CPU returns to the core of AI! Who are the big winners?

牛牛課堂

joined discussion · ·

Surging prices for North American computing power rentals! With Neocloud making frequent moves, how should this sector be viewed?

"Price hikes" — undoubtedly the most critical keyword in this year's AI industry chain.

This trend is gradually spreading to the rental prices of computing power. According to monitoring data from New York-based data provider Ornn, the spot rental prices of NVIDIA’s entire range of GPUs in cloud data centers have soared to varying degrees in recent months. The hourly rent for its most advanced Blackwell series chip has reached $5.17.

Coincidentally, in early AprilBreaking the Depreciation Curse! GPU Rental Fees Soar, How Far is Neocloud’s ‘V-Shaped’ Rebound?, it was mentioned in an article that, according to the latest data released by semiconductor research firm SemiAnalysis, the one-year lease contract price for H100 rebounded strongly from a low of $1.7 per hour in October 2025 to $2.35 in March this year, marking a rise of nearly 40% over the period.Since the publication of that article, $NEBIUS (NBIS.US)$、 $CoreWeave (CRWV.US)$、 $IREN Ltd (IREN.US)$both rose by about 59%, $Oracle (ORCL.US)$while another increased by over 29%.

From rising underlying computing costs to explosive terminal stock prices, the market has already responded with real investments.So the question arises: What is the core logic of the Neocloud sector at this stage? What are the latest developments in this industry? Which companies are worth paying close attention to? Let’s break it down step by step.

What is the core logic behind the Neocloud sector?

First, if traditional cloud computing is likened to an 'all-encompassing department store,' then Neocloud is more akin to the 'computational water, electricity, and gas' exclusive to the AI era.

Their underlying logic is to proactively 'simplify': rather than attempting to encompass all complex cloud services, they are custom-built from the start for AI computation. From GPU clusters, high-density data centers, liquid cooling systems, to IB high-speed networks and bare-metal delivery, every design element aims for one ultimate goal — delivering AI computational power to customers in the 'fastest, most stable, and most cost-effective' way possible without interruption.

So why has Neocloud (emerging cloud) become the standout黑马 in this year's AI industry chain?The core lies in 'structural gaps' and 'absolute pricing power':

Supply side: The 'capacity pain period' of traditional cloud — AI applications are booming, but the old data centers of traditional cloud giants cannot handle high-power AI computations. Building new data centers requires a cycle of over three years. Neocloud takes advantage by cutting unnecessary features and focusing on providing pure and efficient GPU computing infrastructure, perfectly filling the market gap.

Demand side: The 'computational flywheel' of Token economics — With the widespread adoption of AI applications, the surge in demand for large model token usage directly drives exponential growth in the need for underlying computational hardware.

Business Model: Strong 'seller’s话语权' —— The supply-demand imbalance has given Neocloud vendors significant control. SemiAnalysis research points out that Neocloud (referring to 'New Cloud,' an emerging cloud computing service model or market category focused on providing high-performance GPU computing power for AI workloads to address structural shortages in traditional cloud providers' AI computing capacity) vendors and hyperscalers hold the upper hand, allowing them to negotiate more favorable terms such as higher upfront payments, better pricing, longer contract durations, and even the ability to choose start and end dates of contracts to align with their inventory situations.

What are the latest developments in this industry?

With the explosive demand for computing power, the Neocloud (emerging cloud) sector is reaching a peak of industrial chain integration. Two recent landmark acquisitions reveal that the industry is rapidly evolving from 'pure hardware rental' to 'integrated hardware and software':

1. Nebius acquires Eigen AI: Upgrading from 'computing power landlord' to 'full-stack AI powerhouse'

Nebius, as an emerging AI cloud platform, originally focused on providing GPU clusters and Token computing power factories; Eigen AI, on the other hand, specializes in model inference and optimization.

Underlying logic: This means that Nebius is no longer satisfied with merely renting out bare-metal GPUs but instead integrates Eigen’s inference and post-training optimization layers to package GPUs, inference engines, auto-scaling, and fine-tuning pipelines, delivering a ready-to-use enterprise-grade AI ecosystem platform directly to customers.

2. IREN acquires Mirantis: Completing the software puzzle, targeting 'full-stack AI cloud'

As a pioneer in computing power transformation, IREN announced the acquisition of Mirantis, a veteran player from the previous generation of cloud computing infrastructure (a core participant during the OpenStack era), marking a qualitative leap in its AI cloud service capabilities.

Underlying logic: With the acquisition of Mirantis, IREN has completely broken through the limitations of the single-GPU rental model, gaining the ability to deliver full-stack AI cloud services encompassing 'bare metal + containers + management platform' on top of its self-built computing clusters. The market widely believes that this strategic move represents the most substantial step forward in IREN's pursuit of catching up with first-tier players like Nebius and CoreWeave.

Moreover,Looking ahead this month, CoreWeave and IREN will release earnings after the market closes on May 7, while Nebius will announce its results after the market closes on May 13.Due to the recent strong stock price trend, market sentiment has already priced it in. Therefore, investors need to closely monitor whether management can deliver better-than-expected results and demonstrate solid execution to investors.

Which companies are worth focusing on?

previouslyMeta, Anthropic and other giants double down on AI computing power! What’s the outlook for the new cloud (Neocloud) track?The article also outlined the core North American vendors in Neocloud as follows:

According to CMB Securities’ research report, considering the ranking of cloud service capabilities and evaluating current cloud service agreement orders, key attention will be given to leading Neocloud vendors listed on US stock exchanges with cloud service agreements exceeding tens of billions. $CoreWeave (CRWV.US)$ 、 $NEBIUS (NBIS.US)$ 、 $IREN Ltd (IREN.US)$as well as$Oracle (ORCL.US)$ Specifically:

1. CoreWeave: The Aggressive New Giant in AI Cloud Infrastructure

$CoreWeave (CRWV.US)$The business model is essentially that of a computing power broker.Unlike traditional cloud service providers, the company does not offer CPUs, storage, or security services but focuses solely on renting out GPU computing power.Customers are mainly divided into two categories,one being tech giants like Microsoft and OpenAI, which require urgent capacity expansion and contribute the vast majority of revenue; the other being smaller AI companies unable to build their own GPU clusters. The company’s business model means that the ability to secure GPUs determines customer stickiness.

CoreWeave specializes in bare-metal services, directly providing hardware computing power.CoreWeave offers fast, reliable, large-scale bare-metal GPU services—providing thousands, or even potentially millions, of GPU devices at once. CoreWeave eliminates the traditional virtualization layer (hypervisor) and runs Kubernetes directly on physical machines. This architecture removes the additional overhead and performance loss caused by virtualization, achieving the lowest latency and highest performance. CoreWeave's analysis shows that traditional virtualized environments can waste up to 65% of effective GPU computing capacity, and the bare-metal architecture aims to solve this pain point.

Currently, CoreWeave is actively expanding its scale, increasing active power capacity to approximately 850MW, with contracted power capacity reaching 3.1GW.Among this, CoreWeave will develop a new data center campus in Kenilworth, New Jersey, with a capacity of up to 250MW. This is CoreWeave’s first greenfield data center project specifically for artificial intelligence, with the first phase expected to be delivered by 2026. The company expects the vast majority of contracted power capacity to be operational within the next 12 to 24 months.

Currently, CoreWeave holds a large number of major client computing power rental orders,Including but not limited to: (1) A major cooperation agreement with NVIDIA worth up to $63 billion. It commits to purchasing all of CoreWeave's unsold computing power capacity over the next decade. (2) Announced an agreement to provide Meta with $142 billion worth of AI cloud infrastructure, along with an expanded cooperation worth an additional $210 billion. (3) Reached a new cooperation agreement with OpenAI valued at $65 billion. The total contract amount has surged to $224 billion. (4) A multi-billion-dollar order agreement with Anthropic.

Additionally, it is worth noting that CoreWeave possesses industry-leading capabilities in acquiring GPUs.CoreWeave is NVIDIA’s first Elite Partner (the highest level) and also the first AI cloud provider to offer NVIDIA RTX PRO 6000 Blackwell GPUs on a large scale.

2. Nebius: Vertically Integrated AI Infrastructure Provider

A full-stack infrastructure provider in Europe. $NEBIUS (NBIS.US)$ Originally founded as Yandex N.V. in 1989, after the Russia-Ukraine conflict, all non-Russian assets of Yandex were spun off to become Nebius, including a data center in Finland, $2.3 billion in cash reserves, and a range of other divisions. The company resumed trading on the exchange on October 21, 2024, positioning itself as a European AI infrastructure and service provider, building full-stack infrastructure including GPU clusters, cloud platforms, tools, and developer services.

Following its transformation, Nebius’ market capitalization has rapidly increased.NVIDIA participated in the financing.As of Q4 2025, Nebius operates seven data centers across Europe, the US, and the Middle East. Nebius will continue to expand its capacity, with projected contracted power capacity exceeding 3.0 GW by the end of 2026 (up from last quarter’s guidance of 2.5 GW), with connected power ranging from 800 MW to 1 GW.

Nebius is a Preferred Partner of NVIDIA and is also one of the six global cloud service providers on NVIDIA's Reference Platform Cloud Partner (Reference Platform NCP).

Nebius provides a full-stack platform,integrating software toolchains with the core goal of building an end-to-end cloud service platform capable of serving as a 'production-grade inference platform.' Nebius not only serves top AI tech companies that train their own models but also attracts many small or medium-sized development and research teams within large enterprises. These customers have relatively limited technical capabilities and are more willing to pay for an 'out-of-the-box' complete solution, thus placing greater value on the comprehensive software abstraction layer services provided by Nebius.

3. Iren: Green computing power infrastructure provider based on renewable energy

From mining facilities to computing power plants. $IREN Ltd (IREN.US)$ Founded in 2018 by brothers Daniel Roberts and William Roberts in Sydney, Australia, the company’s initial mission was to become a sustainable Bitcoin mining firm powered entirely by renewable energy. In response to volatility in the cryptocurrency market, the company decisively initiated a strategic transformation, focusing investment in the AI cloud services sector, while temporarily halting the expansion of its Bitcoin mining operations.

The company's latest guidance maintains a projection to achieve $3.4 billion in annualized run-rate revenue (ARR) from AI Cloud services by the end of 2026.Among this, Microsoft contracts are expected to contribute approximately $1.94 billion in ARR, with around 63,000 GPUs deployed at the British Columbia campus anticipated to bring in about $1.5 billion in ARR.

Iren owns seven self-built data center campuses located in British Columbia, Canada, as well as Texas and Oklahoma in the United States; currently, four data centers are operational, with a running capacity of 810MW and contracted electricity exceeding 4.5GW.

The Childress campus Horizon 1 (50MW IT load) has been recently completed, while construction and equipment procurement for the Horizon 2 site are ongoing; Sweetwater 1 (1.4GW) is expected to be powered up by April 2026, and Sweetwater 2 (600MW) is planned to go live by the end of 2027; the existing 160MW capacity in British Columbia, Canada, is being continuously transitioned from ASICs to GPUs, with the goal of completing this by the end of 2026.

Mining firms transitioning into AI benefit primarily from their vast and low-cost power and land resources.Bernstein analysts pointed out that mining companies' grid-connected power can reduce AI data center deployment time by 75%, with very low capital expenditure required to retrofit existing facilities, making them ideal partners for AI cloud providers. This model allows mining companies to quickly enter the market and seize the opportunity.

Iren leverages its mining network, access to low-cost electricity (3.5 cents/kWh), grid connection capabilities, and internal engineering expertise to develop a vertically integrated high-density data center. This vertical integration model eliminates reliance on third-party hosting providers, and more importantly, removes all associated counterparty risks, enabling Iren to debug GPU deployments faster and maintain full control over execution and uptime.

4. Oracle: Traditional database giant transitions into an AI cloud infrastructure powerhouse

$Oracle (ORCL.US)$ At the annual investor conference held on October 17, 2025, in Las Vegas, clearer guidance updates were provided regarding the gross margin of artificial intelligence infrastructure projects, the latest order situation, and the company’s long-term financial goals, further boosting market confidence:

GPU Cloud Gross Margin:For example, taking AI workload infrastructure projects (such as data centers), a six-year AI infrastructure project with total revenue of $600 billion could achieve a gross margin of 35%, and this level of gross margin 'is also relevant even for the largest customers.'

GPU Cloud Large Agreement Status:Within 30 days of the previous quarter, Oracle's cloud computing division, Oracle Cloud Infrastructure, had booked $65 billion in new commitments, including a $20 billion deal with Meta, while the latest $65 billion in orders all come from clients outside OpenAI.

Long-Term Financial Target Guidance Raised:Revenue is expected to reach $225 billion by fiscal year 2030, significantly higher than the average analyst estimate of $198 billion. Additionally, the company expects adjusted earnings per share (EPS) to reach $21 at that time, compared to the current consensus estimate of $18.5. Previously, at the end of 2024, Oracle announced that in FY2029, annual revenue would grow to at least $104 billion.

Summary

Overall, China Merchants Securities believes that the rise of AI Neocloud is fundamentally driven by the overflow demand for AI computing power due to structural mismatches in supply and demand, the outsourcing demand for assets brought about by changes in capital structure, and the combined effects of evolving technical requirements and ecosystem interest restructuring. Both domestic and overseas AI giants have a continuous incentive to seek computing power leasing firms for outsourcing their computing needs.

The bank believes that the long-term value of Neocloud vendors stems from three core competencies: First, control over scarce computing resources (GPUs, electricity, etc.), which determines their scalable growth potential; second, deep AI infrastructure optimization capabilities, which determine the cost structure, utilization, and profitability per unit of computing power; and third, the ability to manage financial leverage, which determines the efficiency of business returns.

404K Views