IREN, NBIS, and CRWV: The Rise of Neoclouds in the AI Infrastructure Boom

The cloud industry has long been dominated by major players like AWS, Azure and Google Cloud (we can also add Oracle and Alibaba in that group, but they don’t have the scale of the top three, yet).

In recent years, the explosive growth of artificial intelligence has created an unprecedented demand for specialized computing power, particularly high-performance GPUs capable of handling massive training and inference workloads.

Traditional cloud giants like AWS, Azure, and Google Cloud are dominant in general-purpose services, but building specialized computing power from scratch takes a significant amount of time and capital, as the demand is quite urgent. Here, the neoclouds step in to fill the gap.

Neoclouds share several core similarities that distinguish them from the incumbent cloud providers. Unlike the traditional clouds, which offer hundreds of managed services across a broad spectrum of computing, storage, and networking, neoclouds are highly specialized in GPU-centric infrastructure for AI workloads. Their primary clients include the hyperscalers (the traditional cloud providers themselves - Microsoft, Google, Amazon), Meta (not a cloud provider, but a hyperscaler), as well as large AI startups like OpenAI, xAI, etc. Emerging customer segments, such as sovereign AI initiatives and traditional large enterprises, are also beginning to appear.

At their essence, neoclouds procure GPUs, house them in data centers, and sell access to the raw computing power, supported by high-speed networking and optimized software layers for AI tasks.

Clients often prefer them for "bare metal" access and potentially lower pricing compared to traditional platforms' overhead. Importantly, neoclouds are infrastructure-first - they supply the "shovels" for the AI gold rush without directly competing against their customers by developing rival consumer-facing AI models.

Like electricity companies, power here is everything. When people talk about these three companies, they often mention capacity in terms of MW and GW. Roughly, 100 MW of capacity can translate to about $1 billion in annual revenue at maturity, and for newer GPUs, it can even reach $1.2-1.3 billion.

It is important to distinguish contracted capacity (future power under agreement, basically the backlog) from connected capacity (power ready for GPU deployment), and monetized capacity (actual revenue-generating GPUs). Gaps between these metrics often stem from delays in GPU delivery, cooling systems, or legacy operations.

 Revenue per GPU hour further differentiates them, reflecting pricing power, efficiency, and service sophistication. For H100/H200 rates, the pricing varies:

• Nebius: $3.00 – $3.50 (the highest because it includes a software element to its services);
• CoreWeave: $2.40 – $2.90 (mass volume);
• IREN: $1.90 – $2.50 (they charge the lowest, due to their pure “bare metal” service and cost advantages).

IREN started as a Bitcoin miner and still derives most of its current revenue from mining, though it is aggressively transitioning towards AI. Its mining capacity has scaled to 57 EH/s with the potential to generate around $1 billion in annual revenue from its legacy business. However, the management does not plan significant further expansion of mining beyond this level, instead channeling focus toward AI cloud services.

IREN sells 100% of its mined Bitcoins rather than holding them, providing a certain cash flow backing, but the volatility in BTC prices brings a large degree of uncertainty to this revenue line. Recent guidance targets AI-related annualised run-rate revenue (ARR) of around $3.4 billion by the end of 2026, potentially expanding further, with some projections reaching over $3.7 billion supported by a 150,000 GPU fleet.

IREN’s legacy business works on ASICs, so the company is also in the process of replacing ASICS with GPU infrastructure, which causes one-off impairments.

IREB is the most asset-heavy among the trio, owning its data centers outright. Not only that, but the data centers are also located in remote areas, primarily, with a surplus of energy procured at ultra-low costs (~$0.033 per kWh). This vertical integration eliminates co-location fees and supports lower GPU-hour pricing, a potential long-term competitive advantage.

Financially, IREN exhibits high shareholder dilution, perhaps being their greatest sin, with shares outstanding growing from 20.6 million in June 2021 to 298.2 million recently.

Margins benefit from asset ownership and Bitcoin as a cash cow, though aggressive depreciation (over 5 years) makes operating profit margins appear weaker. This approach is more conservative than CoreWeave's but less so than Nebius's. Software capabilities remain light; IREN focuses primarily on physical infrastructure rather than advanced managed services.

CRWV is basically the poster child of the neoclouds. Similar to IREN, CRWV also has cryptocurrency origins, having mined Ethereum, but unlike IREN, CRWV has fully exited crypto mining. It entered space with existing GPU expertise and has become the largest by operations and market capitalization.

Reported revenue reached $5.12 billion for 2025 (up significantly from prior years), with guidance for at least $12-13 billion in 2026, with the potential of this number to go even higher. The revenue backlog stands at a massive $66.8 billion.

To become the largest neocloud out there, CRWV fueled its ascent with a solid amount of debt ($30 billion in debt and quarterly interest expenses of $388 million, eating a huge chunk of its income statement). Much of this debt is collateralized by GPUs, but that holds risks. Furthermore, CoreWeave employs an asset-light model, leasing space from operators like Equinix and Digital Realty rather than building outright, a contributing factor to the high leverage (lease liabilities). This contrasts with IREN's ownership approach, where they hold the assets and also IREN’s preferred equity financing strategy.

Data centre locations are near major cities, incurring higher electricity costs but reducing latency for end users, which can also be an overhang for margins. Overall, profitability remains challenging due to rent, high interest, and elevated power expenses.

The accounting is the most aggressive of the three, with GPUs depreciated over 6 years (vs. IREN's 5 and Nebius's 4). This can flatter short-term metrics but raises questions about prudence. On the software side, CRWV utilizes Kubernetes orchestration for coordinating large GPU clusters, though its software capabilities are less sophisticated than Nebius's stack.

NVIDIA holds strategic ownership (~9%, similar to its stake in Nebius), granting preferential GPU access and underscoring industry ties.

Share dilution has occurred (from 180 million in 2022 to 436 million in 2025), but at a lower scale than IREN, though still quite massive.

Nebius lacks the crypto mining history the other two have, emerging instead as a spin-off from Yandex (Russia's tech giant, equivalent of Google/Uber) following the 2022 Ukraine invasion. The international assets—including a major Finnish data center, cash, and engineering talent—formed the core assets of the company. This gives Nebius a clean, non-crypto foundation and strong European presence, with expansion into North America for global reach. It also gives them a relatively clean balance sheet with some cash in hand.

Revenue for 2025 was approximately $529.8-530 million (479% growth), with 2026 expectations around $3-3.4 billion and ARR guidance of $7-9 billion by year-end. A recent Meta deal has boosted the backlog to around $49 billion (including prior agreements), with up to $27 billion from Meta alone over five years, starting significant revenue in 2027 (dedicated $12 billion tranche plus up to $15 billion optional). Microsoft also features prominently, contributing to customer concentration (~80% of booked capacity from Meta and Microsoft combined).

NBIS operates a hybrid data center model—owning key sites like Finland while leasing others—resulting in co-location fees higher than IREN but lower than CRWV.

Its architecture reportedly delivers 3x more compute per MW than peers, thanks to a proprietary software stack (Aether platform managing full-stack from silicon, Token Factory for optimized low-latency token-based billing). This enables higher revenue per GPU hour and functionalities like efficient data flow management, disruption fixing, and preventing data starvation. NBIS also designs its own cooling systems, reducing reliance on third parties like Dell or Supermicro. Non-AI businesses (data labelling, robotics, edtech) contribute only ~10%.

The balance sheet is relatively healthy: lower debt (~$4.9 billion total, net ~$1.2 billion; debt/assets 39%, net debt/assets 10%) and limited historical dilution, funded partly by Yandex spin-off cash and customer prepayments. Recent capital raises (including convertible notes post-Meta deal) and NVIDIA's strategic ~8-9% investment (plus launch partnership for Vera Rubin platform) provide support.

Depreciation is the most aggressive (4 years), signalling prudence. Profitability faces pressure from sales hiring and R&D on software, with less contract transparency than peers. NVIDIA's involvement mirrors its CoreWeave stake, enhancing credibility.

All three face sector-wide risks.

The biggest risk is, of course, a potential AI winter, reducing demand and GPU pricing. Another risk is space-based data centers disrupting terrestrial models, but that’s more of a worry in the not-so-near future. Other issues are related to accounting, especially around depreciation and circular economies. This fear can be reinforced by the excessive customer concentration: IREN with Microsoft (55% of 2026 revenue), NBIS with Meta/Microsoft (80%), CRWV with OpenAI/Meta (~55% of backlog).

Company-specific vulnerabilities differ.

CRWV's leveraged balance sheet (highest debt/assets at 60%, net at 37%) poses bankruptcy risk in a downturn, with aggressive accounting and high interest amplifying pressure, making CRWV the stock with the greatest downside here.

NBIS and IREN, despite being in a healthier situation, still need funding, and there is a high chance they will tap the capital markets for further funding (read “stock dilution”).

But as much as there are risks, there are also advantages. IREN's infrastructure ownership and low-cost power create cost leadership and defensiveness, while NBIS's software stack delivers efficiency and differentiation (higher pricing, optimization).

From a risk/reward lens, all offer high upside in the AI infrastructure supercycle, but downsides vary. CRWV represents the aggressive play—scale and backlog promise hypergrowth ($12-13B+ revenue in 2026), yet debt and accounting warrant caution; an AI winter could severely impact equity.

NBIS is the efficiency play, with a software moat, pure-play focus, and strong utilization driving potentially superior long-term margins, though execution on global expansion and customer reliance introduces risks.

IREN emerges as a safe option, blending Bitcoin cash flow, asset ownership for margins, and vast power runway, tempered by dilution and slower AI monetization.

Execution on capacity ramp, utilization, and customer diversification will ultimately determine which neocloud captures the most value from the AI boom. With hyperscalers still capacity-constrained, these "shovels" providers are positioned for substantial, if volatile, growth through 2026 and beyond. The differences in their approaches—asset-heavy vs. software-enhanced vs. scale-leveraged—ensure they will shape the neocloud landscape in complementary yet competitive ways.