NeoCloud’s Big Three: Which Offers Greater Investment Value—NBIS, IREN, or CRWV?
TechFlow Selected TechFlow Selected
NeoCloud’s Big Three: Which Offers Greater Investment Value—NBIS, IREN, or CRWV?
Computing power shortages will persist for three to five years. Nebius, CoreWeave, and IRON—the three NeoCloud providers—each possess distinct competitive moats and will all benefit. Among them, Nebius stands out for its engineering-driven DNA and software-stack strategy, giving it the strongest long-term potential; its revenue could reach $50 billion by 2030—nearly 14 times the Wall Street consensus.
Navigating Web3 tides with focused insightsCompiled & Translated by TechFlow
Guest: Nate (Founder, Endicott Invests)
Host: Steven Fiorillo
Original Title: The Neocloud Discussion with Nate Endicott: NBIS, IREN, CRWV
Podcast Source: steven fiorillo
Air Date: May 24, 2026
Editor’s Introduction
Nebius, CoreWeave, and IRON—these three “NeoCloud” companies may represent the most certain yet longest-overlooked link in today’s NVIDIA GPU compute supply chain. In this episode, we welcome Nate, founder of Endicott Invests, who systematically dissects the business models, debt structures, customer compositions, and team DNA differences among the three firms.
Nate’s core thesis is that the compute shortage will persist for at least three to five years. Hyperscalers physically cannot build out data centers fast enough on their own—and NeoClouds are not substitutes but essential complements. He projects Nebius’ 2030 revenue at 5 GW × $10/MW = $50 billion (implied by its current market cap), nearly 14x Wall Street’s consensus estimate of $3.677 billion.
A more critical signal: On June 1, Nebius raised prices for H100/B200 GPUs by 30–70%, directly refuting the bearish narrative that GPUs rapidly depreciate and compress margins.
Key Quotes
Market Size & Cycle Outlook
- “I’ve never been this bullish before. AI’s current penetration level is equivalent to the internet in 1996—large corporations are still building basic ‘websites’ as toys; twenty years later, every company had a website. AI is no different—we’re still at the far lower-left corner of the S-curve.”
- “Everyone wants to know what maintenance mode looks like. Let me tell you: it looks like Microsoft, Amazon, and Google holding capital expenditures flat just to maintain existing scale—while every other company ramps up infrastructure spending.”
- “TSMC won’t ramp capex to this magnitude for a one- or two-year boom cycle—and neither will Micron. They have the highest visibility into the cycle; their capacity expansion signals to the market that this is not a short-term cycle.”
Differentiation Among the Three NeoClouds
- “CoreWeave is the OG—the earliest entrant. It started mining Ethereum in 2017, pivoted post-2018 crash to supplying compute to Microsoft and early OpenAI, and now has over 1 GW of activated power. It focuses on training, commanding the highest unit pricing and revenue.”
- “IRON originated as a Bitcoin miner and transitioned to GPU compute last year. Its moat lies in land and power—it has secured 4.5 GW of contracted power, the largest among the three—but has signed only one major customer (Microsoft) so far. A second major contract remains conspicuously absent.”
- “Nebius sits between the two, uniquely betting on software stack. Arkady has publicly stated four or five times that Nebius aims to become the fourth hyperscaler. Recent acquisitions—including Egen AI and Clarif AI—are all aimed at building Token Factory, its inference platform.”
Nebius’ Team & History
- “Team is my top investment criterion. IRON’s founders are two Australian salespeople; CoreWeave’s founder has only modest technical credentials; Nebius alone fields an engineering team that has worked together for 20 years, led by a CEO who is a mathematician. This DNA shapes corporate culture and product trajectory.”
Moats, Pricing Power & Margins
- “In Nebius’ last earnings report, they disclosed that each GPU is contested by four customers. Starting June 1, H100 and B200 pricing for new customers rose 30–70%, passing memory and GPU cost increases straight through to downstream clients—boosting gross margins instead of compressing them.”
- “Their GPU depreciation cycle is five years. H100/B100 units are nearing five years of service—but prices are still rising. Any price increase post-depreciation flows straight to 100% net profit—a mathematically irresistible equation.”
- “Hyperscalers cannot simply replicate NeoClouds. AWS data centers were designed for low-intensity cloud services: rack weight limits, cooling, and network configurations differ entirely. You can’t shut down cloud services overnight and switch to AI workloads.”
Future Growth Sources & Methodology
- “Last quarter, Nebius disclosed that contract revenue from non-hyperscaler clients—excluding Microsoft and Meta—grew 3.5x quarter-on-quarter. This signals that mid-market enterprises—not just giants—are beginning large-scale compute procurement. Markets focus exclusively on hyperscalers, overlooking this second wave of demand.”
- “A rough rule of thumb for estimating NeoCloud revenue: 1 MW = $10 million in annualized revenue. Nebius’ 2030 guidance of 5 GW implies $50 billion in revenue.”
- “Nebius also holds four subsidiaries that add valuation upside: 25% stake in ClickHouse (valued at ~$5 billion based on its $20 billion January valuation), 83% in Avride (autonomous delivery robots), plus Toloka and TripleTen. These constitute additional ‘sum-of-the-parts’ valuation levers—and serve as collateral for low-cost debt.”
Secondary Opportunity: The Network Layer
- “The most underappreciated signal in NVIDIA’s latest earnings was its networking business’ outperformance. If racks won’t stand up, you still need firewalls, switches, and wireless access points (WAPs). Arista Networks is the pure-play network-layer play; Cisco derives 42–44% of revenue from services and security (post-Splunk acquisition)—both paths are viable.”
Steve: Should we start with “What is a GPU?”
Nate: Given the size of my NVIDIA position, I should at least know what a GPU is. I’m often asked how to choose between CoreWeave and Nebius. My standard answer: I invested in Nebius—not because I know who’ll win, but because I prefer its financial structure, especially its debt profile.
Steve: Full disclosure—I hold Nebius, but no CoreWeave or IRON. I’d love to absorb more from you. My original logic was simple: compute shortages + massive RPO (Remaining Performance Obligations) mean AWS, Google Cloud, Azure, and Oracle must offload demand. NeoClouds also carry large contract backlogs—smaller than the giants’, but already substantial. Then I looked at balance sheets: Nebius looks far stronger than CoreWeave, and then at long-term growth. Please walk us through all three.
Nate: Full disclosure: I hold a strong preference for Nebius—over 50% of my portfolio. My investing style is highly concentrated; I adopted the same approach when backing Palantir. But at the sector level, I believe all three will win—demand and compute bottlenecks are real and structural. I initially viewed Nebius as a 12–24 month trade. Now approaching one year in, it’s evolved into a 3–5–10 year holding—as my understanding of the scale of compute demand keeps getting revised upward.
NeoClouds sit between hyperscalers (AWS, GCP, Azure, Oracle) and lab/enterprise customers as specialized GPU cloud providers. This category exists because hyperscalers physically cannot build out power, cooling, networking, and GPU clusters fast enough. CoreWeave is the OG—launched in 2017 mining Ethereum, pivoted post-2018 crash to supply compute to Microsoft and early OpenAI. IRON, too, began as a Bitcoin miner and shifted to GPU compute last year. I’ll cover Nebius separately—its engineering depth will surprise you. Other smaller players include Cipher, WULF, and Galaxy—all tiny by market cap; these three dominate.
I emphasize Nebius because I value its team. CEO Arkady is a mathematician. In 1993, he founded Yandex in Russia—before Google existed. In 2003, Google offered $130 million to acquire it; he declined. That same year, Yandex IPO’d at an $8 billion valuation—and peaked at $30 billion in 2022. This history proves Arkady is both principled and visionary—and a world-class mathematician. Between 2003 and 2022, Yandex grew into Russia’s Google—running search, autonomous driving, and diverse businesses—and built its own data center in Finland, operational around 2012–2013. When the Russia-Ukraine war erupted in 2022, Arkady and the entire company publicly opposed it, decided to sell Russian assets and relocate from Russia—and relisted on NASDAQ. The entire company and engineering team moved to the Netherlands. So Nebius’ stock chart shows a multi-year gap—that’s its delisting period. Investing in Nebius means backing a mathematician who said “no” to Google, “no” to his home country, and relocated his company to another nation to list. All engineers moved with their families—this culture and mission-driven ethos mirrors what I valued when investing in Palantir.
Steve: I had no idea about this history. I mainly look at GPU trends, financials, balance sheets, forward EPS, and revenue growth. CoreWeave looks intriguing—but I hate its stacked debt. As my view on AI’s adoption timeline keeps pushing further out, I’m now considering positions in all three.
Nate: I lead with background because most people don’t know it. They sold Russian assets for roughly $5.5–6 billion, and used those proceeds to build new data centers by end-2024.
Steve: What’s the main difference among the three?
Nate: Three dimensions. CoreWeave entered earliest, has the highest activated power, and generates the highest revenue—focused on training (e.g., OpenAI’s GPT-5 training runs on CoreWeave). IRON sits between training and inference. Nebius focuses on inference—more sustainable, though lower-priced than training; CoreWeave thus commands premium pricing.
In power scale: CoreWeave has activated >1 GW; IRON has locked in 4.5 GW—its biggest moat is land and power—but curiously has signed only Microsoft as a major client. A second major contract remains delayed—an oddity. All three have Microsoft contracts. CoreWeave also serves OpenAI; Nebius serves Meta; IRON currently serves only Microsoft. For IRON holders, the next major contract is the key catalyst—the longer it drags, the stranger it becomes.
Differentiation-wise: CoreWeave bets on scale + ClusterMAX Platinum certification; IRON on low-cost renewable energy + vertical integration (own land and power); Nebius on software stack—its Token Factory inference platform is central. Recent acquisitions of token-optimization and model-optimization firms like Egen AI and Clarif AI all aim to push toward hyperscaler status. Arkady has repeatedly emphasized on earnings calls that Nebius intends to be the “fourth hyperscaler”—the fifth, if counting Oracle.
Last quarter, Nebius disclosed that contract revenue from non-Microsoft/non-Meta clients grew 3.5x sequentially—a strongly bullish signal. Why? The further downstream you go, the more customers need software services—and higher margins. Contracts with Microsoft and Meta are bare-metal: hyperscalers bring their own container orchestration and scheduling software, buying only raw compute—low-margin. Mid-market enterprises need full software stacks—driving significant margin elasticity for Nebius.
Nebius also adds valuation via four subsidiaries. ClickHouse is a data analytics and labeling platform—comparable to Databricks—valued at $20 billion in its January financing round; Nebius holds 25%, worth ~$4.2 billion. In private markets, its value could double or triple over the next three years—Nebius’ stake could reach $10–15 billion (assuming minimal dilution; actual dilution may reduce its stake to 15–18%). Avride builds autonomous delivery robots, deployed on campuses like Columbus and partnered with Uber; Nebius owns 83%, with a valuation range of $5–8 billion. TripleTen (education platform) and Toloka (data labeling) are smaller—valued in the low millions—contributing little to valuation but adding optionality. Thus, Nebius’ valuation must reflect both its GPU cloud business and the sum-of-the-parts valuation of its subsidiaries. CoreWeave has only top-tier compute; IRON is earliest-stage in compute but holds the largest land-and-power commitments—its AI cloud revenue stands at only $20–30 million; Nebius sits in the middle—revenue exceeds IRON’s but power capacity lags, requiring continuous self-builds. Permit approvals in New Jersey and Pennsylvania pose major hurdles, and capex is high.
Steve: We just received all earnings reports—including NVIDIA’s two days ago, plus hyperscaler reports (excluding Oracle). Combined with Jensen Huang’s commentary, NVIDIA data, memory (Micron, SanDisk), and TSMC updates—how do you interpret these for NeoClouds?
Nate: I’ve never been this bullish before. Our use of AI remains extremely early. Every company will have a website in 20 years—and AI will follow the same path. We’re at the far lower-left corner of the S-curve—equivalent to the internet in 1996. I believe NeoClouds will explode over the next three years: owning GPUs guarantees 10x growth, because customers simply need compute. Only after five years will software stacks and service quality determine winners—and Nebius’ investments here prepare it for that phase.
As for whether memory and GPU price hikes will squeeze Nebius’ margins—Nebius announced yesterday that H100 and B200 pricing for new customers rises 30–70% effective June 1, fully passing through costs. Nebius’ medium-term EBIT margin guidance is 15–20%; year-end EBITDA margin guidance is 40%. I pay less attention to EBITDA, but EBIT reflects operating margin—this is a useful benchmark.
Steve: All signals point to compute—neither CPUs nor GPUs are sufficient. Highest-visibility indicators are TSMC and Micron: servers need memory; chips need fabrication. TSMC’s largest customer is NVIDIA; NVIDIA’s largest customers are Microsoft, Amazon, Alphabet, and Oracle. TSMC’s sharp capex expansion isn’t for a one- or two-year cycle—and neither is Micron’s. If they’re expanding capacity, my prior cycle assumptions were wrong. I’ve already revised my AI-cycle target twice—and am about to revise it a third time. When ChatGPT launched, I projected NVIDIA’s cycle at 18–24 months; a year and a half ago, I extended it to 24–36 months; now I’m asking myself: will this slow before 2030? I’m uncertain. Alphabet says it will spend $190 billion on capex this year—and even more next year. No slowdown until 2028.
Nate: I agree with your assessment. I first viewed Nebius as a tactical trade—but deeper market understanding, combined with observing capex growth curves and shifting dialogue, has transformed it into a long-term holding. Everyone talks about hyperscalers now—but next comes the entire Fortune 500 and S&P 2000: every mid-sized company will say, “We need compute; we need agent training.” That’s true early-stage adoption. Nebius already partners with Shopify, Cloudflare, and CrowdStrike. The next wave is “every company needs compute”—equivalent to “every company needs a website.” Executives and VPs think: let the giants spend hundreds of billions first to figure out what works—then we’ll follow. Nebius’ 3.5x sequential growth in non-hyperscaler contracts is precisely that signal.
Steve: I used to model GPUs using RAM-cycle logic—but now I see GPU demand won’t stop, memory won’t stop, and capex acceleration won’t stop. Giants won’t double or grow 50–60% annually forever—but 10–20% annual growth remains very plausible. Overlay NVIDIA’s data with token economics, and I must believe CoreWeave, Nebius, and IRON revenues will expand dramatically.
Nate: A simple compute-to-revenue conversion: 1 MW = $10 million in annualized revenue; 100 MW = $1 billion. Bare-metal is cheaper; other offerings are pricier—but this is the baseline. So Nebius’ 2030 guidance of 5 GW implies $50 billion in revenue. Even analysts’ consensus of $3.677 billion represents ~100x growth from last year’s ~$500 million. Next year’s consensus growth is +550%; the year after, +219%—based on 17 analysts’ estimates. CoreWeave is similarly explosive: +147%, +97%, +58%—with absolute revenue in the tens of billions.
Regarding the bearish narrative on GPU depreciation—Nebius explicitly stated in last quarter’s earnings that each GPU is contested by four customers.
Steve: I have an original theory. At a recent conference, Jensen noted many customers still run H100s from five or six years ago. I suspect that once hyperscalers receive the next-gen Vera Rubin GPUs, they’ll lease out older ones—not scrap them. Nebius and CoreWeave will do the same. Ultimately, this compute flows to non-tech firms like Coca-Cola and Walmart.
Nate: Excellent point. Nebius’ GPU depreciation cycle is five years. H100/B100 units are nearing five years—but prices keep rising. Any price increase post-depreciation is 100% net profit.
Steve: What’s the most misunderstood aspect of this sector?
Nate: “Hyperscalers can replicate NeoClouds.” AWS was architected in the early 2000s for cloud services—far less demanding on compute intensity, cooling, and network configuration than AI workloads. AWS can’t shut down cloud services overnight and switch to AI workloads. Racks must be reconfigured to handle higher heat and load. Nebius, CoreWeave, and IRON built their data centers from scratch for next-gen AI workloads. Hyperscalers have experience and capital—but cannot simply copy-paste. That’s the #1 misconception. The #2 is Michael Burry’s claim that “GPUs will soon be obsolete”—already refuted by markets and reality: Nebius raised prices 30% yesterday. Hyperscalers will certainly compete—but that’s irrelevant for the next three years. After five years, software and services will decide winners—so I favor Nebius’ software-stack positioning. For the next three years, buying any NeoCloud blindly beats the market.
Steve: Among the three, who’s most likely to win—given overlapping customers?
Nate: Long term (five years), Nebius. IRON’s founders are two Australian salespeople; CoreWeave’s founder has only modest technical credentials; only Nebius fields a 20-year co-working engineering team led by a mathematician—this DNA dictates culture and product direction. But the likelier outcome resembles AWS/GCP/Azure’s triopoly: 40/40/20 or 35/35/30. Short term (two years), CoreWeave leads—thanks to Platinum certification and RPO near $100 billion. Nebius’ RPO is ~$40 billion; IRON’s number is unclear. One structural distinction: IRON operates fully owned, self-run data centers—but with weak software; CoreWeave uses colocation—multiple firms share one facility, compressing margins; Nebius straddles both—owning and operating some facilities while collocating others. Nebius guides 75% owned-and-operated by 2030—gradually eroding IRON’s moat while preserving margins. IRON holds near-term land-and-power advantages—but over five to ten years, Nebius wins.
Steve: Must only one win—or can all three grow revenue and EPS?
Nate: All three grow. All are bullish through 2030—and possibly beyond. Post-2030 may be even more bullish, once the world grasps how to apply AI—that’s the real starting line. All three will win—even smaller players like Cipher, WULF, and Galaxy.
Steve: Final question: Beyond compute, where’s the next wave of opportunity?
Nate: NVIDIA’s earnings delivered a knockout surprise in networking—that’s the signal. My largest position is Arista Networks.
Steve: Heartbroken—I’m a Cisco shareholder.
Nate: I like Arista’s vendor-agnostic stance. Networking capex has longer conversion lag than NeoCloud or GPU firms—but its earnings beat was unequivocal. GPU results barely met or slightly exceeded expectations.
Steve: Networking already dominates NVIDIA’s revenue—your point stands. Without racks standing, you still need firewalls, switches, and wireless access points (WAPs). I know Cisco better than Arista—I’ve negotiated with them. Cisco has transformed from hardware to services and security: 42–44% of quarterly revenue now comes from those segments. Its Splunk acquisition strengthens security offerings—Umbrella is particularly powerful—and hardware is subscription-bundled. Both stocks will rise. Non-technical people underestimate how big the switch business is. Few homes wire Ethernet anymore—but offices require switches: signal degrades beyond 300 feet, so enterprises need a main data center plus multiple IDF wiring closets. Data centers require more rack-terminated Ethernet cables. The networking opportunity is larger than imagined.
Nate: Arista Networks is also up 100%. A safer answer is “the next Micron, the next Memory, the next NeoCloud.” If you believe in AI infrastructure, Nebius, Micron, Arista, Cisco, NVIDIA, and ASML will all win. Personally, I overweight Arista in the network layer.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
深潮TechFlow
