Profile. Qumulo brought in Doug Gourlay in July, appointing him as president and chief executive. Gourlay also joined Qumulo’s board of directors. The business has an on-premises scale-out file system and cloud-native product that can be used in the general file-focused unstructured primary data market.
Qumulo has more than 1,000 customers in 56 countries, with several adjacent and overlapping markets. Segments it serves include cloud file services; public cloud unstructured data; high-performance computing (HPC); file lifecycle management and orchestration; and GenAI training and inferencing, for example. Gourlay’s stated mission is to accelerate the business’s growth but says there is a constant tension as Qumulo and its leadership considers its markets. Which are worth adopting and which are niche – for now, the mid-term, or long term?
Gourlay tells Blocks & Files he has to consider things including the company’s product set, the core culture of the company, and market developments before he can steer Qumulo in the right direction.
He says: “There was an obvious gap that the company did have, and that gap was it didn’t have a good story, it didn’t have a good talk track of where it was and where it was going. And a lot of our customer base wanted to know that the boat had a rudder and which way it was pointing.”
There are advantages in being a niche file-type product: “It’s not a niche product, it’s a enterprise-wide primary storage offering. And the downside of that is the niche offerings command significant price premiums because they do something in that niche that somebody thinks is valuable. Their go-to-market organization knows exactly what to call on, and they talk to this person, that person goes, wow, that’s amazing. That solves so much for me that I’m willing to pay you a metric ton of cash for.”
The more general-purpose products have lower prices. Should Qumulo move to these higher-priced niches? Gourlay says: “If I turned to just doing a specialty offering, we would lose the top-line revenue that the company depends on. If I just rotate it tomorrow to only do cloud, even though the unit economics are better, the margins are better, the deal velocity is better, I kill the company.”
Cloud and hybrid cloud
Blocks & Files sees Qumulo as a hybrid – on-premises and public cloud – company. Gourlay responds: ”Our customers are hybrid. And our cloud story is very simple. I can state in less than 30 seconds and probably one sentence: our job is to align our technologies with our customers’ priorities. Our job in cloud, therefore, is to give our customers the ability to make a business or economic decision about where their data and workloads reside and to eliminate all the technology barriers to that decision.”
He doesn’t buy into simple cloud repatriation stats, saying the situation is much more nuanced than we might think: “Michael Dell … posted to LinkedIn … four months or four weeks ago, which was 85 percent of the customers we polled are repatriating from the cloud. They’re going back on-prem.”
“OK, I have a book called How to Lie with Statistics. All right, if I tell you that data, it might be 100 percent accurate, but I’m not telling you how many workloads are moving. Somebody moving one workload? Check! You’re repatriating from the cloud.”
But it’s only a single workload, out of tens or hundreds. “There’s workloads that make sense in the cloud, and there’s workloads that don’t. Now, if I’m a cloud-only company, I’m going to tell you a completely different story. If the whole world is going cloud then, if you’re not going cloud, you’re a Luddite. If the answer isn’t cloud, the question is wrong. It’s cloud-first. It’s cloud-only.”
For now, Gourlay says: “Cloud, this quarter, is probably a fifth of our revenue contribution. It’s about 20 to 28 percent.”
It’s likely growing and the need for a hybrid data fabric with good on-prem-to-cloud data movement (and no doubt the other way) is too. He asks rhetorically: ”Can we invest in advanced technologies that allow rapid data replication between the cloud and the on-prem that accelerated and overcome some of the bandwidth delay product issues? So we’re testing those right now, with tremendous early results and further layering on capability that differentiates and maps these use cases.”
He sees similar one-dimensional – and wrong – thinking elsewhere: “If I’m Jensen (Huang), it’s all about the AI center, the datacenter is the AI center, the future is all GPUs and all our type of computing.”
This is half-blind thinking, he says: “These are wonderful stories from myopic points of view. Our customer base has a mosaic point of view, not a myopic one. And they want a system that works for their AI workloads, for their own workloads and their cloud workloads. And they want one that is consistent in operation and capability and capacity, regardless of which modality they’re executing against. That’s our job. That’s our job more than anything.”
He wants Qumulo and its people to have “organizational alignment around a common vision and strategy. A common technology evolution that addresses key priorities within our customers, whether those are AI, whether those are cloud adoption, whether those are a shift to different virtual machine types inside their datacenter, whether these are this ever expanding data set that our customers are getting weighed under.”
“Historically, they were deploying different file systems and different storage systems for each application. One of the great things about what we do is we can consolidate those. One of the great things about the way that we move the data inside of the system is we can start collapsing tiers very cost effectively.”
Having a global namespace (GNS) as an integral part of the product is a vital facility for Gourlay. He is not a believer in adding an external GNS, which is one way he views Hammerspace: “The unfortunate problem of an overlay GNS on third-party systems is no system has the verbs in place to give them an assurance that the data has been durably written to multiple targets before they acknowledge the data. And in that scenario, you end up with a tremendous risk of data loss, which our customers have experienced, multi petabyte data losses.”
“It’s funny that I looked throughout all of our marketing material, and that never came out to me. I see customers drawing up a four tier storage system. I’m like, but guys, I could put a lot of QLCs in here and a caching layer, and I could do hybrid over there. But if I just increase the cache sizes and the SSDs, [then] haven’t I really collapsed two or three of those tiers together, at a similar price point, with a larger storage cluster?”
Gourlay discusses a customer example: “I have a customer who does rocket launches. We store the telemetry for it. They may have had data loss because of organizations that didn’t have the ability to guarantee writes rights – so strict consistency matters when the data matters. Our customer base has data that matters.”
AI training and inference
What about AI where Qumulo, unlike NetApp, Pure Storage, and other competitors, does not support Nvidia’s GPUDirect protocol for feeding data to GPUs for AI work? He makes a distinction between AI model training, which does need GPUDirect, and AI model inference, which does not.
The Qumulo CEO thinks that AI training is a highly niche market. The 15th largest GPU cluster in the world has, he says, just 256 GPUs. “Numbers one through 14 are larger and number 16 onward are smaller. Why do I want to chase such a small market?”
He’s emphatic about this point: “Why do you want me to sit there and compete with four other companies, all chasing 14 companies that are spending enough to be worth calling them? Do you realize that the largest financial institution in the US has eight DGXs and doesn’t know what to do with them. Number one, the largest bank in America, has eight DGXs. Why should I bother with a tiny market?”
There are two successful storage suppliers to his knowledge, but hyperscaler customers are ruthless: “There’s two storage suppliers [that] are being pretty successful, right? One of them is getting kicked out of the largest cluster because they’re building their own. Now, that’s the other problem. These are hyperscalers. Yes, the largest AI cluster in the world is moving from an open storage environment, consuming a commercial product, to them building their own.”
This is a risky sales approach: “If I’m a company that has a 50 percent revenue concentration in one customer, that’s huge risk. I have 1,000 customers. I have over five exabytes to date under storage. I don’t have a single customer worth more than 2 percent revenue to me. I don’t have risk. They do. They have a customer concentration risk with customers who actively want them out of their system.”
”Hyperscalers either want to ram your margin down or get you out and replace you with something they can build themselves. I want customers who love me, who want to keep us in because they love what we do, and who aren’t capable of investing hundreds of engineers of effort in getting us out of their networks and systems.”
This is the situation now. It could change: “My statement wasn’t never. My statement was, I’m not going to chase it now.”
“I need to do this for a different reason. I need to do this because I have a substantial number of the largest ADAS (Automated Driver Assistance Systems) clusters in the world. I’m in seven of the eight largest autonomous driving clusters in the world, two largest research grant recipients in North America. We’re in the largest pediatric critical care facilities … we have customer demand now to start building it for things that are being delivered in the next two to four years, [the] cycle of our customers adopting next generation technologies. So it’s not a never, it’s at a right time.”
“I want to be right-timed. I want to hit a market inflection. I want to hit it when it scales out. To go for the broad customer adoption, for the base we have. I don’t want to build customer concentration risk in a market that wants to evict me. I want it with ones that embrace me. So I’m not saying never, definitely not saying never.”
If not never, when? “I think if you and I are sitting here next year having this conversation, I will happily have product to market that addresses some, if not all of what you and I are discussing.”
A last question from us: “NetApp is building a third ONTAP infrastructure to cope with supplying systems for AI, a disaggregated infrastructure. Dell is aiming to stick a parallel interface on top of PowerScale. Is Qumulo thinking about being able to respond already to these moves; developing something like a parallel architecture for itself?”
Gourlay exudes positivity about this: “I think if you take a look at the architecture that we’re using in our cloud-native offering, one that runs as a series of EC2 instances, backed with EBS and local NVMe, but then using parallel access to object storage in the back end, called S3 in the cloud. I think you see the exact architecture you’re describing in production today.”
****
A CEO like Gourlay can encapsulate key aspects of a company’s offerings and approach, then express them compellingly, not shying away from debate with opposing views, and winning doubters over with strongly argued and logical views. It’s a formidable talent and, when coupled with a clear view of a market and its realities, should enable a business to do very well indeed.
Qumulo is being re-energized and carving out its own messaging under Gourlay. Its competitors are going to face a tougher fight when meeting it in customer bids, and he’ll relish that.
