How Long Will Neocloud Scarcity Last?

The commercialization of any era-defining computing technology typically features a race between suppliers: how long can new suppliers live off of scarcity before abundance eliminates the profit margins?

In other words, how long can business strategies based on scarcity flourish until abundance is reached, and how many of the early suppliers can make the jump from “commodity” product supply to some form of new value add?

That is one question investors have to ask themselves about the rise of third-party high-performance computing facilities operated by the likes of CoreWeave, Nebius and others. 

A related question for investors in artificial intelligence firms is whether the capital intensity of AI changes the traditional sources of value for computing products, which has tended to migrate “up the stack” over time. 

The capital intensity of the AI model training and inference operations might suggest there is a new and possibly somewhat lasting new “moat” value based on infrastructure ownership and control, which has tended not to be the case for most other computing innovations. 

This tension between scarcity-driven value and abundance-driven commoditization arguably has occurred in every new major wave of technology.

A new resource (compute, bandwidth, or power) is initially scarce, allowing early movers to capture "bottleneck rents."

However, incumbents and new entrants eventually respond by flooding the market with supply, destroying those rents and forcing value to migrate to a different layer of the stack.

Google (TPU), Amazon (Trainium), and Microsoft (Maia) have responded to Nvidia’s dominance by building their own silicon, starting to alleviate the physical scarcity of graphics processor untis for their internal workloads.

Open source has in the past functioned as a method of alleviating scarcity. The issue is how well this strategy will work in a capital-intensive AI industry. 

This has played out in past computing technology eras. 

In the late 1970s, computing was defined by integrated, proprietary systems like the Apple II or the Commodore PET. If you wanted the software, you had to buy specific hardware.

For a brief moment, hardware manufacturers enjoyed high margins because they controlled the entire "stack."

In 1981, IBM entered the market using off-the-shelf parts (Intel CPUs, Microsoft OS). That eliminated “scarcity” as other firms flooded the market with "IBM-compatible" PCs, turning the hardware into a low-margin commodity.

As the hardware became a commodity, the value migrated to the two components that remained proprietary and scarce: the Intel microprocessor and the Microsoft operating system (the "Wintel" monopoly).

In the era of cloud computing, server ownership ceased to be a barrier to compute hardware costs. As a practical matter, that meant software or app startups did not need to invest in their own compute facilities, but could rent them from AWS, for example. 

Tech Wave	Initial Scarcity (Rents)	Mechanism of Abundance
Electricity	Localized Generation	The Centralized Utility Grid
PC Era	The Integrated "Box"	Open Architecture & Clones
Internet	Bandwidth / Connectivity	Massive Fiber Overbuild
Cloud	Physical Server Racks	Virtualization & Hyper-scale
AI (Current)	GPU Compute / Tokens	Neoclouds & Model Distillation

The point is that, over time, scarcity is eliminated in the value chain, shifting value creation elsewhere in the product stack. 

So the issue is how neocloud providers such as CoreWeave and Nebius, which essentially have democratized access to high-performance clusters, will flourish going forward, as high-performance computing becomes less scarce and more commodity-like. 

Stage	Commodity Model (The "Plumbing")	Value-Added Model (The "Experience")
Example	Raw Token API, Basic GPU Hosting	AI Agents, Industrial Robotics, Sovereign AI
Revenue Driver	Volume and Scale (Low Margin)	Outcomes and Reliability (High Margin)
Competitive Edge	Lowest Price per FLOP	Trust, Performance, and Workflow Integration
Incumbent Threat	Massive CapEx from Hyperscalers	Regulation and “Reasoning” Moats

The rise of "Neoclouds" (specialized AI infrastructure providers like CoreWeave, Nebius, and Lambda Labs) challenged the assumption that only the "Big Three" hyperscalers (AWS, Azure, Google) owned the “AI as a service” infrastructure business. 

But history suggests the "compute moat" goes away over time, forcing suppliers to find new sources of value. 

Aspect	The "Old" Moat (Pre-2024)	The "New" Moat (Late 2025)
Primary Barrier	Ownership of H100 GPU Clusters.	Proprietary "Reasoning" and reinforcement learning
Strategy	Vertical Integration (Own the DC).	Architecture Efficiency (Train for less).
Infrastructure	Proprietary Cloud (Azure/AWS).	Multi-Cloud/Agnostic (Rent where available).
Value Capture	Selling Compute / Tokens.	Selling Outcomes / Agentic Actions

So investors must make decisions about where scarcity is in the value chain, how long that scarcity will last, and where value shifts within the value chain as a result.

Computing has Shifted from Work to Life and Now Begins to Augment Life

I think we generally miss something important when pondering how artificial intelligence will shift job functions from repetitive, lower-order tasks to higher-order cognitive tasks, even displacing many cognitive tasks, with consequent impact on jobs. 

Across three major computing eras: the personal computer era (roughly 1970s–1990s); the internet era (1990s–2010s) and the coming AI era (2010s–present), computing's pervasiveness has increased steadily.

Where we first used PCs to accelerate routine work tasks ("doing things faster"), we later used the internet to accelerate knowledge acquisition ("learning things faster") and then playing, shopping and traveling, while demolishing many geographic barriers.  

The shift was from “computing for work” to “computing for life.”

But AI should be even more pervasive, allowing us to optimize outcomes ("doing things better"), and shifting computing from intentional interactions to anticipatory (autonomous) action. So computing shifts from tool to “collaborator.” PCs and software were tools we used. In the AI era computing will augment and amplify human capabilities. 

To be sure, we might argue that all general-purpose technologies have augmented human senses or capabilities in some way (muscles, sight, hearing, cognitive tasks, speech, transport, staying warm or cool). 

So the movement is something like “work to life to existence.” Sure, we can still ponder what AI means for work, or life. But that likely underplays the impact on normally esoteric thinking about what humans do that is uniquely human. 

AI arguably can automate intermediate cognitive tasks such as basic data analysis, customer service responses and routine decision-making. So yes, AI will reshape work. 

Cognitive Task	Example Tasks	Current AI Capabilities	Extent of Automation
Data Processing and Analysis	Data entry, basic statistical analysis, report generation	AI excels at processing large datasets, generating insights, and creating reports (e.g., tools like Power BI, Tableau with AI plugins, or custom ML models).	High: Routine data tasks are fully or near-fully automated. Human oversight needed for validation and complex interpretation.
Pattern Recognition	Fraud detection, image classification, trend identification	AI uses machine learning (e.g., neural networks) to identify patterns in financial transactions, medical imaging, or market trends with high accuracy.	High: AI often outperforms humans in speed and scale, but human judgment is required for context or anomalies.
Basic Decision-Making	Customer service responses, inventory management, scheduling	AI-powered chatbots (e.g., Zendesk, Intercom) handle routine queries; algorithms optimize schedules or stock levels.	Moderate to High: Routine decisions are automated, but complex or ambiguous cases require human intervention.
Content Generation	Writing emails, creating marketing copy, summarizing texts	Generative AI (e.g., GPT models, Jasper) produces coherent text, summaries, or creative content based on prompts.	Moderate: AI generates drafts or suggestions, but human editing is needed for nuance, tone, or originality.
Diagnostic Tasks	Medical diagnostics, legal research, technical troubleshooting	AI assists in diagnosing diseases (e.g., IBM Watson, Google Health), analyzing legal documents, or identifying system errors.	Moderate: AI provides accurate recommendations, but final diagnoses or decisions require human expertise.
Predictive Modeling	Sales forecasting, risk assessment, customer behavior prediction	AI models (e.g., regression, deep learning) predict outcomes based on historical data with high precision.	High: Predictions are automated, but humans must interpret results and make strategic decisions.
Language Translation and Processing	Real-time translation, sentiment analysis, speech-to-text	AI tools (e.g., Google Translate, DeepL) provide near-human-quality translations and analyze sentiment in texts or speech.	High: Routine translations are nearly fully automated; human input needed for cultural nuances or specialized contexts.
Routine Problem-Solving	Technical support queries, basic coding, process optimization	AI resolves common IT issues, generates simple code (e.g., GitHub Copilot), or optimizes workflows.	Moderate: AI handles standard cases, but novel or complex problems require human creativity and reasoning.

But AI will affect not only work, but almost all other elements of human life. In the PC era computing automated and digitized work and personal projects.

In the Internet era computing enabled new forms of creativity, commerce, and community.

In the AI era we’ll see augmented human intelligence, senses, and capabilities.

Also, compared to the earlier impact of PCs and the internet, it is possible that AI will produce outcomes sooner than has been the case in the past. 

Where we might argue that PCs produced widespread change over a two-decade or three-decade period, where the internet arguably produced fundamental changes over a two-decade period,, some believe AI will achieve widespread change in as little as a decade. 

The IBM PC, for example, was released in 1981. It wasn’t until about 2000 that half of U.S. households owned a PC. 

In 1983, perhaps 10 percent of U.S. homes owned a PC and about 14 percent of those homes used a modem to connect using the internet, according to Pew Research. At that point, it was all-text bulletin boards and the visual browser and multimedia internet had not yet been invented. 

It was not until 2000 or so that half of U.S. consumers said they used the internet. 

Year	PC Adoption (%)	Internet Adoption (%)
1995	36	14
2000	51	41.5
2010	76.7	71
2016	89.3	87