For those who have followed me for a while, you will remember I tweeted a lot about Nebius Group early this year; I was very bullish. I took a big position in out of the money calls; I sold them in early June for a 570% gain. In June I wrote that I still liked the company; I was unsure what to do next. Buy stock; or go further out on options. Looking back, waiting and focusing on other names was not my best choice. If I had pushed into longer-dated strikes, my account would probably be up about +500% this year. From my Q1 entries the stock nearly 5x’d at the highs; the move has been strong. Even after that run, I see fresh upside. The reflexive loop is in our favor. The almost 20 billion dollar deal with Microsoft works as a great de risking; thus, I once again entered the stock.
Nebius Group is a spin-off from Yandex; often described as Russia’s Google. The business the market cares about today is AI infrastructure. In plain terms that means Nebius builds and runs data centers; fills them with the right chips; and rents out computing power so companies can train and use AI models.
“Full-stack” means one vendor handles the whole chain. Nebius secures land and power; designs and builds the site; installs servers with high-end GPUs; runs cooling; adds the software layer so customers can launch jobs and store results; and provides support to keep workloads online. Fewer vendors to coordinate; fewer failure points.
Most modern AI runs on Nvidia GPUs. Nebius buys and deploys those GPUs at scale; so customers can bring existing tools and models with minimal changes. That reduces integration risk; speeds up time to value; and helps keep utilization high.
Nebius also has proof that large buyers will use its capacity. A recent agreement with Microsoft signaled that the product meets top-tier requirements and that demand is backed by real contracts; not just interest.
Around the core there are smaller pieces. Avride builds software and testing for autonomous vehicles; think robotaxi and logistics. The data labeling arm cleans and tags datasets so models learn correctly. Managed databases such as ClickHouse let customers analyze data quickly without running their own servers. These support the ecosystem; they are not the main driver of the equity story right now.
In this article I will focus on the AI full-stack infrastructure; not because the other parts do not matter; but because investor sentiment today is centered on the infra engine that turns power and GPUs into cash.
Reflexivity and the business model
My understanding is that a full-stack AI infrastructure business like Nebius can land at roughly 20–30 percent FCF margins once sites mature; they have stated that the payback time per GPU is under 2,5 years; and that it should go lower as their services get refined. On paper this makes modeling look simple; take the cash they have; translate it into how many GPUs they can buy; apply the payback; and you get a view on revenue and free cash flow. In practice the key is the reflexive loop between belief; funding; capacity; and demand. If investors believe there will be demand for Nebius’s full-stack approach; Nebius can raise money; buy more GPUs; secure power; and build faster. The larger footprint then attracts bigger customers and longer contracts; that validation lowers funding costs; the stock moves up; and they can raise again on better terms. Belief funds capacity; capacity earns validation; validation attracts more belief.
This is exactly what they have been doing step by step. After the Microsoft deal they raised billions across debt and equity; yes this can imply dilution; but the important point is that the dilution is smaller than the cash flow generated by the added capacity if payback stays sub-2,5 years and ROIC clears the cost of capital. In that case per-share value still compounds. So the real question is simple; will there be enough demand for their products; will existing customers scale; and will there be more deals like Microsoft. If the answer is yes; the reflexivity does most of the work.
How big can inference and training get?
There has been a lot of guesswork about how large inference and training will become; I have seen numbers like 500 billion in a few years. Near term the spending picture is simple; policy pressure and flagship partnerships are pushing companies to commit massive capital; the checks keep getting bigger. The real question is the long term; and that is hard to model. Maybe algorithmic innovation makes LLMs much cheaper to train; or maybe chip advances unlock even more demand for AI.
Here is how I think about it. I already use AI for a lot; coding; researching stocks; drafting. If most people used AI to the same extent I do; we would probably already see something like a 100x demand jump. And I am not even close to the top of the usage curve. People are already running fully automated businesses that burn thousands in tokens every day. Then comes the deeper question; why do we need humans to do most cognitive work over the long run. A lot of it is repetitive; process driven; and boring. AI will take a growing share of that. Automating large parts of human cognitive work is worth a very large number.
And it will not stop at matching human capability. If we can exceed it; we will. As long as there are things we do not understand; diseases we want to cure; and discoveries we still want to make; capital will keep flowing into better models; better training runs; better inference footprints. That is the core of the demand view; the ceiling keeps moving up because the use cases compound and the value of faster; cheaper; smarter cognition compounds with them.
I do not think AI spending slows until it starts moving electricity prices in a meaningful way; models will keep improving; chips will keep improving; the bottleneck shifts to power generation and the grid. We converge toward a world where electrons, land, cooling, and transmission become the gating factors; we are not there yet. New data centers are still finding power; governments are aligning incentives; nuclear and renewables pipelines are growing; demand growth is still outrunning the constraints. Long term I see no reason the market would not sit in the multi-trillions; the spend spans training, inference, memory, networking, and the full physical stack that turns megawatts into tokens and outputs.
Where does NBIS land
So if sector upside is tied to how much energy we can produce; near term there is a lot of runway; long term it is almost unlimited. The next question is who captures that value. I think Nebius has good odds; for a few concrete reasons.
First; the talent pool. The core is ex-Yandex engineering; a large share are top Russian-trained engineers now building from hubs in the EU and Israel. That creates a differentiated workforce; deep systems talent at a lower all-in cost than many US peers; with a bias toward building real infrastructure rather than slideware. The integration challenge is real when you add hires elsewhere; but my net read is this is a competitive advantage. The company itself is the Yandex non-Russia spin now focused on AI; the structural break and relocation are done.
Second; the full-stack approach. Nebius does land; power; build; operate; and sells capacity up the stack; tightly coupled to Nvidia platforms. That lets them price sharply and still earn attractive paybacks because they control more of the bill of materials and the software layer that drives utilization. The partner language and positioning with Nvidia are public; the proof is that a top-tier buyer just chose them.
Third; validation and distribution. The Microsoft agreement was the step change; it moved them from “credible challenger” to “contracted supplier” for an A-grade buyer; and it anchored the capital cycle for the next leg of builds. With that in place, the bar for other hyperscalers; model labs; and national clouds to follow is lower. I would not be surprised to see similar agreements announced over the coming quarters.
Finally; funding reflexivity. Post-deal they tapped markets for scale; including convertible notes and an equity raise; which lowers their weighted cost of capital and supports faster GPU procurement. As long as incremental returns clear the coupon and the equity cost, per-share value compounds even with some dilution.
Put together; differentiated talent; full-stack cost position; marquee validation; and funded runway. That is a good place to land in a sector where the limiting factor is shifting toward electrons; not narratives.
Pricing the company
So now to the trickiest part; pricing. We need to balance the huge long-term upside with the real risks that still exist. First; technology regime risk. In any new field the underlying stack can change quickly; a breakthrough in models or hardware can reset the economics. Second; narrative risk. Sentiment can flip overnight; NBIS fell roughly 50 percent on the DeepSeek scare; it was irrational in my view, but it proved that reflexivity cuts both ways. Third; platform dependency. Ask around and most will say Nvidia will not lose share; over a long horizon no single vendor outshines forever; a leap in chips or even quantum could change the game. Fourth; execution. Power, builds, delivery, uptime; miss here and contracts wobble. Fifth; competition. Hyperscalers and new entrants will fight on price and bundles. Finally; capital markets. If panic hits, the funding loop reverses and multiples compress.
Against that backdrop my numbers. In Q1 2025 I marked fair value around 15 billion; 10 billion for the AI infrastructure segment; 5 billion for Avride, ClickHouse, and the rest. With Microsoft validating the stack and lowering funding risk, 10 billion for the core is now too low. My updated fair value is about 30 billion; reflecting a cleaner risk profile and a longer contracted runway. That still leaves upside from today’s price; especially if they keep landing A-grade customers and sustaining sub-2,5 year paybacks.
It is worth noting that a valuation of 30 billion could be conservative; if one or more Microsoft-sized deals land, fair value could push toward 40–50 billion almost immediately as contracted demand lowers risk and funding costs. On the other hand industry standards are not set; the sector could rhyme with memory in the 1980s. I think those odds are low; threat of new entrants is low; threat of substitutes is somewhat low. Intra-industry rivalry is medium; buyer and supplier power are high. Net-net, early-mover advantage plus Nebius’s unique talent pool should support okay margins long term; not astronomical ones. This is worth watching because once the market matures; or even if enough investors merely think it has matured; margin compression will enter the conversation; growth will be questioned; multiples can wobble. Volatility can be very high; the reflexive nature of financing cuts both ways.
End note
All in all I think Nebius is a very favorable position; they are in an exploding sector that I believe will keep compounding for a very long time. They seem to have leadership that knows what they are doing; and they appear to optimize capital for shareholders. There are risks; but after the great de risking and the clear affirmation that the tech works, I see the name as underpriced again. If you liked this write up, please share it and subscribe to my Substack.
Disclosure and disclaimer
This article is for information and education only; it is not investment advice; it is not a recommendation to buy or sell any security. Do your own research; verify all numbers and assumptions; consider your personal objectives; risk tolerance; and constraints. Past performance is not indicative of future results; forward-looking statements involve uncertainties; actual outcomes may differ materially. Investing involves risk; including the possible loss of principal. I may hold positions in securities mentioned; I may buy or sell at any time without notice. Nothing here constitutes legal; tax; or accounting advice. Consult a licensed financial adviser before making any investment decisions.