As SaaS Seat Model Collapses, Inference Surpasses Training: Decoding the Diverging Fates of Neocloud Giants CoreWeave, Nebius, and IREN

1. Structural Fission of AI Computing Demand: Industrial Implications of the Anthropic Phenomenon

According to the latest research report released by SemiAnalysis in early May 2026, Anthropic's annualized recurring revenue (ARR) was only approximately $1 billion in December 2024, reached $9 billion by December 2025, and has since surged to $44 billion as of May 2026. In other words, the company's revenue base has increased nearly fivefold in the past five months, adding an average of $96 million in net ARR per day. Meritech Capital partner Alex Clayton noted in public comments that he has studied the IPO filings of more than 200 public software companies and has never seen such a growth curve; notably, Anthropic's ARR had only just jumped from $1 billion to $3 billion when he made that comment, and the growth rate has accelerated rather than slowed since then.

Even more noteworthy are the marginal metrics disclosed in the same report: the gross margin of Anthropic's self-operated inference infrastructure has jumped from 38% a year ago to over 70% currently. This means the company is not only rapidly expanding its revenue base but also improving its unit economics, contrary to the stereotype that "AI companies can only burn cash through financing."

However, Anthropic's own infrastructure simply cannot carry the scale of computing power required behind this explosive growth. A common misconception in the market is that Anthropic's computing power comes from Amazon Web Services and Google Cloud and has no direct relationship with so-called "Neoclouds." This observation is not entirely accurate. Anthropic's computing structure is a microcosm of the current AI industry: Amazon and Google handle the leading training loads, but demand that spills over the proprietary capacity of hyperscalers naturally flows to third-party providers. Anthropic's multi-year computing contract with CoreWeave in April 2026, and reports of talks with Nebius in May, are direct manifestations of this spillover mechanism.

Understanding this spillover logic is crucial for evaluating the Neocloud sector. The business foundation of the "Neocloud Big Three"—CoreWeave, Nebius, and IREN—is not derived from directly serving frontier AI labs like Anthropic, but from capturing the computing gap created by the explosive growth of the entire AI industry. This article will systematically analyze two structural changes driving the re-rating of Neocloud valuations, the differentiated positioning of the three companies in the industry chain, and the role of power supply chain bottlenecks in reshaping the competitive landscape, ultimately summarizing an analytical framework for long-term investors.

2. The First Structural Change: The Collapse of the SaaS Seat-Based Business Model

For the past two decades, the core profit model of the SaaS industry has been built on "seats." The logic is quite simple: if a company has 100 employees who need to use Salesforce, Salesforce charges a monthly fee based on 100 seats. This model is highly attractive to investors due to its stable and predictable revenue, allowing companies like Salesforce, ServiceNow, and Workday to build market capitalizations in the tens or even hundreds of billions of dollars.

However, the emergence of generative AI has completely broken this logic. A specific scenario illustrates this rupture: suppose a company has only one engineer, but he uses an AI programming tool like Cursor, allowing him to assist in refactoring 5,000 lines of code within an hour, triggering hundreds of API calls and consuming tens of millions of tokens. In this context, "seat count" and "actual computing consumption" have become completely decoupled. For SaaS companies, continuing to charge by seat means an employee pays a $20 monthly subscription fee, while the backend computing cost consumed could be as high as $50—this is the fundamental reason why AI-native tools like Cursor and Claude Code faced losses in their early stages.

To address this structural contradiction, the industry is collectively shifting toward a hybrid model of "base subscription fee + usage-based billing": companies pay a fixed monthly fee as an entry threshold, while token consumption exceeding the quota is priced separately. Currently, leading companies such as OpenAI, Anthropic, Microsoft, and HubSpot have completed or are undergoing this transition. Gartner predicts that by the end of 2026, 70% of enterprises will prefer usage-based billing models over pure seat-based systems.

This trend also explains how Anthropic achieved a fivefold increase in ARR within five months. Its growth momentum comes primarily from the exponential expansion of token consumption by existing enterprise customers rather than simple new customer acquisition. Data publicly disclosed by Anthropic shows that the number of enterprise customers paying more than $1 million annually has doubled in the past two months; eight of the Fortune 10 companies are its paying customers; and its Claude Code product reached an ARR of $2.5 billion in February 2026, with enterprise users accounting for over 50%. Together, these figures convey a core fact: the business model of the AI era no longer sells seats, but tokens—the value of a seat is fixed, but the demand for tokens is theoretically infinite.

3. Second Structural Change: Inference Computing Historically Overtaking Training

If the collapse of the seat-based model is a shift at the business model level, the overtaking of training by inference computing power is a fundamental reshaping of the infrastructure demand structure.

To briefly explain these concepts: training is the process of teaching a large model existing knowledge, while inference is the process of the model assisting users in solving specific problems after deployment. In the past two years, the market's focus on AI computing power has been mainly concentrated on the training side—OpenAI's hundreds of millions of dollars spent training the GPT series and Meta's construction of large gigawatt-scale AI data centers all fall into this category. However, the industrial landscape is changing rapidly.

According to the latest forecasts from Gartner and Deloitte: in 2023, training accounted for 67% of AI computing power while inference accounted for only 33%; by 2025, the two will each account for half; by 2026, inference will have overtaken training to reach 67%; and by 2030, the share of inference will reach 70%. Data disclosed by NVIDIA itself is even more direct—currently, 80% of enterprise AI budgets are spent on inference, with only 20% invested in training.

Why does this reversal have structural significance? The fundamental reason is that the consumption curves of the two are completely different. Training is a one-time R&D expenditure—computing consumption ends once the model is trained; however, inference is a continuous operational expenditure—as long as users are still calling the AI, inference power will continue to be consumed. The arrival of the Agent era further amplifies this trend by several orders of magnitude: in the past, a user asking ChatGPT a question triggered only one LLM call; but now, asking an Agent to assist in analyzing a financial report might automatically trigger dozens of search, read, verify, and summarize calls, each step consuming tokens. This is the fundamental reason why the highly agentic Claude Code product from Anthropic consumes more than ten times the computing power per user compared to traditional ChatGPT users.

A key question arises here: since inference demand is so vast, why are Neoclouds benefiting instead of hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud? If this question is not clarified, the entire Neocloud bullish argument will not hold. In fact, the logic of inference spilling over from hyperscalers includes three independent but mutually reinforcing layers.

The first layer is the short-term supply-demand mismatch: Hyperscaler capacity has long been filled by their own flagship customers—Microsoft Azure serves OpenAI and Anthropic, Amazon Web Services serves Anthropic, and Google Cloud supports both Anthropic and its own Gemini. Even if these giants maintain hundreds of billions of dollars in annual capital expenditures, construction speed still cannot keep up with demand growth. Microsoft's $17.4 billion contract with Nebius and $9.7 billion contract with IREN are direct evidence of this spillover demand—if Microsoft's own data centers were sufficient to meet demand, there would be no need to procure from third parties.

The second layer is the structural difference in price advantage: The Tokenomics report released by Deloitte in January 2026 clearly stated: Tokenomics The report clearly points out that for medium-scale token workloads, Neocloud prices are 30% to 80% cheaper than traditional hyperscalers. This difference does not come from price wars but from generational differences in infrastructure. Neoclouds have no legacy baggage—their data centers were designed from the start with the GPU as the core, without the need to carry redundant costs to support traditional general-purpose cloud services. Amazon, Microsoft, and Google are adding AI facilities on top of existing clouds, while Neoclouds are pure AI-native; the unit cost structures of the two are completely different.

The third layer is customers' active demand for supplier diversification: Large AI customers like Anthropic are wary of over-reliance on a single cloud provider—once locked in, bargaining chips are lost. Therefore, they actively diversify their sources of computing power, leaving a portion of demand for neutral third-party providers like CoreWeave and Nebius. While Anthropic signed a historic order for 1 million TPUs with Google, it also engaged with CoreWeave and Nebius; the core motivation is not necessarily a lack of Google's computing power, but ensuring it always has diverse supply channels.

With these three layers of logic combined, the explosion in inference computing demand is not an accidental opportunity for Neoclouds, but an inevitable result of changes in industrial structure.

4. Differentiated Positioning of the Three Neocloud Giants: Training Machines, Inference Platforms, and GPU Landlords

4.1 CoreWeave: The Training Leader Dedicated to Frontier AI Labs

CoreWeave's (CRWV) market positioning is very distinct—it is one of the few GPU cloud infrastructure providers currently capable of simultaneously serving the world's four major AI R&D giants. Meta, Anthropic, OpenAI, and Google have all signed major computing contracts with CoreWeave in 2026, a customer structure that was hard to imagine in early 2025 but has become a reality in 2026.

Regarding financial data, CoreWeave's accumulated Remaining Performance Obligations (RPO) reached $66.8 billion as of the end of December 2025, a significant increase of more than $50 billion compared to the end of 2024 (over 4x year-over-year growth). On April 9, 2026, CoreWeave added a new $21 billion contract to the $14.2 billion contract signed with Meta in September 2025, bringing Meta's total committed amount to $35.2 billion, with the contract period extended to December 2032. Combined with the multi-year contract signed with Anthropic during the same period, Wells Fargo estimates that CoreWeave's actual contract scale has exceeded $90 billion.

The company's revenue guidance for 2026 is $12 billion to $13 billion, representing a median year-over-year increase of 140%; management further expects annualized revenue to reach $17 billion to $19 billion in 2026 and to break through $30 billion in 2027. To support this growth target, the 2026 Capital Expenditure (CapEx) guidance has jumped from $14.9 billion in 2025 to $30 billion–$35 billion.

One point worth clarifying is that CoreWeave does not "only do training." Its software stack is also actively moving into the inference field, including CKS (CoreWeave Kubernetes Service), SUNK (Slurm on Kubernetes), and the acquisition of Weights & Biases—all strategic moves to strengthen inference capabilities. However, its customer structure determines that the current business focus remains on training—OpenAI, Anthropic, and Meta typically use CoreWeave to train frontier models first, with inference loads being a natural extension after training is completed. The core assumption for investing in CoreWeave is essentially the judgment that "training demand will continue to increase."

In terms of financial risk, CoreWeave's total debt as of the end of 2025 was approximately $21 billion, plus $34 billion in off-balance-sheet lease obligations, resulting in a highly leveraged capital structure. Q4 single-quarter interest expense reached $388 million, more than double the $149 million in the same period of 2024. Q4 adjusted EBITDA was $898 million, with an EBITDA margin as high as 57%, indicating strong unit economics in the core business; however, the GAAP net loss widened to $452 million, showing that the pressure of capital investment on short-term profits has not yet eased. Management expects Q1 2026 to be the low point for profit margins (with adjusted operating profit of only $0–$40 million), gradually recovering to low double digits for the full year 2026, with a long-term target set in the 25%–30% range.

Customer concentration is another risk factor that cannot be ignored. Microsoft contributed 62% of CoreWeave's revenue in 2024 and still accounted for 67% in 2025, reaching as high as 71% in Q2. Although management expects Microsoft's share to fall below 50% as contracts with OpenAI, Meta, and Anthropic are executed, calculations by Next Platform suggest that in the expanded $87.8 billion backlog, Meta accounts for 40.1% and OpenAI for 25.5%, together totaling nearly 65%. In other words, customer concentration has not disappeared; the protagonists have simply changed. In early May 2026, the Wall Street Journal reported that OpenAI's internal user and revenue data failed to meet internal targets, causing CoreWeave to drop nearly 3% in a single day, illustrating the market's sensitivity to this exposure. Additionally, institutional investor Magnetar sold 1.28 million shares on May 1, 2026, cashing out $154 million, which is also an internal signal worth noting.

CoreWeave will release its Q1 2026 financial report after the market closes on May 7, 2026. According to LSEG market consensus, Q1 revenue is expected to be $1.9 billion to $2.0 billion, with adjusted operating profit expected at $0–$40 million. Three key observation metrics include: whether the quarterly growth rate of RPO can exceed 50% (the market expects 35%–40%, with an unexpected breakthrough being a strong bullish signal); whether unit economics turn positive for the first time; and whether the OpenAI computing delivery schedule is proceeding as planned.

4.2 Nebius: An Inference-First Full-Stack AI Platform

Nebius Group's (NBIS) positioning is completely different from CoreWeave's. If CoreWeave is likened to a "training machine," Nebius's core positioning is an "inference-first full-stack platform." Its flagship product, Token Factory, adopts a usage-based billing model per token, aligning perfectly with the external billing logic of large model manufacturers like Anthropic and OpenAI.

To strengthen this strategic direction, Nebius announced on May 1, 2026, that it would acquire Eigen AI for $643 million—a company incubated by MIT's Han Song Lab whose core technology, Activation-aware Weight Quantization (AWQ), allows a single NVIDIA GPU to output more tokens. In a time of extreme computing scarcity, this technology directly determines the unit earning efficiency of Nebius's inference business.

At the contract level, Nebius currently holds a contract backlog of nearly $50 billion, mainly from three major orders: a $17.4 billion contract signed with Microsoft in September 2025 (according to Nebius's 13-F SEC filing, Microsoft has committed approximately $6.96 billion in upfront payments, with the contract delivered in nine batches, the first two of which were delivered as scheduled in November 2025 and February 2026); an expansion of the original $3 billion Meta contract to $270 billion in March 2026; and NVIDIA's $2 billion investment as a strategic investor in March 2026. A detail in the $270 billion Meta contract is easily misinterpreted: only $120 billion belongs to Meta's proprietary procurement, while the other $150 billion is a commitment from Meta as a "backstop buyer"—Nebius will first attempt to sell this capacity to other third-party customers, and Meta will only cover the unsold portion. This structure makes Nebius's actual customer concentration more dispersed than the surface numbers suggest. Regarding revenue guidance, Nebius's full-year 2026 revenue is expected to be $3 billion to $3.4 billion, a growth of approximately 600% from the $530 million base in 2025; the annualized ARR target for 2026 is set between $7 billion and $9 billion. Q1 2026 consensus revenue expectation is $389 million, a year-over-year increase of 600%.

In the entire Neocloud bullish narrative, the most critical positioning is that the only specific product currently verifiable by financial data for "usage-based token billing" is Nebius's Token Factory. However, the company has not yet publicly disclosed Token Factory's specific revenue contribution, customer count, or share of ARR. This means that Nebius's Q1 earnings release before the market opens on May 13, 2026, will be a litmus test for the entire Neocloud industry narrative—if the customer count and ARR share for Token Factory do not show a significant jump, it will shake not only the bullish logic for Nebius alone but the foundation of the entire usage-based token billing narrative.

On the risk side, Nebius's stock beta is as high as 2.97, meaning the stock could drop 3% when the broader market drops 1%; it is a high-reward, high-volatility growth stock, not a ticker that one can hold long-term without concern.

Additionally, Nebius continues to raise funds through ATM (at-the-market) offering mechanisms, exerting continuous dilution pressure on existing shareholders.

Wolfe Research initiated coverage on the stock in late April 2026 with a fair value range of $80 to $170—a range width of nearly 100%, directly reflecting the high market uncertainty regarding its future value.

Furthermore, data shows that Nebius has missed revenue expectations in three of the past four quarters; the May 13 earnings report must deliver on its promises, or the stock price reaction will be quite violent.

Besides the share of Token Factory revenue, the focus of the earnings observation also includes whether the customer structure is further diversifying and the upgrade progress of the Aether platform. Aether is Nebius's full-stack cloud management platform—essentially its underlying "operating system"—and is an independent but complementary core product to Token Factory. Goldman Sachs's raising of its price target to $205 lies in its confidence in the execution of Aether and Token Factory.

4.3 IREN: A GPU Landlord Focusing on Physical Infrastructure

IREN's (IREN) market positioning is the most unique. If CoreWeave targets training and Nebius targets inference, IREN represents another business model—the bare-metal landlord, or "GPU landlord." Unlike the former two, it does not delve deep into the software stack but focuses on owning physical infrastructure such as land, power, and renewable energy assets.

IREN's primary customer is Microsoft—in November 2025, the two parties signed a 5-year contract worth $9.7 billion, which, once deployment is complete, will bring in approximately $1.94 billion in annualized recurring revenue (ARR), with an EBITDA margin as high as 85%, a top-tier level among Neocloud peers.

However, this contract contains a numerical structure that is often misinterpreted. From a capacity perspective, the contract occupies only 10% of IREN's total computing capacity—but from a revenue perspective, its $1.94 billion ARR accounts for approximately 57% of the company's $3.4 billion annualized ARR target for its AI cloud business in 2026. The same contract leads to very different conclusions from two perspectives: it is dispersed in terms of capacity (IREN still has 90% of its capacity available for sale), but from a revenue perspective, IREN's current AI cloud business is highly concentrated on a single customer: Microsoft. This contrast between capacity and revenue stems from differences in the unit economics of IREN's internal businesses. The 10% capacity used by Microsoft is loaded with NVIDIA's highest-end GB300 GPUs and utilizes liquid cooling technology, contributing approximately $9.7 million in annual revenue per MW; however, most of the remaining 90% capacity is still running Bitcoin mining—using ASIC miners rather than GPUs, with an annual revenue per MW of only $1 million to $3 million. The unit earning efficiency of the two differs by three to ten times.

IREN's competitive advantages are tangible: it has the lowest financial leverage among the three, owns its own land and grid interconnection rights, has renewable energy assets capable of self-generation, and can use cash flow from Bitcoin mining to subsidize capital expenditures for new data centers. However, its drawbacks are equally apparent: its software layer is nearly nonexistent, resulting in revenue per MW of only about one-fourth that of CoreWeave (CoreWeave can reach $10 million to $12 million per MW); equity has been severely diluted—IREN's outstanding shares were only 20.6 million in June 2021, but have ballooned to nearly 300 million within five years, a dilution of more than 14 times. Most critically, its AI cloud business has yet to truly take shape, leaving it with little bargaining power in the current environment of extreme computing power scarcity.

Therefore, while IREN's strategic pivot is correct and its low leverage and proprietary power resources constitute a genuine moat, a clear catalyst is required for the market to re-rate it at the valuation multiples of its Neocloud peers—specifically, AI cloud revenue must structurally surpass Bitcoin mining revenue. Until then, the market will likely continue to value it as a "Bitcoin miner with AI operations."

5. Power Supply Chain Bottlenecks Reshaping the Competitive Landscape

An analysis of the Neocloud sector would be incomplete if limited to the differentiated positioning of these three companies. A variable widely underestimated by the market is the power supply chain bottleneck, the impact of which is inconsistent across the three firms.

According to reports from Bloomberg, TechSpot, and other media outlets, one-third to one-half of the AI data centers planned for construction in the U.S. by 2026 may face delays or cancellations. The core issue is not a lack of funding or GPU shortages, but rather that the power infrastructure itself cannot keep pace with demand growth—lead times for high-voltage transformers have surged from 24 to 30 months prior to 2020 to five years today. This supply bottleneck stems from the U.S.'s rapidly increasing reliance on Chinese electrical equipment, with transformer imports from China rising from approximately 1,500 units in 2022 to over 8,000 units in the first 10 months of 2025.

On the surface, this risk appears bearish for all Neocloud companies, but a deeper breakdown reveals a completely asymmetrical impact.

For CoreWeave, this is a double-edged sword. Its asset-light model—primarily leasing data centers from Equinix, Digital Realty, and others—means that new capacity coming online is directly constrained by third-party supply chain bottlenecks. Conversely, its $66.8 billion in signed contract value is locked in; even if power shortages intensify, the contractual value remains unchanged, which actually enhances the scarcity of existing contracts.

For Nebius, the impact is two-fold. Nebius currently operates data centers across North America, Europe, and the Middle East; in an environment of worsening power shortages, this online capacity becomes a scarce asset, which is a positive. However, to scale from its current size to its target of 5GW by the end of 2030, Nebius will also have to wait for transformers and join the queue. In other words, the power bottleneck is a short-term tailwind but a long-term challenge for Nebius.

For IREN, this represents both the greatest upside and the greatest risk. Its core advantage lies in its secured land and grid interconnection rights, plus renewable energy assets for self-generation—theoretically allowing part of its capacity to avoid complete reliance on the third-party transformer supply chain. However, because IREN's AI cloud business is not yet mature, this advantage currently serves only one client, Microsoft, making it difficult to leverage into bargaining power for more commercial contracts.

This exogenous variable—the power supply chain bottleneck—is reshuffling the internal competitive landscape of Neocloud. Players who already control power resources will pull ahead, while those still waiting for transformer deliveries will face execution risks.

6. Four Evaluative Questions and Five Key Indicators: A Reusable Neocloud Analytical Framework

Having grasped the structural changes, the differentiated positioning of the three companies, and the impact of power bottlenecks, investors need more than a simple "buy/sell" conclusion; they need a reusable judgment framework. The following four often-overlooked market questions constitute the core logic for Neocloud portfolio allocation.

Question 1: Are you betting on training or inference? Training is a one-time R&D expenditure, while inference is the continuous consumption of computing power once a model is deployed. If investors agree with the structural trend of inference surpassing training, the primary beneficiaries will be inference leaders rather than training leaders. Examining by this standard: CoreWeave's customer base leans toward training; Nebius's Token Factory is a pure inference product; and IREN is an infrastructure generalist with relatively shallow software depth.

Question 2: Do you believe "usage-based billing" will become the new foundation for SaaS valuation? If investors subscribe to this narrative, it implies that Neocloud's mid-sized enterprise customer base will see exponential growth—according to Deloitte's Tokenomics report, mid-sized token workloads represent the segment where Neocloud holds the greatest cost advantage over hyperscale cloud providers. In this scenario, players like Nebius—with "complete software stacks and relatively diversified customer bases"—stand to benefit most. Conversely, if investors believe hyperscalers will ultimately dominate and Neocloud will merely serve to fill overflow demand, then CoreWeave, with its ties to the four major AI labs, offers greater certainty.

Question 3: How much downside risk can you tolerate? The risk profiles of these three companies are distinct: CoreWeave is "high leverage + high contract certainty" ($21 billion in total debt, $34 billion in off-balance-sheet leases, but a $66.8 billion contract backlog with blue-chip giant clients); Nebius is "low leverage + continuous equity dilution" (relatively healthy financial structure, but ongoing ATM offerings dilute existing shareholders); IREN is "lowest leverage + highest customer concentration" (the healthiest capital structure of the three, but Microsoft alone accounts for 57% of its AI cloud ARR target). There is no absolute hierarchy among these three risk profiles; the key is whether they align with an investor's individual risk appetite.

Question 4: What are the true litmus test indicators for this earnings season? The answer to this question directly forms an actionable watchlist.

• The first set of indicators targets CoreWeave, to be released after the market close on May 7, 2026: whether the quarterly growth rate of the contract backlog exceeds 50% (market expectations are 35%–40%); whether unit economics turn positive for the first time; and whether OpenAI's compute delivery schedule remains on track. Failure to meet any of these three indicators will fuel the bearish narrative.
• The second set of indicators targets Nebius, to be released before the market open on May 13, 2026: whether there is a significant jump in the number of Token Factory customers and its share of ARR. This single figure carries the credibility of the entire inference usage-based billing narrative.
• The third set of indicators targets IREN, to be tracked in subsequent earnings reports: when AI cloud revenue will structurally surpass Bitcoin mining revenue. Until this threshold is reached, it will be difficult for the market to treat it as a pure-play Neocloud peer.

7. Risk Warnings and the Boundaries of the Investment Framework

Finally, several important boundaries must be clarified. First, all three Neocloud companies are currently in the eye of the earnings season storm; any earnings miss could trigger 15% to 25% single-day stock price swings. The analytical framework provided in this article is not a guide for short-term entry points but a tool for constructing allocation logic over the next 6 to 12 months. For short-term traders, the marginal utility of this framework is limited, but for long-term investors, earnings-induced pullbacks may instead represent buying opportunities. Second, this framework is not a definitive answer. Every investor’s risk tolerance, capital size, and overall portfolio allocation differ—using the same four-question framework, different investors may arrive at vastly different allocation outcomes; this is a normal occurrence rather than a failure of the framework. Third, all price targets, contract amounts, and earnings guidance in this article are based on the latest data from early May 2026. Market conditions and company fundamentals change over time; readers encountering this article several months later should independently verify the latest progress.