The Best Argument for Sustainable Neocloud Role in the AI Ecosystem

Perhaps the “best” argument for a permanent role for neocloud service providers is the relevance of enterprise private cloud inference services, which arguably cost less than public cloud providers. 

While hyperscalers (AWS, Azure, Google Cloud) dominate general-purpose computing and model training, neoclouds (CoreWeave, Nebius, Hut 8 and others) have some advantages for inference operations, at least for the moment. 

How sustainable that advantage might be is the question. But the logic for long-term relevance generally rests on a few key arguments. 

Enterprises increasingly are wary of sending sensitive proprietary data (financial records, healthcare data, IP) to public multi-tenant clouds for inference.So neocloud providers can offer private services that , satisfy strict compliance (GDPR, HIPAA) and data sovereignty requirements that public clouds arguably struggle to guarantee.

The former bitcoin miners are essentially energy arbitrage firms. They have spent years securing the world's cheapest power contracts and building high-density cooling infrastructure. So they might have an advantage supporting services using high-performance graphics processing unitss (Nvidia H100s/Blackwell) at a lower marginal cost than hyperscalers.

Some believe neoclouds can undercut hyperscaler GPU hourly rates by 30 percent to 50 percent.

Neocloud providers also offer advantages in portability. Since neoclouds typically provide raw compute (Kubernetes/bare metal), enterprises can build portable inference engines. If a company wants to move their Llama 3 or Mistral inference workload from CoreWeave to Nebius to chase a lower price, they can do so easily.

Even if long-term evolution remains unclear, and even if questions about sustainability remain, virtually all forecasts project near-term revenue growth, though differing in the magnitude of those forecasts. 

Provider	2025 Est. Revenue / Run-Rate	2026 Forecast / Target	Key Enterprise Strategy
CoreWeave	~$8.0 Billion	$10B+ (High visibility from backlog)	The market leader. Deep partnership with Microsoft (serving as Azure's "overflow" valve) allows them to capture massive overflow inference demand.
Nebius Group	~$500–550 Million	$3.45B – $7.0B	Aggressive growth. Heavily focused on "AI Cloud 3.1"—a full-stack platform specifically targeting enterprise model deployment and inference.
Iris Energy	~$300 Million (AI specific)	~$3.4 Billion (Annualized AI Revenue target)	"Mega-deal" focus. Recently secured a $9.7B deal with Microsoft, validating the model of miners acting as sovereign infrastructure partners.
Hut 8	~$150–200 Million (HPC)	~$450 Million (Annualized NOI from new lease)	Infrastructure-first. Pivoted to a "landlord" model for AI, signing a $7B/15-year lease deal to host AI workloads, providing extreme stability.
Hive Digital	~$100 Million (Annualized Run-Rate)	~$140 Million (HPC Run-Rate Target)	The "Double Helix." Maintains a strong Bitcoin mining arm while upgrading legacy fleets to Tier 3 data centers for boutique enterprise inference.

CoreWeave arguably has the greatest degree of revenue visibility, though, based on its backlog, which might be as much as two orders of magnitude greater than other competitors. 

Provider	2024 Revenue	2025 Revenue (Forecast)	2026 Revenue (Forecast)	Notes [Sources]
CoreWeave	$1.92B	$5.0-5.1B zacks​	$10.9-14.9B spglobal​	$55B backlog; GPUaaS leader marketwise​
Lambda Labs	~$425M ARR (late)	$520M+ (FY Sep) finance.yahoo​	N/A	$500M ARR mid-2025; cloud GPU growth sacra+1
Crusoe Energy	$276M tsginvest​	$998M ainvest+1	~$2B tsginvest​	Stargate project; stranded gas power tsginvest​
Hut 8	N/A	~$650M (annualized) ainvest​	N/A	AI expansion; Q4 $162M quarterly ainvest​
HIVE Digital	N/A	N/A	BUZZ HPC $140M ARR hivedigitaltechnologies​	285% YoY Q2 growth hivedigitaltechnologies​
Vast Data	~$200M ARR (early)	N/A	$600M ARR calcalistech+1	AI storage; 5-7 yr contracts calcalistech​
Soluna Computing	N/A	N/A	N/A	Q3 2025 rev. up 37%; 100MW AI sites barchart+1

All that noted, perhaps the best rationale for a continued value chain role for the neoclouds is demand for private inference capabilities needed by enterprises.

Clear AI Productivity? Remember History: It Will Take Time

History is quite useful for many things. For example, when some argue that AI adoption still lags, that observation, even when accurate, ignores the general history of computing technology adoption, which is that it takes longer than most expect. 

Consider a widely-discussed MIT study that was also widely misinterpreted. Press reports said the study showed AI was not producing productivity gains at enterprises.

So all we really know is that pilot projects have not yet shown productivity gains at the whole-enterprise level. And how could they? 

Much has been made of a study suggesting 95 percent of enterprises deploying artificial intelligence are not seeing a return on investment.

There’s just one glaring problem: the report points out that just five percent of those entities have AI in a “production” stage. The rest are pilots or limited early deployments. 

That significant gap between AI experimentation and successful, large-scale deployment arguably explains most of the sensationalized claim that “only five percent of enterprises” are seeing return on AI investment. 

It would be much more accurate to say that most enterprises have not yet deployed AI at scale, and therefore we cannot yet ascertain potential impact. 

But that is not unusual for any important new computing technology. Adoption at scale takes time. 

Consider the adoption of personal computers, ignoring the early hobbyist phases prior to 1981, which would lengthen the adoption period. At best, 10-percent adoption happened in four years, but 50-percent adoption took 19 years. 

It took at least five years for the visual web to reach 10-percent adoption, and about a decade to reach 50-percent usage. 

For home broadband, using a very-conservative definition of “broadband,” (perhaps 1.5 Mbps up to perhaps 100 Mbps), it took seven years to reach half of U.S. homes.  

Technology	Commercial Start (Year)	Time to 10% Adoption	Time to 50% Adoption	The "Lag" Context
Personal Computer	1981 (IBM PC launch)	~4 Years (1985)	~19 Years (2000)	High Lag. Slowed by high cost ($1,500+), lack of connectivity (pre-internet), and steep learning curve (DOS/early Windows).
Internet	1991 (WWW available)	~5 Years (1996)	~10 Years (2001)	Medium Lag. Required physical infrastructure (cables/modems) and ISP subscription growth. "Network effects" accelerated it rapidly in the late 90s.
Broadband	~2000 (Cable/DSL)	~2 Years (2002)	~7 Years (2007)	Medium Lag. Replaced dial-up. Dependent on telecom providers upgrading last-mile infrastructure to homes.
Smartphone	2007 (iPhone launch)	~2 Years (2009)	~5-6 Years (2012-13)	Low Lag. Piggybacked on existing cellular networks. High replacement rate of mobile phones accelerated hardware turnover.
Tablet	2010 (iPad launch)	~2 Years (2012)	~5 Years (2015)	Low Lag. Benefited from the "post-PC" era ecosystem. Familiar interface (iOS/Android) meant zero learning curve for smartphone users.
Generative AI	2022 (ChatGPT launch)	< 1 Year (2023)	~2-3 Years (Proj. 2025)*	Near-Zero Lag. Instant global distribution via browser/app. "Freemium" models removed cost barriers. Adoption is currently outpacing the smartphone and internet.

The point is that widespread adoption of any popular and important consumer computing technology does take longer than we generally imagine. 

AI adoption is only at the very early stages. It will take some time for workflows to be redesigned; apps to be created and redesigned and user behavior to start to match the new capabilities. 

It is unreasonable to expect widespread evidence of productivity benefits so soon after introduction, even if new technologies now seemingly are adopted at a faster rate than prior innovations.