Nebius Group N.V. (NBIS)

A Cheap Unknown AI Play

Nebius Group (NBIS) offers an opportunity to invest in AI infrastructure at a reasonable price. There are plenty of reasons for that. It's a new name (formerly Yandex, which got relisted without the Russian assets). There's no coverage or articles. Eventually, this will change.

---

Nebius Group N.V. (NBIS)

Disclosure: Long

Here are my notes on Nebius Group N.V. (NBIS).

Profile

• Currently re-listed on the Nasdaq. Former Yandex N.V. (YNDX)., aka "Russia's Google''. Yandex was halted on February 28, 2022. Resumed on Oct 18, 2024 under new name/business.  Rebranded as Nebius. 

• Completed divestment of Russian business. Sold Russian assets to a Russian consortium for $5.4b ($2.5b cash and the remainder paid in its own shares). Nebius has $2.4b cash and no debt. There's no more Russian connection. 

• Nebius/Yandex was allowed to maintain a suspended listing provided it decoupled Yandex/Russian assets.

• Nebius, the new company, pivoted to AI cloud infrastructure and aims to be a leader in GPU-as-a-service.

• NBIS is now a pure play in the fast-growing AI infrastructure space.

• CEO: Arkady Volozh, the co-founder of Yandex. EU removed him from the sanction list in March 2024 (Russia-Ukraine war). We have a founder who succeeded before. Yandex was worth $30b at one point.

• The newly listed stock is all over the place. NBIS currently trades at ~$20 a share. It has a $4b market cap. This includes the $2.4b in cash. It has an enterprise value of $1.6b.

• Headquartered in Amsterdam. Additional R&D hubs in Europe, North America and Israel.

• Nebius is a technology company building full-stack infrastructure to service explosive growth of the global AI industry, including large-scale GPU clusters, cloud platforms, and tools and services for AI developers. 

• Full stack: Provides the data center, hardware, cloud platform, AI expertise and service (they are across the value creation chain)

• Nebius provides infrastructure and services to AI builders

• Nebius is 4 businesses:

  •  The core of the group is Nebius AI, an AI-centric cloud platform built for intensive AI/ML workloads, running a network of data centers with thousands of GPUs and a set of services for LLM builders. Nebius provides AI and ML practitioners the compute, storage, managed services and tools they need to build, tune and run their models.

  • Toloka AI - data partner for all stages of AI development from training to evaluation: humans training machines

  • Avride - one of the most experienced self-driving teams: driverless cars and delivery robots

  • TripleTen - a leading edtech player training people for tech roles.

  • Owns ~28% of ClickHouse which is valued at $2b (stake = $560m)

• 1,300 employees. 850 AI/tech/cloud engineers

• The business in a nutshell: Providing compute for AI. Build data centers with providing AI compute infrastructure and value-added services to the AI industry

• Back story:

  • Formerly Yandex N.V. in Europe. It was the parent company of Yandex.

  • Hundreds of engineers, half of senior management, and employees, left Russia. Including Arkady Volozh. He strongly condemned the war.

  • Built on the foundation of Yandex N.V.

  • Managed separation from Russia. Divestment of Russian assets completed on July 15, 2024. 

  • No more Russian connections.

  • Even though it has the feel of a start-up, this is not a new start-up or company. It comes from a strong tech heritage. It has people + capital + tech.

  • $2.4b in cash (result from divestment. No debt.) 

  • State of the art infrastructure + AI engineers

  • Allows to scale quickly and boost compute capacity

• Demand for AI infrastructure will be massive and sustainable

• It's all about compute. Compute is what underpins the current revolution in GenAI. Increase demand for compute will fuel demand for this industry. There's been substantial growth and it's still early innings.

• Compute resource is the engine that powers applications, services, model training and inference.

• Compute power alone does not guarantee a better AI model, but one school of thought in the industry is that more of it generally equates to more capable models.

• Finnish data center has 14,000 GPUs (paid and deployed).  20,000 by year-end. Plans to triple to 60k GPUs. Nebius also has an AI cluster at Equinix data center in Paris (co-location).

• Aiming for 240k+ in the medium term. (Nvidia H100 equivalent). 

• Nbius is rapidly expanding in the US, where many clients are. New GPU Cluster in Kansas City will service customer workloads from early 2025 (capacity of 35k GPUs after expansion). By comparison: xAI data center just opened with 100,000 GPUs. It's believed to be the largest cluster of GPUs. It will be used to train the AI model behind Grok)

• Nebius provides its customers with access to NVIDIA GPU resources, enabling them to efficiently train and deploy complex AI models and handle compute-intensive workloads. (GPU as a Service (GPUaaS) or Artificial Intelligence as a Service (AIaaS)

• Because Nebius is relatively new in the space, they have a flexible offering (short-term 1yr contract and on demand). I'm guessing the goal is getting customers, servicing these customers as they scale to become larger customers with more sophisticated needs.

Numbers

• Q4-2024 ARR: $170-$190m vs Q4-2023 $21m

• 2024E revenue: $120m-$130m

• Estimated 2025 ARR: $500m-$1b. (corresponding $500-700m for 2025E revenue)

• But the opportunity is multi-billion dollar annual revenue potential 

• 40+ clients (10 in 2023)

• Growth: Existing clients growing consumptions + new clients + expanding capacity

• Expected to triple power capacity to 75MW by the end of 2025 or early 2026

• Nebius expects the revenue to increase at least 5x and become Adj EBITDA positive in 2025

• 2024 CAPEX: $1b

• 2025 CAPEX: $600m- $1.5b (most of it will be invested in NVIDIA GB200 GPUs, as well as expansion of our DC capacity)

• I assume the $2.4b cash will be used for CAPEX/operations. They will need it for growth.

• I didn't explore deeply the other businesses: Toloka (data for GenAI), Triple-Ten (ed-tech, AI retraining), AVride (autonomous driving for cars and delivery robots). I'm not sure how material they are at the moment (it doesn't seem to be in terms of CAPEX).

• I believe Nebius core AI business is what's driving the value. The rest is free upside if something substantial comes out of it (monetization to material profit). 

• This company is an under the radar AI play. It's relatively unknown. It's a new listing (well re-listing). There's no broker or analyst research at the moment. There's no articles. There's nothing in Barron's or WSJ. 

• It will take some time to ramp up ARR and name recognition.

Questions, Risks, & Concerns

• How superior is Nebius' offering? Why them?

• Need to secure access to land, power, and GPUs. None of this is easy.

• Is there enough power to run these data centers? At a decent cost?

• AI infrastructure is capital intensive. It will cost billions to build. 

• What's the depreciation and replacement cost on these GPUs and infrastructure? I guess the question is how much is maintenance CAPEX? and growth? Going forward.

• Companies throw more compute power at training LLMs, they aren’t seeing enough performance gains to justify the extra cost. Microsoft has a $10b AI business with an estimated spend of $80b. Except for Nvidia, most AI businessesis are losing money. And Microsoft is one of the better ones. Similar story at Alphabet, Meta and Amazon.

• The CAPEX is immense. The returns are hard to gauge. How much and when? What's the return on investments on these billions spent? Is there enough return to justify that level of investment (The AI ROI Debate)? It remains to be seen. Once the end user sees value, and he's willing to pay for it, it will drive ROI.

• There's a growing disconnect between what companies spend on AI architecture and revenues being generated from AI products and services. This might come into play eventually. Delivering on AI products and services is difficult.

• Productivity gains have to be clearly demonstrated. It will get there. But it will take time to see productivity growth.

• ChatGPT has been on the scene for 2 years. Progress has been slower than expected. AI was supposed to transform economies around the world. People's jobs were supposed to be at risk. These predictions are taking longer than expected. We are in the first inning of the AI era.

• Pre-ChatGPT: As compute increased, capabilities went up exponentially. This dynamic was given a term: “Scaling laws.” Right now they have hit a plateau. Researchers are looking for a breakthrough.

• Yandex N.V. reached ~$30bn market cap in 2021 (must take into consideration the divestment).

• With large massive infrastructure buildout, will demand meet capacity? (supply and demand rarely perfectly match)

• But if you take a company like Nebius, the spending infrastructure is backed by massive demand.

• “The risk of under-investing is higher than the risk of over-investing.” - Sundar Pichai, CEO of Google

CoreWeave (private)

CoreWeave is a close comparable. It's in the “GPU-as-a-service” space.

• CoreWeave expects $2b in revenue this year, up from $500m in 2023 (source The Information)

• Cisco invested in CoreWeave at a $23b valuation

• CoreWeave closed a secondary share sale also at a $23b valuation (Bloomberg)

• More than triple the value of the company in the past year. Up from $7b

• CoreWeave is Nvidia backed

• Possible IPO next year (2025)

• Raised $14b so far in equity and debt. Raised in $7.5b in debt

• 550 employees

• Data centers: 28 in 2024. 10 more in 2025

• Claims lower on-demand prices than any major cloud company

• Rents out GPU to companies.  “GPU-as-a-service”.

• Even Microsoft relies on CoreWeave for OpenAI. So is Mistral.

• Microsoft is expected to spend $10b on CoreWeave's data centers. (The Information)

• Signed a deal with Core Scientific for 500MW worth of data center capacity that could be with $8.7b over 12 years

• Benefiting from the growing interest in the business of cloud infrastructure for GenAI

• Cloud provider that builds clusters designed to power AI workload

• Makes it possible to use GenAI at scale

• Provides access to data centers and high-powered chips for AI workload

• Boom in AI applications such as ChatGPT and training LLMs has boosted demand for cloud-computing services

• How do you secure electricity supply in a shrinking power market?(Does it even matter what they pay for power in an AI race? Probably not.)

If we go with the idea that CoreWeave trades at ~10x sales, what does that mean for Nebius? 10x sales is a crazy multiple. You need growth to make it work. CoreWeave is growing really fast. Going from $500m in 2023 to an estimated $2b in 2024. Reuters reported that CoreWeave is looking at a $35b valuation target for its 2025 IPO.

What about Nebius? Nebius is newer and smaller. 2024E revenue is expected around $120m-$130m, up from $27m (new company). In their investor presentation Nebius stated an estimated $500-700m in revenue for 2025 (ARR: $500m-$1b). That's at least an increase of 5x in revenues and aim to become Adjusted EBITDA positive.

Let's assume the business gets to a run-rate of $1b in revenues, which is feasible in a year or two. What's that business worth? $5b? Add cash and subsidiaries. $7b? With a margin of safety you are getting a $30 share in a year or two. With the share trading at $20, that's a 50% upside. If you are looking for an AI position before it gets expensive, this is a good one.