 7 megawatts of additional capacity, which the provider may pursue at the same site and could bring the Atlanta location to 10.75 total megawatts. Now looking back, the guidance we issued in July 14th, 2026 contemplated 18 megawatts of HPC capacity by year end. The eight we had plus 10 more to develop. Atlanta is the first 3.75 of that 10. If you add the 39 megawatts under lease, that's roughly 57 megawatts in our combined footprint going into 2027. That's real opportunity ahead of us. Now to the quarter. On July 16th, we began trading on the NASDAQ global market under the ticker QMLS. We did it as a direct listing. Our registration statement was declared effective on July 14th, and we rang the opening bell of the NASDAQ market site on July 29th. On July 17th, the day after we began trading, we were approved as an NVIDIA cloud partner. Two things in the same week, and both are about access. One to the capital markets, the other to the supply chain that the capital buys from. Thank you for joining us. A 224% increase from Q1. It's worth highlighting that most of this was compute we had already sold coming online and revenue more than doubled year over year. Additionally, gross margin expanded to 66% from 55.1%. Scott will talk about that later in today's call. But those three numbers tell one story. AI compute we had already sold came online. and whenever compute comes online, that's when the economics of this business show up. On the commercial side, we signed 21 new direct customer contracts totaling $169.7 million in the second quarter. That brings our total signed contract value to $282.5 million across 40 contracts with a weighted average term of 2.2 years, shorter than what you'll see from the largest players, but that's deliberately so. I'll come back to why later. Thank you for joining us. because it's the real number I'd watch. It isn't revenue. It's signed multi-year demand sitting in front of a company that recognized 6.7 million in the quarter. The distance between those two figures is the entire operating challenge of the business. And it's our job to close it. The only way you close it is by putting compute on the floor at hyperspeed. The customer base has changed too. Direct customer relationships now account for more than 96% of our recurring revenue, up from less than 10% a year ago. We've completed the transition away from dependence on a single marketplace to direct multiyear agreements with customers themselves. With marketplaces used for what they're generally good at, tailored user experiences and filling short-term gaps between term customers as they roll on and off. And customer contracts are getting bigger. If you look at our recent announcements, the trend is clear. What we're signing now is materially larger than what we signed six months ago. As we secure more power, we expect that trend to continue. Now, I want to take a step back for a minute and talk about why we're built the way we are, because there's a lot of noise in this category. We are not a powered cell developer building sites to lease to the next hyperscaler. We are not brokering powered land to data center developments. And we're not brokering GPU clusters. We deploy and operate GPU clusters for our customers at scale. And we made a deliberate choice about how. The traditional way to build AI infrastructure goes like this. And I think this is important to understand. You do an initial data center design and seek data center development approval for about a year. Then you engineer it. And then you finance it. You spend 24 to 36 months in construction. You roll in the GPUs. You spend a few more months commissioning. And three to four years later, you've delivered a token generating cluster. Now, look at the pace that AI is moving. Then look at NVIDIA's chip roadmap. Neither of those paces line up with a four-year build. So we don't lead with four-year builds. We look for pockets of power, call it two to 50 megawatts, where the power shell already exists or can be ready for service soon, and we deploy quickly. Most of that is co-location, and it means we bring capacity online in months, not years. Now, that's hyperspeed. There is a trade in that, and I'll name it. It's gigawatt-scale campuses. Gigawatt scale monolithic builds are important for AI, but they aren't for everybody, and we're not chasing those at the moment. We are currently focused on customers who need capacity today, not those planning for 2030. That market is large and is underserved. What we get in exchange is speed, and in this market, speed is a core differentiator. But speed alone does not win. There are five things customers actually buy. We call it our FACTS framework. First, flexibility. Whether it's around location, vendors, or architectures, such as bare metal versus Kubernetes or Rocky versus InfiniBand, we provide customers that flexibility. Second is access. As I mentioned, our customers struggle to find capacity. So what we do, we provide access and capacity that's available when a customer needs it. And we do that with a fair and clear price that customers can grow into. That's our cost structure. Next is trust. We're building long-term partnerships. Through reliability and SLAs or isolation and data privacy, we're here to support our customers' needs. Lastly, it's speed. And all five matter, but one of them decides most deals right now. customers just aren't opening with price. They're opening with, when can I be running? What we see is an industry that's compute constrained top to bottom, from the largest providers down to the startups that can't get enough compute to keep pace with their customer demand. When supply can't keep pace with customer demand, speed stops being a feature. It becomes the decision. That's the near-term reason we go after small pockets of power. It is the fastest way to get a customer running. There's a second thing that comes from building this way. Because we typically deploy into pockets of existing power rather than financing for your builds, we don't need to lock in every megawatt into longer-term five-year contracts. Most of our contract value is multi-year, but part of the book comes up for renewal each year. and GPU prices have been rising, not falling, which means in a market like this, we'd rather have some capacity repricing than all of it locked to rates we set years ago. And that's deliberate. But we're also making a longer term bet on that same footprint. Today, distribution is a means to an end. Over time, we think it becomes the end itself. As workloads shift from training to inference, they get more particular about where they run. For example, budget starts to matter more. Latency begins to matter. Proximity to the end users starts to matter. So does the security and compliance posture of a specific location. So when customers start choosing in those terms, a network of sites in different places stops being an operational complexity and becomes a differentiated product. Now, we'll see how quickly that develops. In the meantime, the same footprint is doing exactly what we need it to do today, delivering supply at the highest speed. That's why we built this company, Inference First. Tokens are the output. And getting tokens to customers faster than anybody else in our business is winning today. Getting them to the right place is the business we're building towards tomorrow. Now, let me put it together. In the second quarter alone, we more than tripled our GPU fleet. We signed 21 new contracts. We gained access to the capital markets and a place inside Nvidia's partner ecosystem. Our go-to-market motion is working. We activate capacity, sign demand, and deliver the supply. Those are the three dials, and all three move this quarter. Thank you. Now I'll turn it over to Scott to walk you through our financial reports in more detail. Over to you, Scott.

Thanks, Mike. Good afternoon, everyone. I'll take revenue first, then walk down the income statement, then finish with cash and the balance sheet. First, revenue. Revenue was $6.7 million in the second quarter, up 118%, compared to $3.1 million from the same period last year. The driver is compute power, which grew to $5.6 million from $1.3 million a year ago, up 328%, or $4.3 million, as GPUs activated primarily on term compute contracts. representing 84% of total revenue in the second quarter compared to 43% from the same period last year. Our legacy lines moved as expected. Together, mining, hosting, and cryptocurrency mining were about 16% of revenue this quarter, down from roughly 39% in the first quarter and 57% a year ago. We expect this downward trend to continue as we activate more GPUs. Thank you for joining us. Thank you for joining us today. Down to the operating line. Operating loss for the quarter was $7.7 million compared to $2.2 million a year ago. The loss widened, and I want to be specific about why, because the reason is favorable. Depreciation and amortization were $6.9 million in the quarter against $1.1 million a year ago, and up $4.2 million quarter over quarter. That sequential step is the direct consequence of the HPC assets Mike described coming online and driving top line and margin growth. Adjusted EBITDA, which excludes interest, taxes, depreciation, and non-cash items, was a loss of $0.8 million compared to a loss of $0.3 million a year ago. Sequentially, the loss narrowed from $2.8 million in the first quarter. That's a non-gap measure. The most directly comparable gap measure is net loss. and a full reconciliation is in today's release. The year-over-year comparison reflects the public company and personal investments we have made. The sequential improvement reflects revenue beginning to scale against them. General and administrative expenses were 4.3 million, up 1.9 million from a year ago, driven by increased headcount, costs associated with becoming a public company and scaling up infrastructure operations. Sales and marketing expenses were $0.9 million up from $0.4 million a year ago, driven by increased headcount and costs associated with promoting and marketing our HPC services. And below the bottom line, net loss was $22.8 million compared to net income of $12.1 million a year ago. The year-over-year comparison is significantly affected by non-cash accounting in both periods. The prior year result included a $14.5 million non-cash gain related to the remeasurement of our investment in the cloud minders upon acquiring the remaining interest. In the current quarter, the largest single item is a $19.2 million non-cash loss associated with the issuance of convertible notes, which did not represent a cash outflow. The charge was partly offset by $6.2 million in non-cash fair value gains on the notes and related option. Net interest expense was $2 million, primarily related to loan draws and lease financing. Thank you for joining us. Investing activities used $36.3 million, primarily purchases and deposits of HPC and related equipment. 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. That includes $19.9 million that had been restricted pending our public listing, which became available following the direct listing on July 16th. On the balance sheet, total assets were $215 million up from $91.7 million on December 31st, driven by finance rate of use assets, property and equipment, and deposits on power equipment. On the liability side, we carry $55.5 million of convertible notes payable, a $38.7 million additional convertible notes option liability, $18.9 million of protocol loans, and finance lease liabilities of $45.8 million. Thank you for watching. In our financial statements, you'll see a related but narrower figure. Remaining performance obligations of $173.1 million as of June 30th. That's the accounting measure, direct customer compute contracts only, and only what was signed as of quarter end. Two last points. First on unit economics. Our most recent Blackwell contracts are generating between 18 and 20 million of annualized revenue per megawatt. Across our installed base, the blended figure is closer to 16 million per megawatt. The gap is pricing power. Our deployments are currently focused on the Blackwell chips. The price per GPU hour has been firming across the market, and customers will pay for GPUs in capacity that come online quickly. All three of these factors are working in our favor, and the revenue we earn per megawatt is widening as we scale. moderator. We are ready to take questions.

We will now begin the question and answer session. Please limit yourself to one question and one follow-up. If you'd like to ask a question, please press star 1 to raise your hand. To withdraw your question, press star 1 again. We ask that you pick up your handset when asking a question to allow for optimum sound quality. If you are muted locally, please remember to unmute your device. Please stand by while we compile the Q&A roster. Your first question comes from the line of Michael Donovan with Compass Point. Your line is open. Please go ahead.

Thank you, operator. Good afternoon, guys. Congrats on the progress. I thought we could discuss the AI XP opportunity. How were economics structured across Cumulus and site partners, particularly around CapEx contributions and revenue sharing? And can we get a sense of timing for bringing on initial sites of the 25 sites that you indicate in the presentation?

Yeah, thanks, Michael. I appreciate it. So I want to keep those somewhat separate because Thank you so much for joining us. Thank you for joining us. So that being said, the AIXP initiative and those sites that we announced are still in progress. That project is making good progress. They've been posting pretty good updates online. The collective groups are responsible for bringing the land, the power, the shell to us. Our responsibility at that point is to, once connectivity and power and cooling is ready, we roll in the racks and bring the GPUs online. So that's the role we play in that arrangement. The sites are Energizing in phases right now. And really, just to be a little more specific, the first site in Wichita is really the one we're focused on today. The additional sites are still a little further out in our pipelines.

That's helpful. I appreciate it. Thank you.

Your next question comes from the line of Brett Knobloch with AI Infrastructure. Your line is open. Please go ahead.

Hi guys, thank you for taking my question and congrats on the quarter. As we look at the remainder of the year, I'm just curious on the power ramp. I think previously you talked about getting up to 16 or 18 megawatts, so I guess is that still the target? And then the $16 million revenue per megawatt figure of the existing base, at what point would you expect that to start to show up or all those contracts to be up and running where we see that on the eight megawatt figure that you guys disclosed.

Yeah, thanks. I'll answer the first part. And Scott, if you want to round anything out after I'm done, please feel free. But on that, so we look back at the 18 megawatts that we guided to on July 14th, we had put out a plan to hit an additional 10 megawatts on top of the eight megawatts that we already had. So right, so Thank you for joining us. That additional capacity, we just announced the 3.75 that we've added to that capacity. So we're continuing to push really hard on that and still see a fair amount of smaller scale capacity out in the market that's available RFS this year. And to give you a little more specifics there, our preference is to find, call it one, two megawatts that's available RFS this year. but can expand beyond that. So if it's a 3.75 that has the capability to expand to 10, that's a much better fit for our longer term needs. However, we're also seeing smaller pockets of say stranded one to two megawatts that are still out there that haven't been prioritized. So we still see some good options out to hit the 18 that we were looking back at July. And I think you had a second part of the question.

Maybe just on the CapEx front, of all the contracts you've signed so far, I guess how much maybe GPU purchases or CapEx or GPU-related CapEx is left for you guys? And then I'll hop back in here. Thank you.

Okay. And I think you asked about the revenue per megawatt as well. So, Scott, I'll take the revenue per. You can jump in on the CapEx piece. But we are seeing, I think that the... The latest economics are coming into the 18 to 20 million per megawatt on the revenue side of the equation. And then Scott, do you want to comment on the CapEx side of that?

Sure. Yeah, all of the CapEx that is scheduled to go into the 8 megawatts has been ordered and financed. And as Mike alluded to, Mike mentioned, we expect all 8 megawatts to be active and revenue producing by the end of the year.

Thank you, guys. Thank you.

Your next question comes from the line of Bill Papagnastasio with Chardin. Your line is open. Please go ahead.

Yeah, good evening, gentlemen. Some of the progress and the inaugural earnings call of the public company. First one for me, as was mentioned in the print that Cumulus has signed a total of 21 new direct contracts in the quarter. Maybe we can just spend some time speaking about what customer profiles you're targeting as the company scales. Are you looking for particular enterprise customers and what goes into that decision?

Yeah, thanks, Bill. And we're excited to get this first one in front of us and behind us. And so thanks for that. Yeah, so the new contracts, we've announced that is made up of a total of eight customers. We have announced a couple of those customers where the customers have been willing to. But generally speaking, what we're seeing as far as those types of customers, it's a combination of things. Some of it is some of our marketplace partners, which we've been working with for the past couple of years. Some of it is with the, and the way I explain this is They're the AI inference GPU model as a service, generated by AI as a service type companies that are seeing really strong customer growth. They've usually gone out and raised significant capital from brand name are all venture capitalists. So for argument's sake, I'll say they've gone on race to say 100 plus million from A16Z or Sequoia or the likes of those categories. And they're in a rapid growth phase with a lot of customer demand and they need the compute to satisfy it. I would say that's where a lot of the near-term growth is coming from. We call those the AI natives. And then in our pipeline, we do have a lot of interest and continue to look at the enterprise segments. However, the enterprise segments tend to make decisions and move a little Thank you for joining us. We are bringing more supply and that word's getting out to the market. And as we are increasing our organizational maturity level, the quality of our customers, we also see increasing. So that's another positive note I'll highlight.

Great. Appreciate that color. And then hoping you'd be able to provide an update on the GPU financing environment. Have you seen any changes in The way I've explained this pretty consistently over the last couple of months is it seems like the capital markets side of the equation is moving

Thank you for joining us. A portion of that debt has come from USDI and the Permian Labs Agreement. Some of that has come from traditional lease agreements with companies like Tech Finance. So that's been working really well, but we're also seeing new announcements from our peers and from our partners about other creative financing opportunities and options. And really, from our standpoint, it's about aligning the right capital to the right asset, whether it's a data center, power agreement, or GPU or other, and then the right customer. and then understanding what the right cost of capital is. And yeah, I'd say there are more options now than ever that we're seeing out there and that's encouraging.

Appreciate that. And then last one for me, apologies if I missed this, but can we just touch on the prior guidance figure of 300 million exiting in the year? Did you mention that that was still in play, and can you speak to the demand funnel from July 14th when that guidance figure was published? Are you seeing the demand funnel kind of grow at the same faster or slower clip than was previously expected?

Yeah, so we did put the guidance on July 14th with a $300 million ARR target. We are not reaffirming any of the guidance today, but what I can tell you is we're We're not currently demand constrained. We're not able to satisfy all the customer requests that our sales team is in marketing are stirring up and bringing in and our partners are bringing in. So really the largest constraint to our growth right now is the land powered shell in the near term. And so that's why we've got a big focus on our land powered shell team on going out and finding that additional power. specifically in that July guidance was 18 additional megawatts or 18 total megawatts into the year on top of the eight we already had. So a lot of focuses on that both at the end of the year but also into 27 and even 28. I feel really good about the land-powered shell team and the team in itself. I think that we've done a great job at Having capital markets and the land-powered shell teams focus on those efforts while the operations team is heads down focusing on execution and delivering the contracts and bringing in the revenue on the contracts we've announced. So that's working really, really well. I just think that's a good point to highlight.

Great. Thank you for the callers.

Your next question comes from the line of Alan Klee with Maxim Group, LLC. Your line is open. Please go ahead.

Yes, hi. Could you talk a little about the competitive environment of companies you were going up against for GPU compute?

Yeah, thanks, Alan. Absolutely. I talk a lot about this in a lot of our historical materials and also I think in our recording that the differentiation is really about the what we are not. There are a lot of neoclouds. Thank you so much for joining us. However, there's just a lot of near-term demand of customers who are calling and saying, hey, we would love Compute yesterday if you could get it to us. That is a theme. We hear it from our sales team, we hear it from our customers, and we hear it from the market. It's not just us. If you listen to our peers' calls and you listen to hyperscaler calls, I mean, everybody is saying we are... restricted, growth constrained by supply. I think it's the exact words that some of them are using right now. So our differentiated approach right now is really, it's a combination of things. And I refer to it as a right to play, right? The right to play on the space is, do you have access to the capital? Do you have the LAN PowerShell to get the chips into? Do you have the partnerships in the supply chain to procure those chips? And I highlight the NVIDIA cloud partner relationship. That's a key piece of it, but also our OEM relationships with the likes of the Super Micros, Lenovo's, Dell's, et cetera, right? And then it's, can you bring those online and operate those effectively and do you have the team to do it? And this is one of my favorite things to say as a former public company CTO is it feels great to have a CTO I can really trust and lean on. So you put those pieces together and I think we check the boxes, right? So then it's, well, what is the additional differentiation that we're seeing and what's making us win these contracts? And it really is to me speed. Speed and fair costs. We can get it to our customers quickly because we're not out looking for 150 megawatts at a time and our customers are open to a few megawatts of compute at a time, if not less. And we can move that a lot faster. So I think to summarize it, it's that sub 50 megawatt thesis where we believe that we can deploy and find pockets of power shell, roll the GPUs, get them online a lot faster if we are focused on smaller to more moderate pockets of compute rather than large, large deployments.

Thank you. One other question. Could you just remind me of when you're trying to finance GPUs for the debt component, how much does it typically amortize? And is there any changes in rates with the current environment that you're seeing now versus maybe six months ago?

Scott, I might defer that to you. I think you can probably speak the clearest and cleanest right now.

Yeah, so most of our financing arrangements are three years. We do have a few that are longer, but most of those are financed over three years, which lines up typically with some of our recent longer-term compute contracts. In terms of our partners, We have some great partners that we've already done some financing relationships with and we expect to continue to grow with them. So as we get bigger, as we deploy more, we expect obviously the cost of capital to go down. But that's really what we're focused on is finding the lowest cost of capital in each case.

Thank you.

Thank you for attending. You may now disconnect.