NeoCloud's three giants: NBIS, IREN, CRWV, which one has more investment potential?

Organized & Compiled by Deep Tide TechFlow

Guest: Nate (Founder of 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

Editorial Introduction

Nebius, CoreWeave, IRON—these three "NeoCloud" companies are arguably the most certain yet overlooked links in the current NVIDIA GPU computing chain. This episode features Nate, founder of Endicott Invests, who systematically dissects the business models, debt structures, customer bases, and team genetics of these three companies.

Nate’s core judgment is that computing power shortages will persist for at least 3 to 5 years. Large-scale cloud providers physically cannot build data centers fast enough. NeoCloud is not a substitute but a necessary complement. He projects Nebius’s 2030 revenue guidance at 5 GW × $10/MW = $3.68B (matching current market cap), nearly 14 times Wall Street’s consensus of $5B.

A more critical signal is that Nebius raised prices on H100/B200 GPUs by 30%–70% starting June 1, indicating that the narrative of GPU devaluation and margin compression is being overturned by reality.

Key Quotes

Track Space and Cycle Judgment

“I’ve never been so optimistic before. AI’s current penetration rate is comparable to the internet in 1996, when big companies were still playing with small toys like web pages. In 20 years, every company will have a website. AI is the same—we’re just at the very beginning of the S-curve.”

“Everyone wants to know what the maintenance model looks like. I tell them, it’s basically Microsoft, Amazon, and Google capital supporting their existing scale, while all other companies are increasing spending on infrastructure.”

“TSMC won’t ramp up capital expenditure for a one- or two-year boom, Micron won’t either. They have the highest visibility on cycles. Their expansion signals to the market that this isn’t a short-term cycle.”

Differentiation Among the Three NeoClouds

“CoreWeave is the earliest OG, starting in 2017 with Ethereum mining, then shifting in 2018 after the crash to providing compute power for Microsoft and early OpenAI. Currently, it has over 1 GW of active power, focusing on training, so it commands the highest prices and revenue.”

“IRON was originally a Bitcoin miner, transitioning to GPU last year. Its moat is land and electricity. It has secured 4.5 GW of contracted power, the largest among the three, but it’s only signed one major client—Microsoft—and the second big deal has been delayed, which is unusual.”

“Nebius falls between the two, the only player betting on the software stack. Arkady has publicly mentioned four or five times that they aim to be the fourth large-scale cloud provider. Recently, they acquired Egen AI and Clarif AI, both for building a token factory and inference platform.”

Nebius Team and History

“I value the team most in my investments. The founders of IRON are two Australians with sales backgrounds. CoreWeave’s founders have an average technical background. Only Nebius has a team of engineers who have worked together for 20 years, and the CEO is a mathematician. This gene shapes the company culture and product direction.”

Moats, Pricing Power, and Profit Margins

“Nebius’s last quarterly report mentioned that each GPU has four competing clients. Starting June 1, H100 and B200 prices increased by 30–70%, directly passing memory and GPU cost increases downstream, which should improve gross margins.”

“Their GPU depreciation cycle is five years. H100/B100 has been in service for nearly five years, yet prices are still rising. After depreciation, any price increase yields 100% net profit—an extremely profitable model.”

“Hyperscalers cannot simply copy NeoCloud. AWS data centers are designed for low-intensity cloud services, with different rack load, cooling, and network configurations. They can’t just shut down cloud services overnight and switch to AI workloads.”

Future Growth Sources and Methodology

“Nebius disclosed last quarter that contract revenue from clients other than Microsoft and Meta grew 3.5 times quarter-over-quarter. This indicates that mid-sized enterprises outside the hyperscaler giants are beginning to purchase compute power at scale. The market’s focus remains on the giants, overlooking the second wave of demand.”

“A rough formula to estimate NeoCloud revenue: 1 MW = $10 million annualized revenue. Nebius’s guidance of 5 GW by 2030 corresponds to $50 billion in revenue.”

“Nebius also has four subsidiaries adding valuation increments: holding 25% of ClickHouse (valued at about $50 billion based on a $20 billion valuation in January), 83% of Avride (autonomous delivery robots), and stakes in TripleTen and Toloka. These are additional valuation chips and collateral for cheap debt.”

Secondary Opportunity: Network Layer

“NVIDIA’s network business is the most underestimated signal in this earnings report. Without racks, firewalls, switches, and wireless access points (WAP), data centers can’t operate. Arista Networks is a pure network layer play, while Cisco derives 42–44% of revenue from services and security (after acquiring Splunk). Both paths are viable.”

Steve: Should we start from “What is a GPU”?

Nate: I hope that with my NVIDIA position size, I should at least understand what a GPU is. I’m often asked how to choose between CoreWeave and Nebius. My standard answer is I invested in Nebius because I’m less certain who will win, but I prefer Nebius’s financial structure, especially its debt profile.

Steve: I’ll disclose I hold Nebius, not CoreWeave or IRON. I want to learn more from you. My initial logic was simple: compute power is insufficient, with large remaining performance obligations (RPO). AWS, Google Cloud, Azure, and Oracle will need spillover capacity; NeoCloud also has large backlog contracts, though not as big as the giants. Looking at the balance sheet, Nebius looks much better than CoreWeave, and then I consider long-term growth. Please analyze each.

Nate: First, I prefer Nebius, with over 50% position. My investment style is highly concentrated; I also bet on Palantir that way. But overall, I believe each company will win because demand and compute bottlenecks are evident. I initially saw Nebius as a 12–24 month trade, but now, after nearly a year, I see it as a 3–5–10 year long-term hold, as my understanding of compute demand has been repeatedly refreshed.

NeoCloud is a professional GPU cloud provider positioned between Hyperscalers (AWS, GCP, Azure, Oracle) and enterprise/lab clients. The core reason for its existence is that Hyperscalers physically cannot build power, cooling, network, and GPU clusters fast enough. CoreWeave is the OG, starting in 2017 with Ethereum mining, shifting in 2018 to providing compute for Microsoft and early OpenAI. IRON also started as a Bitcoin miner, transitioning to GPU last year. Nebius I will discuss separately; their engineering depth is impressive. There are also smaller players like Cipher, WULF, Galaxy, but their market caps are small; these three are the main players.

I especially want to highlight Nebius because I value the team. CEO Arkady is a mathematician. In 1993, he did Yandex in Russia—earlier than Google. In 2003, Google offered $130 million to acquire him, but he refused. The same year, he IPO’d at an $8 billion valuation. At its peak in 2022, Yandex was worth about $30 billion. This history shows Arkady is a serious, visionary leader and a top mathematician. Yandex grew into Russia’s Google, doing search, autonomous driving, and other businesses, building a data center in Finland around 2012–2013. When the Russia-Ukraine war broke out in 2022, Arkady and the company opposed the war, sold Russian assets, and moved out of Russia, re-listing on NASDAQ. The entire team and engineers relocated to the Netherlands. So, Nebius’s stock chart shows a two- to three-year gap during delisting. Investing in Nebius is like betting on a mathematician who told Google no, refused to serve his own country, and moved abroad to go public. The culture and mission are similar to what I valued in Palantir.

Steve: I had no idea about this history. I mainly focus on GPU, financials, balance sheet, forward EPS, and revenue growth. CoreWeave looks interesting, but I dislike its debt load. As my view on AI deployment cycles shifts, I am now considering investing in all three.

Nate: I mention the background because most people don’t know. They sold Russian assets for about $5.5–$6 billion, which they used to build new data centers by late 2024.

Steve: What are the main differences among the three?

Nate: Three dimensions. CoreWeave entered earliest, with the highest active power and revenue, focusing on training (like GPT-5 for OpenAI). IRON is between training and inference. Nebius focuses on inference. Inference has better sustainability and higher unit prices, so CoreWeave commands a premium.

In terms of power scale, CoreWeave has over 1 GW activated; IRON has 4.5 GW locked, with land and power as its largest moats, but strangely, it only signed one big client—Microsoft—and the second big deal is delayed. All three have contracts with Microsoft. CoreWeave also works with OpenAI, Nebius with Meta, IRON only with Microsoft. For IRON holders, the next contract is a key catalyst, and the delay is odd.

Differentiation-wise, CoreWeave has scale + ClusterMAX Platinum certification; IRON offers cheap renewable energy + vertical integration (own land and power); Nebius emphasizes software stack. Their token factory inference platform is a core focus. Recent acquisitions of Egen AI and Clarif AI aim to optimize tokens and models, pushing toward hyperscaler markets. Arkady has repeatedly emphasized building the “fourth hyperscaler,” which should be the fifth overall if you count Oracle.

Nebius’s last quarter disclosed that contract revenue outside Microsoft and Meta grew 3.5 times quarter-over-quarter, a very bullish sign. The reason is that downstream clients need more software services, which yield higher margins. Contracts with Microsoft and Meta are bare-metal, while hyperscalers have their own container orchestration, so they only buy compute, with lower margins. Mid-sized downstream companies need software stacks, giving Nebius more profit elasticity.

Nebius also has four subsidiaries adding valuation: ClickHouse (data analytics and labeling platform, comparable to Databricks), holding 25% with a valuation of about $4.2B based on a $20 billion valuation in January; Avride (autonomous delivery robots, deployed in campuses like Columbus, partnered with Uber), with Nebius holding 83%, valued between $5 and $8 billion; and smaller platforms TripleTen (education) and Toloka (data labeling), worth a few million dollars each, contributing limited valuation but increasing optionality. Overall, Nebius’s valuation is based on GPU cloud plus subsidiary valuation sum. CoreWeave only has top-tier compute; IRON’s compute is early-stage with land and power locked-in, AI cloud revenue around $20–30 million; Nebius is in the middle, with higher revenue than IRON but higher capex needs due to licensing and permits in New Jersey and Pennsylvania.

Steve: We just reviewed all financial reports, including NVIDIA’s from two days ago, plus hyperscaler reports (excluding Oracle). Considering Jensen Huang’s comments, NVIDIA’s data, memory (Micron, SanDisk), and TSMC’s situation, how do you interpret these data’s impact on NeoCloud?

Nate: I’ve never been so bullish. Our AI usage is still very early. Every company will have a website in 20 years, and AI will follow the same path. We’re roughly at the early adopter stage of the internet in 1996. With robotics and autonomous driving added, we’re at the very bottom left of the S-curve. I believe NeoCloud will see explosive growth in the next three years—GPU demand alone can grow tenfold because customers need compute power. In five years, the focus will shift to software stacks and service quality, and Nebius’s layout is preparing for that phase.

Regarding memory and GPU price increases, Nebius announced yesterday that H100 and B200 prices increased 30–70% from June 1, fully passing on cost increases. Nebius’s mid-term EBIT margin guidance is 15–20%, with EBITDA margin at 40% by year-end. I don’t focus much on EBITDA, but EBIT (operating profit margin) is a useful reference.

Steve: All signals point to compute power shortages—CPUs and GPUs are insufficient. The most transparent are TSMC and Micron, as servers need memory, and chips need manufacturing. TSMC’s biggest customers are NVIDIA, whose largest customers are Microsoft, Amazon, Alphabet, and Oracle. TSMC’s massive capital expenditure this year isn’t for a one- or two-year cycle, nor is Micron’s. If they’re expanding capacity, it means I was wrong before. I’ve revised my AI cycle target twice, now considering a third revision. When ChatGPT launched, I estimated NVIDIA’s cycle at 18–24 months; a year and a half ago, I extended it to 24–36 months. Now I ask myself: will this slow down before 2030? I’m unsure. Alphabet said this year they’ll spend $190 billion, and next year even more. By 2028, they won’t slow down.

Nate: I agree with you. I initially saw Nebius as a short-term trade, but as I understand the market better and see the growth curve of capital expenditure and changing dialogue, I now see all three as long-term bets. The next step is the entire Fortune 500 and S&P 2000—every mid-sized company will say, “I also need compute power, I also need agent training.” That’s the real early stage. Nebius is already working with Shopify, Cloudflare, CrowdStrike. The next wave is “every company needs compute,” just like “every company needs a website.” Executives now think they should let the giants spend hundreds of billions first to prove what works, then follow. The 3.5x growth in non-hyperscaler contracts for Nebius is a clear signal.

Steve: I used to think GPU cycles followed RAM cycles, but now I believe GPU demand won’t stop, memory won’t, and capex won’t accelerate further. Giants won’t keep doubling or growing 50–60% annually, but 10–20% growth per year is entirely possible. Overlaying NVIDIA’s data with token economics, I must believe CoreWeave, Nebius, IRON will see substantial revenue expansion.

Nate: A simple formula for compute conversion: 1 MW = $10 million annual revenue; 100 MW = $1 billion. Bare-metal is cheaper, others more expensive, but that’s the baseline. Nebius’s guidance of 5 GW by 2030 equals $50 billion in revenue. Even the analyst consensus of $15B, compared to last year’s roughly $500 million, is a 10x increase in five years. Next year’s market consensus expects 550% YoY growth, then 219%—a consensus among 17 analysts. CoreWeave’s numbers are similarly staggering: 147%, 97%, 58%, all in the hundreds of billions in absolute value.

Regarding GPU depreciation shorts, Nebius’s last quarterly report explicitly states that each GPU has four competing clients.

Steve: I have a theory. Jensen said at a conference that many clients are still using five- or six-year-old H100s. I guess hyperscalers will lease out the new Vera Rubin GPUs instead of discarding old ones. Nebius and CoreWeave do the same. These compute resources will eventually flow to non-tech giants like Coca-Cola and Walmart.

Nate: Good point. Nebius’s depreciation cycle is five years. H100/B100 are nearly five years old, yet prices are still rising. After depreciation, any price increase yields 100% net profit.

Steve: What’s the biggest misconception in this sector?

Nate: “Hyperscalers can copy NeoCloud.” AWS was designed for early 2000s cloud services, with much lower requirements for compute intensity, cooling, and networking than AI workloads. They can’t just shut down cloud services overnight and switch to AI. Racks need reconfiguration to handle higher heat and load. Nebius, CoreWeave, and IRON are built from the ground up for next-gen AI workloads. Hyperscalers have experience and capital but can’t simply replicate. That’s the first major misconception. The second is Michael Burry’s claim that “GPUs will soon be obsolete,” which has been disproved by markets and recent price hikes—Nebius raised prices 30% yesterday. Giants will compete, but for the next three years, this isn’t critical. In five years, software and services will be key, so I favor Nebius’s software stack layout. In the next three years, blindly investing in a NeoCloud can outperform the market.

Steve: Who is most likely to win among the three, given customer overlaps?

Nate: Long-term five years, Nebius. IRON’s founders are two Australian salespeople; CoreWeave’s founders have an average technical background; only Nebius has a team of engineers who have worked together for 20 years, and the CEO is a mathematician. This gene influences culture and product direction. But a more probable outcome is a three-way split similar to AWS/GCP/Azure—say 40/40/20 or 35/35/30. In the short two years, CoreWeave leads because of Platinum certification and nearly $100 billion RPO. Nebius’s RPO is about $40 billion; IRON’s specifics I don’t know. Structurally, IRON owns its data centers fully but has weaker software; CoreWeave uses colocation with multiple companies sharing data centers, impacting margins; Nebius combines self-owned data centers with colocation. Nebius aims for 75% self-owned by 2030, gradually eroding IRON’s moat while maintaining margins. Recently, IRON has land and power advantages, but in 5–10 years, Nebius is the likely winner.

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

Nate: All three will grow. Looking to 2030, all are optimistic. After 2030, the market might be even more bullish—once the world fully understands how to use AI, the real race begins. All three will win, including smaller players like Cipher, WULF, Galaxy.

Steve: Final question: beyond compute, where is the next opportunity?

Nate: NVIDIA’s earnings signal a network business surge—this is a key signal. I am heavily invested in Arista Networks.

Steve: Heartbreaking—I’m a Cisco shareholder.

Nate: I like Arista’s vendor neutrality. Network capex conversion lag is longer than NeoCloud and GPU companies, but their network business exceeded expectations, while GPU was just meeting or slightly exceeding estimates.

Steve: Network revenue is already a major part of NVIDIA’s income—you’re right. Without racks, firewalls, switches, and WAP, data centers can’t operate. I know Cisco better because I’ve negotiated with them. Cisco has shifted from hardware to services and security, with 42–44% of quarterly revenue from those segments. They also acquired Splunk, and Cisco Umbrella is strong in security, with hardware tied to subscriptions. Both will grow. Non-tech folks don’t realize how big the switch business is. No one uses Ethernet at home anymore, but offices still need switches because signals decay over 300 feet, requiring main data centers and multiple IDF wiring closets. Data centers need more rack-mounted Ethernet. The network opportunity is larger than many imagine.

Nate: Arista has also risen 100%. A more stable 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. I personally hold significant positions in Arista at the network layer.