Nscale makes big pitch as UK AI hyperscaler, and Nokia's on boardNscale makes big pitch as UK AI hyperscaler, and Nokia's on board

Nscale could not have picked a more stunning location for its inaugural European data center. Photography shows it perched next to the dark waters of a Norwegian fjord, looking boxy and incongruous among the pines. On the opposite bank, a wall of craggy mountains rises abruptly out of the depths, its peaks dusted lightly with snow. But Glomfjord was not chosen for its scenery. Tucked just inside the Arctic Circle, with hydroelectric power readily available, it is a prime spot for what basically constitutes an artificial intelligence (AI) factory.

Brought out of stealth mode as recently as May, Nscale sees itself as a kind of UK-headquartered answer to the US hyperscalers. But rather than trying to compete against the likes of AWS and Microsoft for routine cloud business, it believes it has spotted a lucrative and ill-served niche amid excitement about generative AI.

Demand for graphical processing units (GPUs), the chips deemed essential to AI, has driven a ninefold increase in the share price of Nvidia, their main provider, since early 2023. But the hyperscalers accounting for about half those GPU sales lack the free capacity to serve the AI needs of many organizations, according to executives at Nscale. They position their company among an emerging crop of "AI neo clouds" that aims to fill the void with GPU-as-a-service offers.

"You have the Tier 1 hyperscalers, but if you go to them and ask for a 10,000-GPU cluster, you're not going to get it," said David Power, Nscale's chief technology officer, at a briefing in London. "They add GPU capability to different regions, and it's typically gobbled up by small players, but they don't have these big, contiguous computing clusters."

The neo clouds

Perhaps the best example of this AI neo cloud model so far is CoreWeave, a US company catering largely to the healthcare sector. After recently closing a $650 million secondary share sale, it has reportedly been valued at around $23 billion, up from just $7 billion a year ago. Nscale's interest in replicating that approach is not hard to understand.

Investors have already bought in. On December 9, Nscale secured a Series A funding injection of $155 million from financiers including Sandton Capital Partners, Kestrel 0x1, Blue Sky Capital Managers and Florence Capital. It is constructing what Power describes as "a 1.3-gigawatt pipeline of assets," which includes another Norwegian data center in the more temperate zone of Stavanger, the country's third-biggest city. As noted by Telecoms.com, Light Reeding's sister publication, that 1.3-gigawatt commitment would today make Nscale the sixth-biggest hyperscaler by capacity, based on a ranking from Data Centre Magazine.

Nscale, however, has a long way to go before it is even close to that figure. Its Glomfjord data center has an operational capacity of 30 megawatts (MW), although Nscale says it is "expandable" to 60MW. The company's immediate goal is to add 120MW in 2025 to the 300MW it currently boasts across Europe and North America. New locations it is considering include the UK.

Investment needs are significant because of Nscale's determination to provide the full monty. "It's the land, the power, the data center, the hardware, the networking, the software stack, the API endpoints," said Power. "All will be developed in-house by Nscale and owned under our brand, as opposed to taking pieces of the stack and then sitting in an third-party data center and having a margin."

But Nscale says it has been able to mitigate financial risks by securing anchor tenants for the capacity it will bring online. It also downplays the difficulty of obtaining GPUs from Nvidia amid constant talk of supply shortages as hyperscalers squeeze out rival buyers. Nscale is purchasing in sufficient quantities to be a valued customer from Nvidia's perspective, insists Power. Sam Sturgess, Nscale's vice president of strategy and corporate development, says a move by AWS and Microsoft to develop their own AI chips gives Nvidia a big incentive to diversify its customer base.

Power reckons there is now a serious GPU alternative in AMD, as well. "AMD's approach is quite interesting," he said. "They have a very open software ecosystem. They integrate very well with common frameworks." The facility in Glomfjord seems largely if not entirely based on AMD's rather than Nvidia's GPUs.

InfiniBand out, Ethernet in

Nscale has also looked to someone other than Nvidia for its networking needs. An obvious choice for connecting clusters of Nvidia GPUs in a data center would be InfiniBand, the technology Nvidia has effectively owned since its $7 billion takeover of Mellanox in 2020. But Power wanted to avoid over-reliance on Nvidia. "That's the only place you can get InfiniBand from and so it's not really an open standard anymore," he said.

Instead, at the new Stavanger facility, the company has opted for an Ethernet-based alternative from Nokia, which stands to do very well out of AI data center business if others share Power's reservations about InfiniBand. Thanks to the efforts of a group called the Ultra Ethernet Consortium, and its RDMA (remote direct memory access) over converged Ethernet (RoCE, pronounced "rocky") technology, the industry now has something with "the same performance characteristics of InfiniBand but greater scalability and interoperability," says Power.

The Nokia "fabric," as the companies describe it, includes Internet Protocol (IP) routers whose essential job is to ensure those GPUs can harmoniously interconnect. And the volume of traffic is significant. Power says each GPU is accompanied by a 400-gigabit network interface card. "You've eight of these per server and so you've got 3.2 terabits of data coming in and out of each of these systems."

While Nokia is likely to face competition from the likes of Cisco and Arista in this market, it evidently views the sector as one of growth for the coming years. "There is probably quite a lot of opportunity for us and other Ethernet providers because it's open," said Paul Alexander, Nokia's country general manager for the UK and Ireland. "I think previously there has been a level of lock-in that you won't get as it becomes an Ethernet-standardized approach."

News of the contract with Nscale comes weeks after Nokia landed a similar deal with CoreWeave. In that case, it provides both IP routers and optical transport technologies for interconnecting data centers in the US and Europe. Anticipated demand from AI data center customers partly explains Nokia's $2.3 billion bid for Infinera, an optical equipment specialist, in June. "AI is driving significant investments in data centers at the moment and one of the key attractions of this acquisition is that it significantly increases our exposure to data centers," said Nokia CEO Pekka Lundmark on a call with reporters at the time the deal was announced.

In a weak market, the IP networks unit, a part of the network infrastructure business group, stood out as Nokia's best-performing unit for the recent third quarter, when sales were up 6% year-over-year on a constant-currency basis, to €581 million (US$609 million). That contrasts with the 17% decline at mobile networks, still Nokia's largest business group, over the same period.

Perhaps the biggest risk for both Nokia and Nscale is that enthusiasm for generative AI starts to fade as the promised benefits fail to materialize. So far, there has been little sign of an AI-fueled economic boom except for Nvidia, which has defied expectations sales growth will eventually slow. But if CoreWeave's momentum is indicative, Nscale's investors have cause to feel confident. And the investments it plans can only be a good thing for Nokia.