Nebius: is it all noise? Or did the bears get it right?

Since the (admittedly poorly timed in hindsight) bullish deep dive I did at the start of July on Substack, the AI world has seen a lot of headlines and Nebius has seen a ~40% drop in the share price. Spoiler alert: most of it, I think, is noise. But, here's what I think actually impacted the factors I mentioned in the deep dive needed to be tracked and where we stand today:

Factor 1: Demand vs. supply

The Meta scare: I talked about this one in the article as I published it shortly after this news came out. To recap: on July 1st Bloomberg said Meta was building a business to resell its excess compute; the neocloud names got hit and Nebius fell ~12%. The fear was mainly a headline that distorted the reality a bit, in my opinion. It basically came down to people reading the headline as "Meta has excess compute, demand is falling, supply is catching up". The truth was, Meta had zero excess compute, Meta was desperate for compute, and Meta just said that they thought compute was such a profitable long-term business, that they wanted to build an AWS-style business to eventually, at some point in the future, to be determined, sell any potential excess compute, bit of a different story. Low and behold, by July 9th Zuckerberg was on record saying Meta "needs all the computing power it can get," and by the 13th Meta had committed another $40bn or so to a single Louisiana site. The excess compute premise collapsed almost as fast as it appeared. On balance, I would say this whole story is mildly bullish for our first factor as it just shows that compute is constrained and demand is still there, no shock.

Amazon raised GPU prices: AWS is probably the biggest landlord on the block and they just raised their rents ~20%, it basically tells you the market can bear it which helps smaller landlords like $NBIS also raise their own. Obviously, this is also bullish and a confirmation of pricing power & demand.

A new >$1bn customer: Reflection AI signed a compute deal through 2029 with Nebius. A fresh name beyond Microsoft and Meta is exactly what you want to see off that concentration fear, so again, bullish. That being said, Reflection AI is nowhere near the level of blue chip customers like the Mag7, so their ability to actually pay is a (soft) question mark because they are a money-losing startup. They are however, backed by Nvidia and founded by two ex-Google DeepMind researchers, and they just partnered with the U.S. Dpt. of Energy, so there is some institutional safety net around them too.

Factor 2: Execution

The asset-light model (capital side): This one had people on FinX split down the middle between "Nebius has noticed their business model doesn't work and need to switch" to "Nebius is the smartest company in history and the next Nvidia". So what happened? Basically on July 15th Nebius unveiled a model where partners fund, own and operate the data centers, and Nebius supplies its architecture and software stack and sells the capacity for a cut. I must admit, I struggle to see the bear case on this one. It's not a business model switch, its a new division and revenue engine on top of the existing business.

I mean the single strongest bear case against this stock is capital. They have to pour billions upon billions on the buildout so there is a constant dilution fear. This new division lets them add capacity without carrying it all on their own balance sheet. So this is, in my view, bullish on the capital side of things but its not all sunshines and rainbows yet: Firstly, they haven't disclosed how the economics of it is going to work. So although people speculate that it will be high margin, we don't have a clue. Secondly, there was a line in the announcement that said that they are 'transferring knowledge to the partner' which I don't think is a smart move if you don't want said partner to eventually learn your trade and cut you out (Breaking Bad anyone?). But that is just my read, and I don't know the details here. Lastly, quality is the whole game for a neocloud, its one of their points of differentiation, as we covered in the deep dive. Problem is that a low PUE and reliability are much harder to guarantee on infrastructure you don't run. So bullish on the capital side but overall, I say this is one to Watch.

$775m of non-dilutive debt: Nebius has recently raised GPU-backed debt that funds the buildout without issuing shares, this helps cut straight at the dilution worry, and it shows that there is bank appetite for funding this expansion. We just have to watch that that debt doesn't run out of control and we run into another CoreWeave situtation. Bullish.

The Alabama lawsuit: Neighbours near the Birmingham site (300MW) are suing to pause it, and a judge let the case proceed. On its own it doesn't move the needle, its a small node in a 3.5GW+ pipeline. But it's a reminder that building in America comes with permitting fights Finland never threw at them. We mention in the article that their American buildout was a risk because it was a new market they didn't know as well as the Finish or European ones more broadly. Not an alarm but mildly bearish / watch.

Factor 3: Revenue mix

The asset-light model (platform side). The same announcement can also be seen as a quiet vote of confidence in the part of the thesis I find hardest to have visibility over now: how much of future revenues will come from Token or the managed platform vs raw GPU rental? This announcement is a vote of confidence that a bigger chunk might come from the higher margin token side because this whole thing only works if third parties believe the Nebius stack, not the raw GPUs, is what's valuable and it monetizes that software layer directly, with capital-light, higher-margin economics because they don't have to own the expensive depreciating assets under. Bullish for the thesis but again, the terms and take-rate haven't been disclosed, no locked-in partners have been disclosed either, so we're still running on conviction, not proof yet.

Kimi K3 (the new deepseek): Remember when deepseek came out and everyone said western AI companies were all dead because this new Chinese model is cheaper to build and run so everything sold off significantly, including Nebius (which went down fairly hard)? Well it has happened again, except its now called Kimi K3. So everyone is freaking out, except, it turns out Nebius Token Factory's entire premise is serving open-weight models by the token. Its value goes up precisely when open-weight models get good enough that enterprises will actually run them instead of paying for a closed API from OpenAI or Anthropic. A frontier-class model you can download for free (Kimi K3, Llama, Qwen, DeepSeek, Mistral, GPT-OSS), all of which Nebius already serves, is literally the single most direct catalyst for that.

The deepseek precedent supports this: cheaper, better models don't shrink compute demand, they expand it, this is the whole Jevons point we talked about in the deep dive. More capable open models pull the value of the AI stack away from the model owners and toward whoever runs the models efficiently and close to demand. That's the neocloud, and more precisely, that's Nebius, because Nebius is the one offering Token Factory while many other neoclouds just sell the raw compute. So the only way to interpret this in my view is as bullish.

One small bearish note to caveat: The efficiency narrative does have one point negative for Nebius over the long-term: it's the training-capex story. If frontier models keep getting dramatically cheaper to build, the "infinite training demand" that underpins the mega dedicated contracts (Microsoft, Meta) gets a question mark over the long run. I don't think it bites near-term at all as those contracts are locked, and demand still outruns supply by a lot but intellectually that's where a Kimi/DeepSeek world could pressure Factor 1 even as it helps Factor 3.

That being said, there is a bullish caveat to the bearish caveat: training is only one part of the equation, and inference is the other. So as models get cheaper to run, and demand to run them grows, demand for inference-based compute grows. And if you recall from my deep dive, inference is easier to run on older chips, which is bullish for the useful economic life of chips, which was one of the main bearish arguments against neoclouds. Bears say that chips have a short useful economic life of 2-3 years and that the capex and dilution will never end, more inference in the future attacks that directly as chips can now keep being useful and earning dollars for longer (maybe 7, 8, 9 years), through inference, thanks to open-weight models.

Net across the two: I'd call it a clear positive bullish note for Nebius, weighted towards confirming the Token Factory bet, with a slight slow-burn asterisk on the training side counterbalanced by the inference side. One smaller (probably insignificant thing to keep in mind): it's a Chinese model, so some Western enterprises will hesitate to run Kimi specifically but the trend (open-weight catching the frontier) is what matters, and that's model-agnostic.

Factor 4: Power, grid & policy (new this month)

A lot of the names sold off, primarily because of the large tech/AI sell-off but the NY moratorium news, helped exacerbate that as people feared the buildout would fizzle out as more states followed NY's lead. I figured I would add a new factor to watch and it's power, grid & policy. Not because I think this news is thesis-breaking or anything, but more so because execution on Nebius' side is one thing, whereas the rules, power and policies that allow for it and are outside of Nebius' control are also worth monitoring separately.

New York's moratorium: Let's start with this one, as all the bears were celebrating the death of the data center expansion, everyone is about to ban new data center construction! The buildout is dead, neoclouds are dead, its all over, everyone is following New York's lead, right? Well, I don't know. Let's start with the basics, what happened?

On July 14th Governor Hochul signed the first statewide moratorium on new large data centers in the US. Here's the first thing though, notice the word 'moratorium'? I saw most people calling this a 'ban' but that's not what a moratorium is. Call me a stickler but they're actually very different things. A ban implies a total ban forever, or until someone unbans it. What this moratorium is is a one-year pause on permits while the state writes tougher rules. The headlines called it a ban and the sector wobbled, but I think the reaction is overblown: it's a pause, not a ban. The most likely outcome in my view is that new data centers will start getting approved after the moratorium as the state will have agreed on the stricter rules it wants them to comply with. New York regulating things, what's the news again?

Also to note: Nebius doesn't build in New York (its US sites are Missouri, Pennsylvania, Alabama); and it's a global company whose best asset is in Finland. Would France or Finland care what one US state did and start following suit? I don't think so. Will red states follow? Pretty hard to imagine. Will even most or a large chunk of blue states follow? Tougher to call outright but if I was a betting man, I would bet no, most blue states won't follow either, at most, some other blue states might follow, probably not even with the moratorium but with copying the stricter rules, and then more capital will flow out of them (or never get to them at all). Again, what's the news? Overall, this might seem headline bearish, but its largely noise for Nebius.

The real risk here: Despite all my snarky remarks, there is one thing that could be a risk for the buildout that came out of the NY moratorium. The actual driver might partially be politics, but it's also electricity costs, and that anger is bipartisan because it hits your wallet no matter which way you vote. Virginia (purple) just taxed data-center power, and Georgia, Texas and Oklahoma are all moving on who pays for the grid to protect consumers. So there is a genuine slow-burn risk with power getting pricier across the industry, nibbling at the cheap-electricity margin a lot of the neoclouds depend on. Mildly bearish on margins over time.

But there is another bullish caveat here because it cuts both ways, because if building gets harder and dearer, whoever already holds secured, largely-owned power (like Nebius' 3.5GW+, >75% owned) quietly gets a wider moat as demand keeps growing and supply keeps getting harder to, well, supply. One more thing, remember how Nebius' edge was not just its efficiency at operating data centers but on the stack on top on which this new asset-light business model depends on? Well, if margins compress on raw-GPU rental because of higher electricity costs, what neocloud do you think is better positioned to survive and thrive, the one selling raw-GPU rental? Or is it the one whose moat comes from the stack on top? Food for thought, but overall, watch the margin, not the ban.

So where does all this leave us?

Zoom out and it's kind of funny: in three weeks the market threw four separate scares at the Nebius thesis: Meta the competitor with excess compute, the capital/dilution problem, a Chinese model that "kills" AI, and the death of the data-center buildout. All of which knocked a big chunk off the price. Work through each one against the factors, though, and not a single one lands the way its headline wanted. Meta turned out to be starved for compute, not drowning in it. Asset-light and the $775m of non-dilutive debt go straight at the capital fear that was the single strongest argument against the stock. Kimi K3 is arguably the most bullish thing that happened all month for the part of the thesis I care most about (margin-expansion via token factory). And the New York "ban" is a one-year pause in a state Nebius doesn't even build in.

If anything, the common theme here makes me more confident, not less: almost every one of these stories rewards the same thing: the stack on top, not the raw GPUs underneath. Asset-light only works if the software layer is the valuable bit. Token Factory only wins if open models get good, and they just did, again. And if power costs squeeze raw-rental margins across the industry, it's again the stack-heavy operator that comes out ahead. That's not four unrelated events; it's the same bet being validated from four directions.

So, more bullish? Yes, certainly, but we still need further evidence when the new earnings come along to prove our thesis. Here's the part that matters to me though: the thesis got stronger while the stock got a lot cheaper (we're under $180 now, versus ~$220 something when I published). Fundamentals up, price down is usually the setup you want, not the one you run from. It doesn't make the risks vanish obviously. The asset-light economics are still undisclosed, the training-capex question is real over the long run, and this thing trades like a high-beta rollercoaster, so if a 40% drawdown ruins your week, it isn't for you. But against the bear/base/bull numbers from the deep dive, the risk/reward at these levels looks better to me than it did at $220, a lot better.

I'm holding, and I added. Sadly, I added in the $190's as I never expected to see it down in the $160's and $170's. But its impossible to perfectly time the bottom, although this certainly seems like it (don't quote me on that though). In any case, it won't matter in 2, 3, 4, 5 years if the thesis plays out and the stock doubles, triples, or more. As always, this is just me thinking out loud, not financial advice. Always do your own work. Next real checkpoint is the Q2 report in August, and I'll be hunting for any colour on those asset-light economics, token factory revenue growth, and proof of execution. Those are what matter in this story, the rest is noise.