$DELL ON AI INFERENCE DEMAND: “Inference has passed training and is pure demand on our industry” Dell expects inference token demand to grow 87x to 3,600 quadrillion tokens by 2030, while training demand grows 5x “Enterprise agentic to be the single largest workload by 2028”

$NBIS Update: Pennsylvania Gigafactory Phase III zoning is OFFICIALLY APPROVED‼️ I called again and managed to get a hold of the Butler Township Supervisor Chairman Mr. Kessler. We had a nice chat, and he confirmed tonight's ordinances passed 2 to 1. Conversation notes 👇 1⃣ Selection Bias Mr. Kessler told me that out of Butler Township’s ~5,000 residents, only 12 people showed up to tonight’s meeting. Vineland has ~60,000 residents. Maybe 100 showed up to protest at the site amendment hearing. Noise ≠ consensus. The loudest voices are almost never representative of the broader community. Most residents are indifferent, supportive, or just not sufficiently opposed to bother showing up. 2⃣ Butler Township WANTS This Datacenter Kessler told me Pennsylvania is “screwy” right now (with Shapiro), but his view is that datacenters are coming regardless. Eventually, municipalities across the state will be asked to accommodate them. Butler Township decided to get ahead of that. Two years ago, the township passed an ordinance designating where datacenters could be built, deliberately selecting an area along I-81 with substantial separation from nearby residents. Rather than wait for the state to come in and say, “We’re putting a datacenter here,” Butler wanted to already be able to say: “We have a spot.” 3⃣ Noise Study Before the vote, Butler Township commissioned an independent noise study to model the datacenter under a range of operating conditions. Kessler said the analysis incorporated roughly 15 years of local temperature data, because the cooling equipment becomes louder as temperatures rise. He noted that temperatures high enough to push the facility toward its loudest operating conditions occurred only three times over that 15-year period. The study modeled the facility at or near full cooling load during extreme heat, including temperatures approaching the local historical high of roughly 101.7°F. It also modeled how noise attenuates with distance from the site. Kessler said the nearest property owner is approximately 3,500 feet away, and cited modeled levels falling to roughly 52 dBA during daytime, and into the mid-20 dBA range at night farther from the facility. The study gave the board enough comfort to tighten the allowable noise limit from 75 dBA to 65 dBA.

🚨 Rumors that $NBIS is eyeing Vienna for a major new AI data center. According to Trending Topics, often described as Austria’s TechCrunch, Nebius is reportedly assessing local AI demand and scouting potential sites after Vienna shelved its EU AI Gigafactory bid. A roughly 300 MW facility is said to be under discussion, which would make it one of Europe’s largest AI infrastructure projects. Nebius declined to comment on Vienna specifically, but confirmed it is continuously evaluating new locations based on power availability, scalability, sustainability and local interests. Still early and unconfirmed, but potentially a major European expansion for $NBIS. h/t @David917S great find!

$NBIS announces proposed private offering of $4.5B of convertible senior notes. • $2.75B of notes due 2030 • $1.75B of notes due 2034 • Potential for an additional $675M Key terms, including the interest rates and conversion prices, will be determined at pricing.

I just published a new article: $TWST: Investment Thesis Explained A fascinating picks-and-shovels play sitting at the intersection of AI and biotech. I truly enjoyed researching this one!

DA Davidson lowered its PT on $NBIS from $250 to $175. The firm believes the current state of Vineland creates enough execution risk to question the year-end ramp and the company's valuation premium. I've read the entire report, so here are my overall thoughts on the issue: ⬇️ First, let's look at the analyst arguments. 1) The Vineland concern is about the gap between connected power and actual revenue-generating capacity. Nebius needs to scale from ~170MW of connected power at the end of 2025 to 800-1,000MW by the end of 2026, while increasing its ARR from $1.9B in March to $7–9B in December. According to DA Davidson, Vineland is a critical piece of that ramp: the site is expected to grow from ~50MW to 328MW. The firm visited the site and attended the recent public hearing (which was open to the public and livestreamed online, so that's not exactly an edge). Their concern is that permitting has already taken longer than expected, including hearings that had previously been postponed, and, based on the construction progress they observed, they don't believe Vineland currently looks positioned to easily reach 328MW of active power by year-end. There's an important distinction here: Bloom Energy fuel cells are reportedly arriving at the site, so Nebius may still be able to classify that capacity as “connected” by year-end. Davidson's concern is whether the actual data center infrastructure will be ready to turn that connected power into active, revenue-producing capacity quickly enough to support the $7–9B December run-rate target. So their thesis isn't simply “Vineland won't have power.” It's that power availability may run ahead of the site's ability to monetize that power. 2) The PT cut is mostly a valuation-premium argument, not just a Vineland estimate cut. DA Davidson believes $NBIS currently receives a premium because investors view Nebius as the highest-quality neocloud: better execution, a stronger capital structure, a more complete software/infrastructure stack and valuable non-core assets. That reputation has translated into $NBIS trading at roughly 0.98x EV/backlog, versus 0.8x for CoreWeave, meaning Nebius carries about a 21% premium to $CRWV. Their argument is basically this: If Vineland experiences a meaningful delay, it would call into question one of the main reasons investors have been willing to pay that premium: Nebius' perceived superior ability to turn contracted power into connected capacity and ultimately into revenue. Therefore, Davidson believes that if those execution issues materialize, $NBIS should trade closer to $CRWV's valuation, rather than maintaining its current premium. That's the core reason they cut the PT from $250 to $175. They also lowered their financial estimates and are now below consensus, particularly for 2027, while pointing to $NBIS' still-negative operating margins as another reason not to give the company the benefit of the doubt indefinitely. Now, my thoughts. 1) I think Davidson is underestimating the diversification of Nebius’ actual deployment pipeline. Davidson mentioned that, when asked about Vineland, management reiterated guidance and said it remains on track to meet all of Microsoft’s contractual commitments. The analyst also explicitly leaves open the possibility that other sites, higher monetization, or developments not yet disclosed could offset weakness at Vineland. That’s where I think they’re underweighting the evidence. In the report, Davidson includes a table showing what it considers Nebius’ data center pipeline across the U.S., Europe and the Middle East. The problem is that, based on my research and that of other members of the community, I believe that picture is incomplete. There's credible evidence pointing to several additional locations and expansions that haven't yet been formally announced, including Estonia, Wales, Spain, India, Singapore, further UK expansion, etc. Yes, some of them will only be deployed in 2027, but there’s also evidence that some could be deployed as early as this year. I understand why Davidson wouldn’t include unconfirmed locations in a formal research model. But if you’re going to make concentration risk one of the central arguments behind a 30% PT cut, then the completeness of the pipeline you’re analyzing matters a lot. And just as Davidson went looking for evidence that Vineland may be delayed, there's also credible evidence available suggesting that Nebius’ broader deployment footprint is much larger than what appears in their table. This matters because, if Vineland is indeed experiencing some delay, something even Davidson cannot say with certainty, while management is still reiterating company-wide guidance and Microsoft commitments, one plausible explanation is that Nebius has enough flexibility across the rest of its deployment pipeline to compensate for timing issues at a single site. This is also consistent with what management has emphasized before: geographic and portfolio diversification is specifically intended to reduce dependence on any one location and mitigate the impact of individual project delays. Funny enough, Davidson makes essentially the same argument when discussing hyperscalers, writing that they're less exposed to individual site delays because they're diversified across many more locations. If Davidson had visibility into Nebius’ full pipeline rather than only the publicly confirmed portion, I wonder whether they would view that concentration risk differently. 2) A site visit from the outside doesn't tell you the commissioning timeline. Davidson repeatedly emphasizes that it visited Vineland and, based on what it saw, doesn't believe the site can “easily” become 328MW of active power by year-end. That is a relevant observation, and I don’t think it should be dismissed. But there’s only so much you can infer from looking at a data center construction site from the outside. I looked closely at the photos included in the report from the "site visit." They don’t appear to have even entered the parking lot, let alone the facility itself. You can see that construction is ongoing, but you can’t see anything that gives you meaningful visibility into the deployment timeline. So I think calling this a “site visit” gives the observation more authority than the photos themselves justify. What ultimately determines when capacity becomes revenue-generating is the commissioning schedule: energization, electrical and mechanical fit-out, equipment installation, phased activation, and the timing of customer workloads. The report doesn't demonstrate that Davidson has direct visibility into those internal schedules. So a site visit can support the conclusion that execution risk exists. It cannot, by itself, establish exactly when specific MW will become active and monetized. And I think Davidson’s own wording reflects that uncertainty. They don’t say Vineland cannot reach 328MW of active power by year-end. They say they don’t believe the site can “easily” be transformed into 328MW of active power by then. It was never supposed to be easy. 3) There’s a big difference between a delay and a thesis-breaking delay. I’m not dismissing the possibility that Vineland is delayed. Davidson is mentioning a credible execution risk that we all have been monitoring for quite some time. But the real question isn’t whether every MW at Vineland becomes active on the exact originally expected timeline. The real question is whether any delay is significant enough to: - make Nebius miss its 800–1,000MW connected-power target; - make it miss the $7–9B December ARR guidance; - interfere with Microsoft’s contractual commitments; - or materially change the economics of the capacity being built. Those are very different outcomes. Davidson hasn't demonstrated that any of them will actually happen. In fact, again, its own report says management reiterated guidance and said Nebius remains on track to meet all of Microsoft’s contractual commitments. So even if we assume Vineland is running behind schedule, that doesn't mean it's material at the company level. And Davidson’s own estimate revisions suggest they’re modeling a relatively small timing impact. 4) The numbers barely changed. The narrative did. DA Davidson cut its PT by 30%, from $250 to $175. But look at what they actually did to their revenue estimates: • 2026 revenue: $3.36B → $3.303B, a reduction of only ~1.7% • 2027 revenue: $10.604B → $9.874B, a reduction of ~6.9% So this isn’t really a story about Davidson dramatically changing its expectations for Nebius’ underlying business (that's why I said their own revisions suggest only a small delay). The overwhelming majority of the PT reduction comes from multiple compression. Their argument is that, if the execution concerns around Vineland materialize, $NBIS should no longer deserve its current premium and should instead trade closer to $CRWV. Davidson points out that Nebius trades at ~0.98x EV/backlog versus ~0.8x for CoreWeave, and effectively uses convergence toward $CRWV's multiple to justify the new $175 target. But I think EV/backlog is far too simplistic a metric to determine whether two companies deserve the same valuation. Backlog tells you how much revenue has been contracted. It tells you much less about the quality and economics of that backlog: margins, contract duration, CapEx required to fulfill it, financing costs, utilization, revenue per MW, customer quality, or ultimately the returns generated on the capital deployed. Two companies can have the same amount of backlog and that backlog can have very different economic value. That said, I don't want to make this a discussion about $NBIS vs. $CRWV. It’s fair if one believes their multiples should converge. The point is that Davidson itself has historically argued that Nebius deserves a premium because of characteristics that have nothing to do with simply having more backlog: its stronger capital structure, full-stack offering, management track record and valuable ancillary businesses such as its ClickHouse stake. I'm using their words. None of those things suddenly disappear because one facility may take slightly longer than expected. [Not to mention that their PT gives no explicit value to Nebius’ non-core business units.] If Vineland ultimately causes Nebius to miss guidance, fail to meet contractual commitments or proves that its execution advantage was overstated, then I agree some de-rating would be justified. But going from “there may be a delay at Vineland” to “Nebius should effectively lose the valuation premium we previously argued it deserved” is a much bigger leap. Davidson also points to Nebius’ still-negative operating margins and broader uncertainty around the neocloud business model as additional reasons for caution. Those are legitimate risks, but they’re not new ones. Nebius also had negative operating margins when Davidson’s PT was $250. The uncertainty around long-term neocloud profitability existed then too. Overall, the valuation framework seems to fit the typical sell-side analysts’ playbook: use the assumptions that best fit your narrative. FINAL THOUGHTS Overall, I think Davidson mentioned a legitimate execution risk at Vineland, but then used several weak arguments to make that risk look much more consequential than the evidence currently supports. To be clear, I don't really care about the $175 PT itself. An analyst can change a PT very easily by changing the multiple they apply. What I care about is the narrative being built around it. An outside site visit isn't visibility into the commissioning schedule. Vineland also isn't binary, it doesn't have to be either 50MW or 328MW by year-end, as capacity can come online progressively. And while Vineland is clearly important for the near-term 2026 ramp, zooming out, the full facility represents only ~8% of Nebius' >4GW YE2026 contracted-power target. It's also not part of the owned-facility portfolio that represents the large majority of Nebius' contracted capacity and, in my view, matters much more to the long-term thesis. That's exactly why diversification matters. Nebius has consistently built its infrastructure portfolio across sites, geographies and deployment models to reduce dependence on any single project. If Vineland slips by a few weeks or months but other deployments compensate and company-wide guidance remains intact, I don't see that as particularly relevant. The threshold for me is simple: does Vineland cause Nebius to miss its connected-power guidance, the $7–9B December ARR, or Microsoft's contractual commitments? So far, we have no evidence of that. Management has actually reiterated all of those commitments Davidson asked about. And that's ultimately my issue with the report: the risk is real, but I think the conclusion is much stronger than the evidence. The financial estimates barely changed, while the PT fell 30% through multiple compression and previously known bearish arguments suddenly received much more weight. Fortunately, we should get much more clarity next week with earnings. To me, this remains exactly what it was before this report: a short-term risk worth monitoring, not a reason for me to change the thesis. I’ll publish my full preview of next week’s earnings over the weekend.

As a $NBIS shareholder, I really love this. Higgsfield AI is one of the best AI video tools out there, and in this short interview, its Founder explains why they chose $NBIS over other GPU cloud providers. 👀 Key takeaways: - Nebius is very friendly to start-ups. If you talk to developers off the record, you'll often hear that larger cloud providers have unclear terms, while Nebius offers full transparency and fair terms with its on-demand GPU offering. As I explained in my previous article, this is particularly attractive for start-ups and smaller developers because they can't predict exactly how their workloads will change. "Nebius has much better flexibility than any other cloud." - Larger cloud providers typically aren't interested in start-ups — they prefer higher-volume customers. "Nebius provides infrastructure and reliability at the same level as Google or Amazon." Again, just like I said in my last article, hyperscalers are not a threat because they often serve different needs. - Nebius uses a pay-per-token model, and the AI Studio is highly versatile in terms of model offerings, supporting everything from ChatGPT to DeepSeek and other LLMs. - It's very easy to work with Nebius. The team is highly responsive, unlike larger cloud providers. - "Nebius' flexibility, combined with its strong support, will likely attract many media companies." - Other cloud providers have lengthy agreements — sometimes up to 200 pages — while Nebius keeps it simple with 3-5 page contracts. Transparency matters. - Most cloud providers require commitments for a fixed number of GPUs per year, which most start-ups can't handle. Nebius, however, offers resource-based commitments, where a minimum spend unlocks discounts instead of requiring a set GPU quota. - Nebius provides a level of customization and flexibility that no other cloud provider offers. Long $NBIS.