深潮 TechFlow

wrote a column · ·

NeoCloud Big Three: Which holds greater investment value—NBIS, IREN, or CRWV?

Compiled and edited by TechFlow Deep Tide

Guest: Nate (Portfolio Manager at Endicott Invests)

Host: Steven Fiorillo

Original Title: The Neocloud Discussion with Nate Endicott: NBIS, IREN, CRWV

Podcast Source: steven fiorillo

Broadcast Date: May 24, 2026

Nebius, CoreWeave, and IRON—the three 'NeoCloud' companies—may represent the most certain yet longest-overlooked segment in NVIDIA’s GPU compute supply chain today. This episode features Nate, founder of Endicott Invests, who systematically breaks down the business models, debt structures, customer bases, and team DNA differences among these three firms.

Nate’s core thesis is that the compute shortage will persist for at least three to five years, and hyperscalers physically cannot build data centers fast enough themselves. NeoCloud providers are not substitutes to be replaced but essential complements filling a critical gap. He provides a 2030 revenue target for Nebius of 5 GW × $10/MW = $50 billion (relative to its current market cap), nearly 14 times Wall Street’s consensus estimate of $3.677 billion.

An even more telling signal is that Nebius raised prices on H100 and B200 GPUs by 30%–70% effective June 1, indicating that the bearish narrative—that GPUs rapidly depreciate and compress margins—is being disproven by reality.

‘I’ve never been this bullish before. AI adoption today is like the internet in 1996—big companies are still building “websites,” which seem like toys now, but in 20 years every company will have one. It’s the same with AI; we’re right at the bottom-left of the S-curve.’

‘Everyone wants to know what the maintenance mode looks like. Let me tell you—it looks like Microsoft, Amazon, and Google spending just enough capital to sustain their current scale, while every other company is ramping up infrastructure investment.’

Taiwan Semiconductor wouldn't ramp up capital expenditures to this level just for a one- or two-year boom cycle, and neither would Micron. They have the clearest visibility into the cycle—when they expand capacity, they're signaling to the market that this isn't a short-cycle phenomenon.

CoreWeave was the original OG entrant, starting out mining Ethereum in 2017. After the 2018 crash, it pivoted to supplying compute power to Microsoft and early-stage OpenAI. It now has over 1 GW of activated power capacity and focuses primarily on training, which commands the highest unit prices and generates the highest revenue.

IRON originally started as a Bitcoin miner and shifted to GPUs at the beginning of last year. Its moat lies in land and power—it has secured 4.5 GW of contracted power capacity, the largest among the three—but so far it has only signed Microsoft as a major client, and a second large deal has been conspicuously delayed, which is highly unusual.

Nebius sits between the other two and is the only player betting heavily on its software stack. Arkady has publicly stated four or five times that he aims to build 'the fourth hyperscale cloud provider.' Its recent acquisitions of Egen AI and Clarif AI are all geared toward building Token Factory, its inference platform.

Team is what I value most in an investment. IRON’s founders are two Australian salespeople by background; CoreWeave’s founders have only modest technical expertise; only Nebius boasts a team of engineers who have worked together for 20 years, led by a CEO who is himself a mathematician. This DNA fundamentally shapes the company’s culture and product trajectory.

In its last quarterly earnings call, Nebius mentioned that every GPU it offers has four customers bidding for it. Starting June 1, it raised prices for H100 and B200 GPUs for new customers by 30–70%, directly passing on the cost increases in memory and GPUs to downstream clients—leading to even better gross margins.

Their GPU depreciation cycle is five years. The H100 and B100 have already been in service for nearly five years, yet their prices are still rising—any price increase beyond depreciation translates directly into 100% net profit. That math is simply too compelling.

"Hyperscalers cannot simply replicate the Neo cloud. AWS data centers were designed for low-intensity cloud services—their rack load capacity, cooling systems, and network configurations are all different. It's impossible to shut down cloud services overnight and switch entirely to AI workloads."

"Nebius disclosed last quarter that revenue from customer contracts outside of Microsoft and Meta grew 3.5x quarter-over-quarter. This indicates that mid-sized enterprises—non-hyperscaler clients—are already procuring compute capacity at scale. The market has fixated entirely on the giants and overlooked this second wave of demand."

"There’s a rough rule of thumb for estimating Neocloud revenue: 1 MW equals $10 million in annualized revenue. Nebius has guided toward reaching 5 GW by 2030, implying $50 billion in revenue."

"Nebius also owns four subsidiaries that add incremental valuation: a 25% stake in ClickHouse (valued at approximately $5 billion based on its January $20 billion valuation), an 83% stake in Avride (an autonomous delivery robotics company), and full ownership of Toloka and TripleTen. These represent additional upside within a sum-of-the-parts valuation and also serve as collateral for low-cost debt."

"The most underappreciated signal in NVIDIA’s latest earnings report was the networking segment beating expectations. You can’t deploy racks without firewalls, switches, and wireless access points (WAPs). Arista Networks is a pure-play play on the networking layer, while Cisco derives 42–44% of its revenue from services and security (especially after acquiring Splunk)—so both paths are viable."

Steve: Should we start with explaining 'What is a GPU?'?

Nate: Given the size of my NVIDIA position, I’d hope I at least know what a GPU is. I’m often asked how to choose between CoreWeave and Nebius. My standard answer is that I’ve invested in Nebius because, while I’m uncertain who will ultimately win, I prefer Nebius’s financial structure—particularly its debt profile.

Steve: Full disclosure—I hold Nebius, but not CoreWeave or IRON. I’d like to learn more from you. My initial thesis was quite simple: there isn’t enough compute capacity, and with such large RPOs (remaining performance obligations), AWS, Google Cloud, Azure, and even Oracle will inevitably need overflow solutions. Neo cloud also has significant contract backlogs—not as massive as the hyperscalers’, but already substantial. Then I looked at the balance sheet and concluded Nebius is far stronger than CoreWeave, especially when considering long-term growth potential. Please walk me through each of the three one by one.

Nate: Full disclosure upfront—I have a strong preference for NEBIUS, with over 50% of my portfolio allocated to it. My investment style is highly concentrated; my earlier bet on Palantir followed the same approach. But at the sector level, I believe every player will win because demand and compute bottlenecks are undeniable. Initially, I viewed NEBIUS as a 12–24 month swing trade. Now, after holding it for nearly a year, I’ve upgraded it to a 3–5–10 year long-term holding, as my understanding of the scale of compute demand keeps getting revised upward repeatedly.

Neo Cloud sits between hyperscalers (AWS, GCP, Azure, Oracle) and lab/enterprise customers as a specialized GPU cloud provider. This category exists fundamentally because hyperscalers physically can’t build out power, cooling, networking, and GPU clusters fast enough. CoreWeave is the original player (OG), starting in 2017 by mining Ethereum and pivoting to supplying compute to Microsoft and early OpenAI after the 2018 crash. IRON also originated from Bitcoin mining and shifted to GPUs in early last year. I’ll discuss NEBIUS separately—their engineering depth is astonishing. There are also smaller players like Cipher, WULF, and Galaxy, all with very small market caps; these three—CoreWeave, IRON, and NEBIUS—are the main ones.

I especially want to highlight NEBIUS because I deeply value its team. CEO Arkady is a mathematician. In 1993, he founded Yandex in Russia—before Google even existed. In 2003, Google offered $130 million to acquire Yandex, but he declined; later that same year, Yandex went public at an $8 billion valuation, and its market cap peaked at $30 billion in 2022. This track record shows Arkady is a serious, visionary leader and a world-class mathematician. Between 2003 and 2022, Yandex grew into Russia’s Google, expanding into search, autonomous driving, and numerous other businesses, and even built its own data center in Finland, which began operations around 2012–2013. When the Russia-Ukraine war broke out in 2022, Arkady and the entire company publicly opposed it, decided to sell their Russian assets, exit Russia, and relist on NASDAQ. The entire company and engineering team relocated to the Netherlands. That’s why NEBIUS’s stock chart shows a two- to three-year gap—it was delisted during that period. Investing in NEBIUS means backing a mathematician who said 'no' to Google, 'no' to his own country, and moved his entire operation to another nation just to relist. All engineers relocated with their families—a culture and sense of mission I also valued when investing in Palantir.

Steve: I had no idea about this history. I primarily focus on GPU exposure, financial statements, balance sheets, forward EPS, and revenue growth. CoreWeave looks interesting, but I dislike its heavy debt load. As my view on the AI adoption timeline keeps shifting further out, I’m now considering investing in all three.

Nate: I shared this background first because most people aren’t aware of it. They raised roughly $5.5–6 billion from selling their Russian assets and used that capital to build new data centers by the end of 2024.

Steve: What are the key differences among the three?

Nate: Three dimensions. CoreWeave entered earliest, has the highest activated power capacity, and leads in revenue—it primarily focuses on training workloads, handling tasks like OpenAI’s GPT-5 training. IRON sits between training and inference. NEBIUS specializes in inference. Inference offers better sustainability, while training commands higher pricing—hence CoreWeave enjoys a premium.

In terms of power scale, CoreWeave has already activated over 1 GW; IRON has locked in 4.5 GW, giving it the strongest moat in land and power access—but strangely, it has only signed Microsoft as a major client, and a second large contract has been slow to materialize. All three have Microsoft contracts. CoreWeave also has OpenAI, NEBIUS has Meta, and IRON currently only has Microsoft. For IRON shareholders, the next major contract is the critical catalyst—the longer it’s delayed, the more concerning it becomes.

Differentiation-wise: CoreWeave competes on scale and ClusterMAX Platinum certification; IRON leverages cheap renewable energy and vertical integration (owning land and power infrastructure); NEBIUS differentiates through its software stack—its Token Factory inference platform is central to its strategy. Recent acquisitions of Egen AI and Clarif AI, both focused on token and model optimization, aim to position NEBIUS as a hyperscaler. Arkady has repeatedly emphasized in recent earnings calls the ambition to become the 'fourth hyperscaler'—or fifth if you count Oracle.

Last quarter, NEBIUS disclosed that revenue from contracts outside Microsoft and Meta grew 3.5x quarter-over-quarter—a very bullish signal. The reason is that as you move further downstream, clients increasingly require software services, which carry higher gross margins. Contracts with Microsoft and Meta are bare-metal deals—hyperscalers use their own container orchestration systems and software, purchasing only raw compute at low margins. Mid-sized downstream enterprises need full software stacks, which significantly boosts NEBIUS’s profit elasticity.

NEBIUS also has four subsidiaries that contribute to its valuation. ClickHouse is a data analytics and labeling platform, comparable to Databricks, which raised funds in January at a $20 billion valuation. NEBIUS holds a 25% stake, equivalent to roughly $4.2 billion. In private markets, this stake could double or even triple over the next three years, potentially valuing NEBIUS’s holding at $10–15 billion (assuming no dilution; actual dilution may reduce its stake to 15–18%). Avride develops autonomous delivery robots, deployed on campuses such as Columbus, and partners with Uber. NEBIUS owns 83% of Avride, which is valued between $5–8 billion. TripleTen (an education platform) and Toloka (data labeling) are smaller, operating at the multi-million-dollar scale, contributing modestly to valuation but adding optionality. Therefore, NEBIUS’s total valuation should be viewed as the sum of its GPU cloud business plus the sum-of-the-parts valuation of its subsidiaries. CoreWeave only offers top-tier compute capacity; IRON’s compute operations are in the earliest stage but have secured the largest land and power capacity. IRON’s AI cloud revenue is only $20–30 million. NEBIUS sits in the middle—its revenue exceeds IRON’s, but it lags in power availability and must keep building its own infrastructure. Permitting approvals in New Jersey and Pennsylvania are major hurdles, and capital expenditures remain high.

Steve: We just received all the earnings reports, including NVIDIA’s from two days ago, along with hyperscaler results (excluding Oracle). Based on Jensen Huang’s comments, NVIDIA’s data, memory (Micron, SanDisk), and Taiwan Semiconductor’s situation, how do you interpret these developments for Neo Cloud?

Nate: I’ve never been more bullish than I am now. Our use of AI is still extremely early-stage. Every company will have an AI presence in 20 years, just like every company has a website today. We’re currently in a phase equivalent to the internet in 1996—early adopters. Add robotics and autonomous driving into the mix, and we’re at the very bottom-left of the S-curve. I believe Neo Cloud will experience explosive growth over the next three years. If you have GPUs, you can grow tenfold because customers simply need compute power. Only after five years will software stacks and service quality become decisive. NEBIUS’s current investments are laying the groundwork for that phase.

As for whether rising memory and GPU prices will squeeze NEBIUS’s gross margins, NEBIUS announced yesterday that H100 and B200 pricing for new customers will increase by 30–70% effective June 1, fully passing through cost increases. NEBIUS’s medium-term EBIT margin guidance is 15–20%, with an EBITDA margin target of 40% by year-end. While I don’t focus much on EBITDA, EBIT reflects operating profitability and is a useful metric.

Steve: All signs point to compute demand—both CPUs and GPUs are in short supply. The clearest visibility lies with Taiwan Semiconductor and Micron: servers need memory, and chips need manufacturing. Taiwan Semiconductor’s largest customer is NVIDIA, and NVIDIA’s biggest customers are Microsoft, Amazon, Alphabet, and Oracle. Taiwan Semiconductor’s significant increase in capital expenditure this year isn’t geared toward a one- or two-year cycle—and neither is Micron’s. If they’re expanding capacity, it means my earlier assumptions were wrong. I’ve already revised my outlook on the AI cycle twice and am about to make a third revision. When ChatGPT launched, I estimated NVIDIA’s cycle would last 18–24 months; a year and a half ago, I extended it to 24–36 months. Now I’m asking myself: will this slow down before 2030? I’m not sure. Alphabet said it will spend $190 billion in capex this year and plans to spend even more next year. There won’t be any letup through 2028.

Nate: I agree with your assessment. Initially, I viewed NEBIUS as a trading opportunity, but as my understanding of the market deepened—and seeing the capital expenditure growth trajectory and shifts in how people talk about AI—I’ve changed my perspective. Right now, everyone is focused on hyperscalers, but the next phase involves the entire Fortune 500 and S&P 2000. Every mid-sized company will soon say, ‘I need compute power; I need agent training.’ That’s when true early adoption begins. NEBIUS is already partnering with Shopify, Cloudflare, and CrowdStrike. The next wave is ‘every company needs compute,’ analogous to ‘every company needs a website.’ Executives and VPs are now thinking: let the giants spend hundreds of billions to figure out what works, then we’ll follow. NEBIUS’s 3.5x faster growth in non-hyperscaler contracts is exactly this signal.

Steve: I used to model GPU cycles like RAM cycles, but now I believe the GPU trend won’t stop—nor will memory demand or accelerating capital spending. Giants won’t keep doubling their spending or growing 50–60% annually, but another 10–20% annual growth is entirely plausible. Overlaying NVIDIA’s data with token economics, I have to believe CoreWeave, NEBIUS, and IRON will see substantial revenue expansion ahead.

Nate: There’s a simple rule of thumb for compute monetization: 1 MW equals $10 million in annualized revenue, so 100 MW equals $1 billion. Bare metal is slightly cheaper, other offerings cost more, but this is the baseline. Thus, NEBIUS’s 2030 target of 5 GW implies $50 billion in revenue. Even the analyst consensus estimate of $3.677 billion represents a 100x increase from last year’s roughly $500 million in revenue. Next year, the market consensus expects 550% year-over-year growth, followed by 219% the year after—that’s the consensus among 17 analysts. CoreWeave shows similarly explosive figures: 147%, 97%, and 58% growth rates, reaching into the tens of billions in absolute revenue.

Regarding the bearish narrative around GPU depreciation, NEBIUS’s last quarterly report explicitly stated that every GPU has four customers competing for it.

Steve: I have an original theory. At a recent conference, Jensen mentioned many customers are still using H100s from five or six years ago. I suspect that once hyperscalers receive the next-generation Vera Rubin chips, they’ll lease out their older GPUs instead of discarding them. NEBIUS and CoreWeave will do the same. Ultimately, this compute capacity will flow to non-tech companies like Coca-Cola and Walmart.

Nate: Good point. NEBIUS has a five-year depreciation cycle. The H100/B100 GPUs are already approaching five years, yet prices are still rising—any price increase beyond depreciation translates directly into 100% net profit.

Steve: What’s the most misunderstood aspect of this sector?

Nate: ‘Hyperscalers can easily replicate Neo Cloud.’ AWS was designed in the early 2000s for cloud services, with far lower requirements for compute intensity, cooling, and network configuration compared to AI workloads. AWS can’t just shut down its existing cloud infrastructure overnight and switch to AI workloads. Racks need to be reconfigured to handle significantly higher heat output and power demands. NEBIUS, CoreWeave, and IRON have purpose-built data centers from the ground up for next-generation AI workloads. Hyperscalers have experience and capital, but they can’t simply copy this model—that’s the biggest misconception. The second is Michael Burry’s claim that ‘GPUs are about to become obsolete,’ which has already been disproven by both market action and reality—NEBIUS raised prices by 30% yesterday. Sure, big tech will eventually enter the competition, but that won’t matter much over the next three years. Beyond five years, software and services will be key—that’s why I’m bullish on NEBIUS’s software stack strategy. Over the next three years, you could practically close your eyes and pick any Neo Cloud provider and still outperform the market.

Steve: Among the three, which is most likely to win? After all, their customer bases overlap.

Nate: Looking five years out, it’s NEBIUS. IRON’s founders are two Australian salespeople; CoreWeave’s founders have only modest technical backgrounds; only NEBIUS has a team of engineers who’ve worked together for 20 years, led by a mathematician CEO—their culture and product direction are shaped by this DNA. However, the more probable outcome resembles the current hyperscaler landscape: a three-way split like AWS/GCP/Azure, perhaps 40/40/20 or 35/35/30. In the short term—two years—CoreWeave leads due to its Platinum certification and an RPO approaching $100 billion. NEBIUS is around $40 billion; I don’t know IRON’s exact figure. Another structural difference: IRON operates entirely with owned-and-operated data centers but has weak software; CoreWeave uses a colocation model where multiple companies share the same facility, which pressures gross margins; NEBIUS sits between the two—combining owned-and-operated with colocation. NEBIUS has guided toward 75% owned-and-operated by 2030, gradually eroding IRON’s moat while preserving profitability. IRON currently holds advantages in land and power access, but over a 5–10 year horizon, NEBIUS is the clear leader.

Steve: Does only one company have to win, or can all three grow revenue and EPS?

Nate: All three will grow. I’m bullish on them through at least 2030. Beyond 2030, the opportunity may be even larger—once the world fully understands how to use AI, that’s when the real race begins. All three will win, and even smaller players like Cipher, WULF, and Galaxy will succeed too.

Steve: Final question: beyond compute capacity, where’s the next wave of opportunity?

Nate: NVIDIA’s earnings report showed networking revenue massively beating expectations—that’s a signal. My top conviction holding is Arista Networks.

Steve: That breaks my heart—I’m a Cisco shareholder.

Nate: I like Arista's vendor neutrality. Network infrastructure has a longer capital expenditure conversion lag compared to Neo Cloud and GPU companies, but networking significantly beat expectations in the earnings report, while GPUs only met or slightly exceeded expectations.

Steve: Networking is already a major contributor to NVIDIA's revenue—you're absolutely right. You can't stand up a rack without firewalls, switches, and wireless access points (WAPs). I'm more familiar with Cisco than Arista because I've negotiated with them directly. Cisco has transformed from a hardware company—42% to 44% of its quarterly revenue now comes from services and security. It has also acquired companies like Splunk, and its security product Cisco Umbrella is very strong. Its hardware is also bundled with subscriptions. Both companies will rise. Non-technical people don’t realize how big the switch business is. Nobody runs Ethernet cables at home anymore, but offices absolutely need switches because signals degrade beyond 300 feet, so enterprises require a main data center plus multiple intermediate distribution frames (IDFs). Data centers need even more rack-end Ethernet cabling. The opportunity in networking is bigger than most imagine.

Nate: Arista Networks has also doubled in price. A safer bet is 'the next Micron, the next memory play, the next Neocloud.' If you believe in AI infrastructure, NEBIUS, Micron, Arista, Cisco, NVIDIA, and ASML will all win. Personally, I’m heavily invested in Arista at the network layer.

258K Views