Sorry to double dip on comments here, but I also meant to ask your thoughts on google’s TPU based structures? Do you see any possibility that these could essentially make all other AI providers essentially obsolete?
Google's TPU-based infrastructure is definitely powerful and highly optimized for AI workloads, especially at scale. But I don’t see it making other AI providers obsolete. Here's why:
1) Not every AI company or workload fits neatly into Google's ecosystem. Many teams value flexibility — especially when working with open-source frameworks or custom hardware. Plus, Google's TPUs aren’t as widely available to third parties as Nvidia's GPUs. Most companies can’t build their own infrastructure around TPUs the way they can with off-the-shelf components.
2) Many enterprises today don’t want to be locked into a single provider. That opens the door for players like Nebius to offer tailored solutions, competitive pricing, and more flexibility. Some companies also need cost-efficient infrastructure or custom deployments — something Nebius can provide, especially for smaller or regional clients who aren’t top priorities for the hyperscalers.
3) While TPUs excel at certain AI workloads (like training large language models), they’re not always the best fit for general-purpose compute, edge inference, or smaller models.
4) The AI infrastructure market is growing rapidly, with room for multiple winners. TPUs might capture share at the high end, but there’s still massive demand across the spectrum — from startups to governments to specialized research.
Thank you, you give me some confidence on holding Nebius during this volatile period. Do you think the tariff has a large impact on Nebius, especially when importing GPU from Taiwan. Many thanks
1GW of compute would generate significantly higher revenue than you projected. Each GPU has a power capacity of 1.2-1.3KW and rents at 3-4$ per hour for older chips, and 6-8$ per hour for newer blackwell chips. Am I missing out on something here?
Even if the entire GPU mix is H100 and H200, at 2.95 - 3.5$ per hour (pricing as per NBIS website) revenue would be significantly higher ($20B+ at 70% utilization)
Question: What is the GPU makeup for NBIS and is it advantageous or disadvantageous? Much has been made recently of CW having 250K Hopper chips as a risk ... and how Blackwell could quickly lower the value of those older generation chips with Jensen saying "Blackwell is 40 times the performance of Hopper." So, I'd love to know more about what clusters Nebius has in production and what is coming online.
During last month's NVIDIA GTC Conference, Nebius announced that it will be an early adopter cloud provider of the NVIDIA Blackwell Ultra platform, offering NVIDIA GB300 NVL72-powered instances. General availability of NVIDIA Blackwell GPU capacity at U.S. data centers is expected to begin in Q2. Additionally, a deployment of the NVIDIA HGXB200 at the company’s Kansas City GPU cluster is expected to be available to customers in Q2 2025.
Great article mate really enjoyed reading thanks for taking the time to put it together. If I come into money in the future you are at the top of my paid subs list
thanks for the deep thoughts and sharing, but 3 main points i can't agree although I also built some positions on this stock before recent dip. 1) given the subsidiaries are at very early stage, can't agree to use sum of parts to value, so only core business should be considered; 2) for neocloud business, unit economics for small clients/short-term contract vs. large/long-term contracts are total different, so don't think Nebius' diversified and small clients focused strategy is a advantages vs. Coreweave, nobody will give up large/long-term contracts if they can, that doesn't make sense, for example, for a 3/5 year contract vs. 1 year contract, with same price, EBITDA margin will be significant lower for short-term conracts, which means coreweave has more flexibility to keep profit with even lower price, that maybe the reason why everybody says Nebius has cost efficiency but it's still loss making. 3) same as point 2, it's not conservative at all to assume a 25% EBITDA margin by comparing to AWS and coreweave, unit economic for AWS/coreweave (large players for large clients with 3/5/10 long contracts) is totally different. just my personal view, maybe totally wrong, but hope to give you a different view, thanks.
1) These subsidiaries are not that "early-stage" — most of them were founded over 10 years ago within Yandex. But even if you exclude them from my valuation model, you still get a significant upside.
2) I'm not sure if you understood what I meant, but Nebius is not focused only on small clients, it's focused on every single potential customer. Its solutions are tailored for all clients, whether small or large, while CoreWeave is particularly attractive to large clients — that's what I tried to explain. The reason Nebius is still "loss-making" is purely due to scale... the company has just started ramping up contracts.
3) I have to disagree with you here. A 25% EBITDA margin seems like a fairly conservative estimate. I'm following your reasoning and assuming a much lower value than hyperscalers and CRWV get — less than half, actually.
Again, thanks for your comment. Debating contrarian opinions is always a pleasure, so I really appreciate it!
I discovered you on X through my own research on Nebius. I opened my first position in early December 2024 and slowly expanded over the months. You write great, clear analyses that match my own observations.
My greatest conviction is in the combination of the various components of Nebius. AV, EdTech and Datacenter capability. For me it is a fullplay AI bet. Notably, Avride and TripleTen give me the conviction to invest in Nebius. I still have some doubts about the data center business. How do you view the company if it consisted only of the data center business, i.e. without Avride, TripleTen and Tolonka. Would you still have as much conviction then?
2) As I mentioned in the article, Nebius will likely burn through that cash quickly during its expansion phase, so I don’t think it’s prudent to use it as an anchor for the stock’s price.
Really well done. I bought this mid twenties months ago and wound up selling out at a small profit after lots of volatility.
I see little doubt that this will trade higher in the long term, however, I am most concerned about the current macro environment muting any positive progress these guys make. Eventually they’ll stand out, but it could take a while if tech stocks continue to get demolished.
Thanks Doug! :) I'm patient, so I'm not worried about external news driving the stock price down, as long as the execution stays on track like we expect.
Thank you, wonderful read
Thanks Sir, I appreciate it!
Sorry to double dip on comments here, but I also meant to ask your thoughts on google’s TPU based structures? Do you see any possibility that these could essentially make all other AI providers essentially obsolete?
No worries at all, appreciate the engagement!
Google's TPU-based infrastructure is definitely powerful and highly optimized for AI workloads, especially at scale. But I don’t see it making other AI providers obsolete. Here's why:
1) Not every AI company or workload fits neatly into Google's ecosystem. Many teams value flexibility — especially when working with open-source frameworks or custom hardware. Plus, Google's TPUs aren’t as widely available to third parties as Nvidia's GPUs. Most companies can’t build their own infrastructure around TPUs the way they can with off-the-shelf components.
2) Many enterprises today don’t want to be locked into a single provider. That opens the door for players like Nebius to offer tailored solutions, competitive pricing, and more flexibility. Some companies also need cost-efficient infrastructure or custom deployments — something Nebius can provide, especially for smaller or regional clients who aren’t top priorities for the hyperscalers.
3) While TPUs excel at certain AI workloads (like training large language models), they’re not always the best fit for general-purpose compute, edge inference, or smaller models.
4) The AI infrastructure market is growing rapidly, with room for multiple winners. TPUs might capture share at the high end, but there’s still massive demand across the spectrum — from startups to governments to specialized research.
Excellent, thank you
Very insightful & well written article - came across this as part of my research on $NBIS. Opened a position today.
Thanks Sir, I appreciate it!
Thank you, you give me some confidence on holding Nebius during this volatile period. Do you think the tariff has a large impact on Nebius, especially when importing GPU from Taiwan. Many thanks
Thanks a lot Ren! What Nebius imports from Taiwan are server racks, not GPUs, so I don’t think it’ll have a significant impact.
Great read, thank you
Thanks for reading!
1GW of compute would generate significantly higher revenue than you projected. Each GPU has a power capacity of 1.2-1.3KW and rents at 3-4$ per hour for older chips, and 6-8$ per hour for newer blackwell chips. Am I missing out on something here?
Even if the entire GPU mix is H100 and H200, at 2.95 - 3.5$ per hour (pricing as per NBIS website) revenue would be significantly higher ($20B+ at 70% utilization)
As I said multiple times throughout the article, the point was to be as conservative as possible! I also believe the upside is much higher.
Question: What is the GPU makeup for NBIS and is it advantageous or disadvantageous? Much has been made recently of CW having 250K Hopper chips as a risk ... and how Blackwell could quickly lower the value of those older generation chips with Jensen saying "Blackwell is 40 times the performance of Hopper." So, I'd love to know more about what clusters Nebius has in production and what is coming online.
During last month's NVIDIA GTC Conference, Nebius announced that it will be an early adopter cloud provider of the NVIDIA Blackwell Ultra platform, offering NVIDIA GB300 NVL72-powered instances. General availability of NVIDIA Blackwell GPU capacity at U.S. data centers is expected to begin in Q2. Additionally, a deployment of the NVIDIA HGXB200 at the company’s Kansas City GPU cluster is expected to be available to customers in Q2 2025.
Great article mate really enjoyed reading thanks for taking the time to put it together. If I come into money in the future you are at the top of my paid subs list
Thank you very much for the support brother! :)
thanks for the deep thoughts and sharing, but 3 main points i can't agree although I also built some positions on this stock before recent dip. 1) given the subsidiaries are at very early stage, can't agree to use sum of parts to value, so only core business should be considered; 2) for neocloud business, unit economics for small clients/short-term contract vs. large/long-term contracts are total different, so don't think Nebius' diversified and small clients focused strategy is a advantages vs. Coreweave, nobody will give up large/long-term contracts if they can, that doesn't make sense, for example, for a 3/5 year contract vs. 1 year contract, with same price, EBITDA margin will be significant lower for short-term conracts, which means coreweave has more flexibility to keep profit with even lower price, that maybe the reason why everybody says Nebius has cost efficiency but it's still loss making. 3) same as point 2, it's not conservative at all to assume a 25% EBITDA margin by comparing to AWS and coreweave, unit economic for AWS/coreweave (large players for large clients with 3/5/10 long contracts) is totally different. just my personal view, maybe totally wrong, but hope to give you a different view, thanks.
Thanks for your feedback Jack!
1) These subsidiaries are not that "early-stage" — most of them were founded over 10 years ago within Yandex. But even if you exclude them from my valuation model, you still get a significant upside.
2) I'm not sure if you understood what I meant, but Nebius is not focused only on small clients, it's focused on every single potential customer. Its solutions are tailored for all clients, whether small or large, while CoreWeave is particularly attractive to large clients — that's what I tried to explain. The reason Nebius is still "loss-making" is purely due to scale... the company has just started ramping up contracts.
3) I have to disagree with you here. A 25% EBITDA margin seems like a fairly conservative estimate. I'm following your reasoning and assuming a much lower value than hyperscalers and CRWV get — less than half, actually.
Again, thanks for your comment. Debating contrarian opinions is always a pleasure, so I really appreciate it!
Amazing article M.V., I would personally be very excited about you writing an article on Avride.
Thanks for your feedback Joshua!
I discovered you on X through my own research on Nebius. I opened my first position in early December 2024 and slowly expanded over the months. You write great, clear analyses that match my own observations.
My greatest conviction is in the combination of the various components of Nebius. AV, EdTech and Datacenter capability. For me it is a fullplay AI bet. Notably, Avride and TripleTen give me the conviction to invest in Nebius. I still have some doubts about the data center business. How do you view the company if it consisted only of the data center business, i.e. without Avride, TripleTen and Tolonka. Would you still have as much conviction then?
Thanks a lot Simon! My conviction would be the same.
Great article. I started a position based on your first article. I might add some more. Thanks for sharing comprehensive research! Highly appreciated
Thanks Sunny, I appreciate your support!
1) Could you compare INOD and Toloka businesses?
2) if they have 2.4B in cash, the stock price below $10 makes no sense, right?
1) I'll take a look at INOD.
2) As I mentioned in the article, Nebius will likely burn through that cash quickly during its expansion phase, so I don’t think it’s prudent to use it as an anchor for the stock’s price.
Thank you.
Thanks, this is great as always! Having an Avride article would be amazing.
Thanks a lot Carlos! :)
Really well done. I bought this mid twenties months ago and wound up selling out at a small profit after lots of volatility.
I see little doubt that this will trade higher in the long term, however, I am most concerned about the current macro environment muting any positive progress these guys make. Eventually they’ll stand out, but it could take a while if tech stocks continue to get demolished.
Thanks Doug! :) I'm patient, so I'm not worried about external news driving the stock price down, as long as the execution stays on track like we expect.