The curious case of Nebius, the publicly traded AI infrastructure ‘startup’

On October 21, a new ticker opened to Nasdaq traders: NBIS, a truncation of Nebius, a fledgling player in the AI cloud infrastructure space.

Casual observers could be forgiven for wondering where this company had come from, as there had been little in the way of the usual fanfare that surrounds most startups’ journey to IPO — no roadshows; no horn tootin’; no confetti-laden ceremonies; nothing, not a peep. That’s because Nebius is an unusual beast: a public company, but a startup in just about every sense of the word.

Nebius has actually been public for 13 years, floating in May 2011 as Yandex N.V. — the Dutch holding company of Russian internet giant Yandex (often dubbed the “Google of Russia”). At the tail end of 2021, Yandex N.V. hit a peak valuation of $31 billion, but in the wake of Russia’s invasion of Ukraine in early 2022, everything changed. Nasdaq halted trading in Yandex N.V. shares that February due to sanctions imposed on Russian-affiliated companies, and a year later Nasdaq said it would delist Yandex altogether. But Yandex successfully appealed on the basis that it was restructuring — a process that would take an additional 16 months to fully complete.

Part of this included offloading all its Russian assets, which was where most of the real business value lay. What remained under Yandex N.V.’s ownership was a random assortment of infrastructure and business units that just happened to be located outside of Russia. This divestment concluded in July, with Yandex N.V. changing its name to Nebius AI, an AI cloud platform replete with its own Finnish data center.

The new business was to be spearheaded by Arkady Volozh (pictured above), the Russian Yandex co-founder and former CEO who was removed from a European sanctions list in March after he publicly condemned Russia’s assault on Ukraine.

The core Nebius business sells GPUs (graphical processing units) “as-a-service” to companies needing “compute” — that is, processing power and resources to carry out computational tasks such as running algorithms and executing machine learning models. Last month, the company debuted a holistic cloud computing platform designed for the “full machine learning lifecycle,” spanning data processing, training, fine-tuning, and inference.

With the restructuring complete, and Volozh free to run the show from the company’s new HQ in the Netherlands, Nasdaq green-lighted Nebius to recommence trading last month. The situation was pretty much unprecedented, though: a public company whose trading was put on pause, only to resume nearly three years later under a new name and entirely different business proposition?

In many ways, it would’ve made sense to have delisted and grown with private capital, the good old-fashioned startup way. But as Volozh explained to TechCrunch earlier this year, building infrastructure is capital intensive, and the easiest and cheapest way to access capital in what is currently one of the hottest spaces in tech is via the public markets. But there was never any certainty on how the public markets would respond to this strange new entity. Nobody really knew what to expect.

A month in, and Nebius has enjoyed a somewhat tepid re-entry to public life; it’s significantly down on its $18 billion market cap before trading halted in February 2022, which was to be expected, and it has since yo-yoed between $3.5 billion and $4.75 billion, with some signs that it is starting to settle.

“We couldn’t predict what would happen, it could be $5 per share, or it could be $50 per share — this has never happened before, nobody really knows how to treat it,” Volozh told TechCrunch in an interview in London this month. “It’s still volatile, but it’s stabilizing, and the good thing is that it has stabilized above the cost of the assets, which means that the market believes we will be able to build a business here. How big a business, we’ll see.”

Nebius competes with all the usual hyperscaler cloud behemoths, though arguably its more direct rivals are other alternative cloud startups such as CoreWeave, which has raised a ton of cash this year. With CoreWeave in the midst of expanding from the U.S. into Europe, Nebius is moving in the other direction, announcing plans this week to extend its presence to the U.S. with a new GPU cluster in Kansas City (on the Missouri side) scheduled to go live in early 2025. The company has also opened “customer hubs” in San Francisco and Dallas, with plans for a third in New York by the end of the year.

But while the cloud infrastructure business is its bread and butter (accounting for two-thirds of its revenue, as per its first earnings report last month), there’s a triumvirate of additional businesses under the Nebius Group umbrella. This includes an autonomous vehicle company called Avride, based in Texas; a Swiss-based generative AI and LLM company called Toloka; and edtech platform TripleTen, located in Wyoming.

Drive time

Avride descends from the international division of Yandex’s self-driving unit, which spun out of a joint venture with Uber in 2020. While Alphabet’s Waymo is now leading the way in the burgeoning robotaxi realm, recently securing a $45 billion valuation, Yandex was an early trailblazer in Russia, with Volozh noting that the company had been on the cusp of beating Waymo to launch the first fully autonomous cars on public roads, before the war put the kibosh on plans.

“They [Yandex] were set to launch the first taxis on public roads with nobody at the wheel, in a real city (Moscow), several months before Waymo launched in San Francisco,” Volozh said. “Journalists were invited to a big event in March, ’22, but that launch never happened. People had to pack all their things and go in a matter of weeks.”

The team that had been working on Yandex’s autonomous vehicle project transitioned over to Avride, a new brand it launched last year, eventually moving to Austin via Tel Aviv.

“This is the same 250 people,” Volozh added.

Last month, Avride announced a significant multiyear partnership with Uber, which saw Avride’s sidewalk food delivery robots land on Uber Eats starting in Austin, though the partnership will also bring Avride’s self-driving cars to the Uber platform later (Uber has signed other similar deals, including with Google’s sibling company Waymo).

While Yandex had sufficiently deep pockets to fund autonomous vehicle projects, Nebius doesn’t — it has a couple billion dollars in the bank from its Russian divestment, and it’s laser-focused on building its cloud infrastructure business. And this is why Volozh says that Avride will need to find additional partners in the longer term.

“They have enough budget for this year and next year,” Volozh said. “We’re financing them, but they need to use this time to find new partners, because it’s very capital intensive to build fleets. It needs real investment.”

Obvious partners might include car manufacturers, but it could be any entity that’s ready to invest billions, with Volozh adding that it would be willing to give up control in Avride if needed.

Toloka, meanwhile, is a platform that specializes in data labeling and quality control for large language models (LLMs) and related AI systems — it’s much like Scale AI, which was most recently valued at more than $13 billion. Toloka has clear synergies with Nebius’s core infrastructure business, but the customers aren’t the same. Nebius works largely with generative AI startups seeking compute, whereas Toloka works with bigger companies such as Amazon and Hugging Face that want to improve their LLMs.

Both Toloka and Avride could eventually follow a similar path to that of ClickHouse, creators of the eponymous open source database management system that spun out of Yandex in 2021. While the commercial ClickHouse entity secured big-name backers such as Index Ventures, Benchmark Capital, and Coatue, Nebius has retained a minority stake.

“ClickHouse became very popular, and we were approached by investment funds to create a business around the open source project. Now they have revenues, and they’re growing,” Volozh said.

TripleTen, on the other hand, is something of an outlier in the Nebius group of businesses, in that it’s pretty much a direct-to-consumer product that offers online coding bootcamps for those wishing to transition into the technology sector. One idea Nebius is dabbling with is to position itself as a provider of a “full stack of services” to AI companies, from data centers and GPU infrastructure, to education. And this highlights the situation that Nebius has found itself in: It’s drawing lines between the different entities it has been left with, and trying to make it all make sense.

For now, TripleTen is breaking even, and Volozh acknowledges that it’s not going to be the big revenue driver that its infrastructure business is — but it has the potential to provide meaningful income and will remain part of the Nebius Group.

“Nebius is a billion-dollar scale business,” Volozh said. “TripleTen — it’s a nice model, but it’s maybe a tens or hundreds-of-millions of dollars business. It’s not a billion-dollar business.”

Parallel compute

As for the core Nebius AI cloud business, the company already has its fully owned data center facility in Finland, with plans to triple its capacity to 75 megawatts. In tandem, the company is building out additional sites at co-location facilities, a move designed to not only increase its capacity, but also to reduce latency by bringing the processing closer to its customers. In addition to the Kansas location announced this week, Nebius had already unveiled a new GPU cluster in Paris that goes online this month.

Further down the line, Nebius plans to build more of its own data centers, both in Europe and the U.S., but given the time it takes, it’s quicker to plug the gap with co-location facilities, which is why it’s forging ahead with a hybrid approach.

“It’s more efficient if we build it ourselves, but to build means a year and a half or two years — it’s a long process, and we can’t wait,” Volozh said. “That’s why we have these co-locations in Paris and Kansas City.”