Is the future of business AI specialized services?
FluidStack is a name that most organizations may not be familiar with, unless they’re operating in the scientific research industry. Nevertheless, the experiences and business model of the company could offer lessons for multiple verticals.
The company was founded by César Maklary and Gary Wu and is one of the many science and technology businesses spun out from Oxford University. In the words of Mike McDonald, VP of product: “They started very much on the genesis of college kids trying to get access to GPUs in the late 2010s.”
“It was very hard to get even the (Nvidia) T4s or 1000 series at that time, so their whole idea was how do we build a marketplace – again, back in 2017 – how do we get all of that capacity online?,” says McDonald.
“So while it started off as a GPU marketplace, it’s been through a couple of iterations (such as) how do we apply that towards gaming? How do we apply that to scientific computing?”
It’s this area where the company laid down the roots of what it has become today, starting with simulation and then moving towards getting its capacity into the hands of AI research labs as a GPU reseller.
“We would go and acquire capacity on a third party – other tier two CSPs or former crypto miners that bought GPU capacity. What [Maklary and Wu] found out is researchers don’t know how to use bare metal GPU capacity and crypto companies don’t know how to build an AI cloud,” says McDonald. “So we filled that gap.”
Subsequently, customers started asking for a more private infrastructure setup using colocation providers, with one such request coming from French generative AI company Mistral.
“Mistral grew and grew and grew and we were deploying some of these large scale 1,000, 2,000 clusters. Then they said ‘hey, take this in house’,” says McDonald. It was at this point the company partnered with Dell and Nvidia to put 160 servers in an Icelandic data center, becoming Mistral’s private cloud provider.
While Mistral is an important client and partner, the target customer base for FluidStack is still research labs. From the company’s point of view, this made more sense than trying to compete with other more generalist cloud service providers, both from a competition point of view and wanting to offer clients a more targeted approach to their sector’s needs.
“Building a multi-tenancy infrastructure for people paying $1 per GPU per hour on a credit card – there’s plenty of other people in that space. What we found is there were a lot of [companies] who weren’t going after the super high end and sort of catering to their every whim,” says McDonald. By contrast, FluidStack has taken exactly that route, which, he says, has made them a better business both technically and from a customer experience point of view.
“It’s forced us to be better at engineering and a more customer-focused organization,” he says. “We’re trying to focus on what the bleeding edge is doing.”
“The other thing we have is this underlying software platform that’s the orchestration layer called Atlas,” he continues. “With Atlas, when one customer hits some weird Nvidia XID error, for instance, we’re able to then go and backfill that particular thing into every other customer.
“A lot of customers are doing similar things and will run into the same problems. We can fix for one of them and the others never even know it was a problem. So we end up having a more effective software platform by picking more sophisticated customers that push us to build better software.”
In other words, specialization pays off for both the customer and the infrastructure provider.
Source link
