It is a little over a year that Oxalya was acquired by OVH. What brought you to sell to a web hosting company?

Oxalya and OVH are two complementary strategic businesses. Oxalya’s problem was investment. OVH’s problem was to find a path to grow in sectors that were not necessarily theirs. Today, OVH is the most important Intel customer in Europe. When OVH’s potential HPC growth came into the discussion. Intel told them that if they wanted to do HPC, they should tell Oxalya. In the HPC community, Oxalya has always had good press. From our side at Oxalya, we were knowing the guys from Intel quite well, since we were selling a lot of Intel configurations to our HPC customers. Intel also knows us well from our innovative components. At the time, in remote visualization for numerical simulation, Oxalya was the only one ahead. Take for instance the image wall deployed by EDF R&D. The wall part was supplied by Barco but the software and infrastructure part was provided and implemented by Oxalya. It is our software that allows the operation of 16 synchronized projectors on a slab of glass that weighs two tons. We also managed and coordinated large collaborative research projects, of up to 18 partners. Anyway, to make along story short, Intel knew us well. For Intel, HPC is a growing segment and a prestigious area to demonstrate technology. So they naturally favored the merger.

So you didn’t think you could continue to grow independently?

Our business logic was software, not hardware, development. We wanted to develop SaaS-enabling tools, to create the software infrastructure capable of hosting other software in SaaS mode. We know how to “SaaSify” an application by installing it locally and making it easily available. Whatever the product from Ansys, ESI, Dassault Systèmes or other vendors, we can do it. We truly master the capacity to automate all the availability processes.

But Oxalya’s vocation wasn’t to invest in a physical infrastructure. We did it once because we needed to, knowing perfectly well that hosting is a mass market. In my opinion, HPC is a bit limited in terms of critical mass to absorb the necessary investments. So, in partnering with a player like OVH, Oxalya got the critical mass. And it was very interesting for Oxalya to lean on OVH considering the commercial opportunities and scalability issues that could follow. OVH manufactures its own servers, on its own assembly lines. If the need arises, we can have, in a few days, 300 additional servers designed to our specifications and plugged into the datacenter. No need to wait 6 to 8 weeks before a manufacturer supplies us.

OVH is experiencing a disruption in dedicated servers following a successful commercial offer. Is this not a limitation of the model and, more importantly, what was the impact on the HPC activity?

Again, yes and no. There is currently a sold out on part of our offers but we are not really out of stock on servers. In fact, we want to change our business model, and that is why we stopped the orders. We need to take some time out to ask ourselves the right questions, to put in place the proper mechanics. Then we’ll reopen the valves. Now if a HPC customer calls us for a large amount of servers, of course, we can deliver. Our supply chains averages between 500 and 1000 servers a day. With a capacity this large, things can go very fast. On an order of several thousand servers, we won’t be ready in 24 hours but then again, in HPC, you never have this type of 24 hours delivery request. However, in a few days, it is quite possible. And in software terms, we have the infrastructure to scale at will.

What capacity do you still have in your datacenters?

We now have 170,000 servers in production, but we have the infrastructure to produce up to a million. With this almost unlimited capacity, we can move very quickly. For Oxalya, this is a very important differentiating asset. A pure player of HPC will now have a hard time catching up, because building a datacenter is complicated but also because energy is the primary cost of a datacenter. OVH has a PUE of 1.09 in its two centers at Gravelines in France and Beauharnois in Canada. And yet, this figure is calculated in the worst case way, i.e. directly out of high-voltage lines. We prefer to use this figure since PUE is not a standard yet. By changing some of the measurements we could still go lower.

Moreover, OVH is now present in 16 countries. In this respect, our new Canadian center is an ideal gateway to the North American market. Plus, we also have our own network – based on a huge 5 Tbps backbone! If one of our HPC customers experiences network congestion, it will be from their side, not ours. And once the data is with us, remote visualization takes over. These networking capabilities increase the value of everything that has been developed so far on the software side.

What is the advantage of having your own interconnection network between your datacenters and access points ?

In terms of SLAs, it is very important to have a network. We can measure our quality of service between our gateway in Hong Kong and our center in Canada. And what we can measure, we can commit to. When you use a standard cloud, you go through an operator network to access it. Then you go in multi-operators mode up to the destination. It’s a best effort route for everyone, which makes offering an end-to-end SLA impossible. I personally worked on a European project mobilizing 15 operators whose goal was to implement end-to-end SLAs. I’m talking about the ETICS project, in which Oxalya brought the HPC use case. Conclusion: it can’t be done! For operators, the availability of their networks and their various routes is core business. They are certainly not going to offer competitors the ability to manage their own routing. It is therefore impossible to commit to an end-to-end SLA today except if you have your own network, which OVH does. This is a huge differentiator.