Neocloud Infrastructure
The Race to Power AI's Compute Needs — emerging GPU cloud providers building at scale outside the hyperscaler giants.
CapEx (2025)
Backlog
TAM by 2029
Queue
Executive Summary
The Investment Thesis
Neoclouds have emerged as the critical infrastructure layer for the AI revolution. These specialized GPU cloud providers are capturing massive hyperscaler contracts as demand for AI compute dramatically outstrips traditional cloud supply. The combined contracted revenue backlog across the five major players profiled in this report nears $100 billion—providing multi-year visibility rare in technology growth equities.1
The structural thesis is straightforward: Microsoft, Google, Amazon, and Meta are spending $380+ billion on AI infrastructure in 2025 alone (up 60%+ year-over-year),2 but they cannot build data center capacity fast enough. Grid interconnection queues now extend to 7 years in key markets.3 IDC projects global AI infrastructure spending will reach $758 billion by 2029, with accelerated servers accounting for over 94% of total market spending.4 This has created a 3-5 year window for specialist providers who control "AI-ready power" to capture outsized share of this buildout.
Yet the sector's capital intensity—characterized by GPU-collateralized debt structures, aggressive capacity buildouts, and concentrated customer bases—creates a distinctive risk profile. CoreWeave's debt-to-equity ratio exceeds 7x (total liabilities of $29.03B / equity of $3.88B = 7.49x); IREN is attempting a 75x GPU deployment expansion in under two years; and several players derive 60-70% of revenue from a single customer.5 The final margins on these hyperscaler deals remain to be seen, as contract ramps are ongoing and full operating costs materialize.
This report provides the educational foundation and comparative analysis needed to evaluate this emerging asset class.
What Are Neoclouds?
Neoclouds are a new class of cloud infrastructure providers focused on GPU-based computing for AI. Unlike traditional hyperscalers (AWS, Azure, GCP) that offer a broad palette of IT services, neocloud companies specialize in delivering massive GPU compute capacity on demand, often with bare-metal performance tailored specifically to AI workloads. They have evolved from niche origins—some began as cryptocurrency miners repurposing their data centers, while others spun out of tech companies—to become dedicated AI cloud platforms.
In essence, a neocloud is "an AI hyperscaler" without the legacy baggage of general-purpose clouds, singularly focused on serving the exploding needs of AI model training and inference.
Why Neoclouds Emerged Now
The tipping point? Generative AI's explosive rise since late 2022, when tools like ChatGPT ignited a surge in GPU demand that traditional hyperscalers couldn't match. Over the past two years (2023-2025), the need for GPUs to train and run AI models has vastly outstripped hyperscalers' capacity and flexibility—leading to months-long waitlists and strained availability. While the big clouds (AWS, Azure, GCP) do offer GPU instances, they're often pricey and limited, forcing AI labs and enterprises to seek alternatives. For instance, as of November 2025, renting an NVIDIA A100 40GB GPU on CoreWeave costs approximately $1.39/hour—versus $3.67/hour on Azure or Google Cloud—a 62% cost advantage for equivalent performance.6
Moreover, hyperscalers have to balance GPUs among many internal and customer uses, leading to waitlists and multi-tenant overhead. Neoclouds seized this opportunity by building dedicated GPU fleets that clients can tap with less friction and lower cost.
Three Structural Drivers
- AI-Driven GPU Bottlenecks: The surge in AI model sizes and usage created a GPU shortage in the market. Established clouds couldn't keep up with skyrocketing GPU requests, so big AI labs turned to alternative providers. Microsoft signed multi-billion-dollar agreements with CoreWeave and others to secure extra GPU capacity for OpenAI. Google backed deals with miners like Cipher to access more GPUs off its own cloud. Neoclouds became the pressure valve for this supply crunch.
- Power and Infrastructure Constraints: Deploying thousands of high-end GPUs isn't just about buying chips—it requires huge power and cooling capacity. Many neocloud entrants had an edge here: Bitcoin miners already operated large-scale data centers with cheap power contracts (often in remote areas with excess energy). As crypto mining economics faltered, firms like IREN and Cipher redeployed their hundreds of megawatts of power infrastructure to host AI hardware. Hyperscalers, in contrast, face longer lead times to expand data center power for GPUs.
- Single-Minded Focus on AI: Traditional clouds must support a wide array of services and customers. Neoclouds differentiate by focusing exclusively on AI workloads—meaning they can optimize everything (network topology, storage, scheduling) for GPU-heavy computing. This lack of legacy overhead allows for specialization and efficiency. In some cases, Nvidia has even given these specialized players preferential access to GPUs to broaden its market beyond the big three clouds.
Business Models: Understanding the Value Chain
Neocloud companies generally pursue one of three primary business models in the AI compute value chain. Understanding these distinctions is critical for evaluating competitive positioning and margin profiles.
Model 1: GPU Cloud (Virtualized GPU Leasing)
This model is closest to a traditional cloud offering—the provider offers on-demand instances of GPUs by the hour (or longer-term reservations) via an online platform. The focus is on infrastructure-as-a-service: stand up large pools of Nvidia GPUs and rent them out in a flexible, multi-tenant cloud environment. Software offerings are minimal beyond basic provisioning; customers typically install and run their own AI frameworks.
CoreWeave is the prime example: originally an Ethereum mining outfit, it pivoted to become the leading GPU cloud, providing everything from single GPU VMs to multi-GPU clusters on demand. Lambda Labs is another player, known for catering to AI researchers with on-demand GPU servers. These providers make money on utilization and scale—much like AWS EC2, but with a niche focus on GPU instances.
Model 2: Bare Metal + Power ("GPU Landlord")
In this model, the neocloud acts essentially as a landlord of powered space for GPUs. The company provides the physical infrastructure—real estate, power connections, cooling, racks—optimized for high-density GPU servers, and the client (often a large tech company or AI lab) either brings their own GPU hardware or exclusively uses hardware installed for them. This is a colocation or dedicated hosting model rather than a multi-tenant cloud.
IREN (formerly Iris Energy) exemplifies this approach. Originally a Bitcoin miner with massive renewable power sites, IREN pivoted to leverage its 3 GW of secured power portfolio for AI infrastructure. Its $9.7 billion Microsoft contract provides dedicated GPU capacity rather than shared cloud instances.7 Cipher Mining (CIFR) operates similarly, with 10-year hosting agreements backed by Google guarantees.
These bare-metal providers deal in megawatts instead of virtual machines—their success depends on securing cheap power and finance, then signing big tenants to long-term leases. Margins may be thinner per MW than on a pay-per-use cloud, but the revenue is more predictable via contracts.
Model 3: Full-Stack AI Infrastructure (GPU + Software Platform)
These players go beyond just providing raw compute—they build a vertically integrated platform combining custom hardware setups with software orchestration layers, managed services, and sometimes unique purchasing models (like token-based consumption or marketplaces for AI models). The idea is to differentiate through software and capture more of the AI value chain, not just commodity GPU time.
Nebius (NBIS) is a prominent public example among AI-focused neoclouds. Spun out of Yandex, Nebius brands itself as a "full-stack AI cloud platform" built for intensive AI workloads. It designs proprietary server chassis and networking in-house to optimize performance. Nebius is moving up the stack with products like Token Factory (inference-as-a-service launched November 2025, supporting 60+ models with claimed 70% cost/latency improvements) and Aether (AI Cloud 3.0, October 2025), which adds enterprise-grade security certifications (SOC 2 Type II, HIPAA, ISO 27001) required for regulated industries like healthcare and financial services.8
Full-stack providers aim to generate revenue not just from renting GPUs, but also from value-added services (managed model hosting, AI toolkits) and ecosystem effects. Without software differentiation, margins may compress toward utility-like levels as GPU capacity commoditizes.
Key Metrics and Unit Economics
Investors analyzing neocloud businesses look at a set of capacity and utilization metrics that might feel more at home in the energy or data center industry than traditional software. Understanding these metrics is essential for evaluating these businesses.
Connected Capacity (MW)
Neoclouds measure their infrastructure size in megawatts of power capacity. As Nebius CRO Marc Boroditsky explains:9
"Contracted is that we actually have a contracted commitment by a supplier to provide us with the energy capacity. Connected is it's in production—we actually have the energy being delivered."
Connected power is live and ready for GPU deployment. For example, CoreWeave reported approximately 590 MW of connected power as of Q3 2025, while Nebius expects ~220 MW by end of 2025. This metric is crucial as a lead indicator: more MW connected means the company can deploy more GPUs rapidly as demand comes in.
Contracted Capacity (Backlog)
Contracted capacity refers to future power under agreement but not yet delivered—essentially the pipeline. For example, CoreWeave has contracted approximately 2.9 GW, Nebius over 2.5 GW, and IREN has a 3 GW grid-connected power portfolio.9 Contracted MW is important because it de-risks the build-out—it's a guarantee of future capacity and cash flow. Investors prefer companies with a high portion of their expansion pre-sold to credible customers, but the gap between contracted and connected capacity also signals execution risk.
Contracted Power Capacity
Based on Q3 2025 filings and company announcements
Revenue per GPU-hour / per MW
This is a unit economic measure of how effectively the company monetizes its hardware. On cloud-like offerings, it might be expressed as revenue per GPU-hour. On a capacity basis, analysts often look at annualized revenue per MW of active capacity. There is a wide range depending on the business model: a pure on-demand cloud with high utilization can generate $9–11 million per active MW in a fully ramped scenario, whereas fixed-rate colocation contracts yield $3.7–8.6 million per MW-year.
Margins & Utilization Sensitivity
McKinsey argues that the bare-metal-as-a-service (BMaaS) model may have fragile economics. Their analysis suggests gross margins are typically 55-65% before depreciation, but after labor, power, and depreciation, gross profit margins drop to just 14-16%—leaving limited margin of safety.10 In this view, if GPU rental prices decline modestly or utilization slips below 80%, returns could flatline. The economics may become even more fragile when debt financing is factored in, as interest costs can erode any residual cushion.
GPU Useful Life & Depreciation
A critical but debated metric is how long GPUs remain economically useful. Companies typically depreciate GPU hardware over 4-6 years for accounting purposes—hyperscalers have extended server depreciation to 6 years. But with NVIDIA releasing new architectures every 1-2 years (Blackwell offers 4-5x faster inference than H100), some argue this overstates useful life and inflates near-term profits. Others counter that older GPUs retain value for less demanding workloads or secondary markets. The jury is still out—investors should understand that depreciation assumptions directly impact reported profitability and scrutinize them accordingly.
Comparative Analysis: Public Neocloud Players
The following comparison highlights the major public neocloud players as of November 2025. Note that data may shift rapidly given the pace of deal announcements and capacity buildouts.
Contract Backlog
Q3 2025 Revenue
Revenue Growth YoY
Debt-to-Equity Ratio
Detailed metrics breakdown:
| Metric | CRWV | NBIS | IREN | CIFR |
|---|---|---|---|---|
| Market Cap | ~$38B | ~$23B | ~$14B | ~$7B |
| Q3 2025 Revenue | $1.36B | $146M | $240M* | $72M |
| Revenue Growth YoY | +133% | +355% | +342% | +198% |
| Contract Backlog | $55.6B | $20.4B+ | $9.7B+ | $8.5B |
| Power Pipeline | ~2 GW | ~2.5 GW | 2.9 GW | 3.2 GW |
| Total Debt | $18.8B | $4.5B | $965M | $1.0B |
| Total Equity | $3.9B | $4.8B | $2.9B | $0.8B |
| Debt-to-Equity | 4.8x | 0.9x | 0.3x | 1.3x |
| Business Model | GPU Cloud | Full-Stack | Bare Metal | Bare Metal |
| Key Customer | MSFT (62%) | MSFT, META | MSFT | GOOG, AMZN |
*IREN figure represents Q1 FY26 (fiscal year ends June; calendar Q3 2025). Data as of November 2025. Source: Company filings and earnings releases.11
Capital Structure and Financing
Neocloud infrastructure is extremely capital-intensive. Building data centers, acquiring thousands of high-end GPUs, and securing power contracts require huge upfront investments. Unlike pure software companies, these businesses look more like industrial projects in terms of cash flows. As a result, their financing strategies have to be creative and robust.
Heavy Use of Debt and Leases
Many neocloud players are financing growth by taking on significant debt, equipment leases, or other forms of non-dilutive capital. CoreWeave pioneered GPU-collateralized debt—an innovative move essentially treating cutting-edge chips like assets to borrow against. In 2024, CoreWeave closed a $7.5 billion debt facility led by Blackstone, using its massive GPU inventory and contracts as collateral. Total debt now exceeds $18 billion.12
Similarly, IREN obtained leases for its GPU purchases—it structured a 36-month lease to finance 100% of the cost for a batch of NVIDIA Blackwell GPUs, at a high-single-digit interest rate.13 This is akin to a sale-leaseback: the vendor or a financing partner pays for the equipment upfront, and the neocloud pays them back over 3 years while using the GPUs to generate revenue.
Equity and Convertible Notes
Equity financing remains important, especially as a cushion for all the debt. Nebius went public to access equity markets and recently completed a $4.2 billion mega-raise combining equity and convertible notes. The convertibles carry low coupons (1-3%) with high conversion premiums (40-50%), minimizing near-term dilution while providing growth capital.14
There are also structured equity deals: Google provided a $1.4 billion payment guarantee for Cipher's Fluidstack contract and received warrants (~5.4% ownership) in return.15 This kind of strategic equity infusion tied to a contract aligns a big tech investor with the neocloud's success. IREN filed for up to $1 billion in convertible notes to fund its AI expansion, which would turn into equity if their stock rises sufficiently.
Power Access as a Moat
Financing alone isn't enough—securing power and grid access is critical, and in many ways is part of the "capital." Having a 100MW substation already built at a site is a tangible competitive advantage. This is why miners with existing infrastructure are valued—they spent years (and lots of money) building out power capacity. Now that capacity can be redirected to AI at much higher economic yield.
Investors should note which players have low-cost, long-term power contracts (ideally renewable or fixed-price power, which gives predictability). IREN's sites are primarily backed by renewable energy (hydro in BC, wind/solar in Texas) often at very low cents per kWh, which could translate to higher margins on AI compute sold.16 Control over power also means the ability to scale: many neoclouds trumpet their GW-scale pipelines (Cipher 3.2GW, IREN 2.9GW, Crusoe 45GW pipeline).
Leverage and Risk Management
High leverage is a double-edged sword. It allows rapid expansion, but it means these companies carry significant fixed obligations. Many have contracted revenue that gives lenders comfort—for instance, Nebius financing its Microsoft deal partly by issuing debt secured by that contract's cash flows. This kind of project financing is becoming common: each big contract can be financed by loans paid back from that contract's revenue, isolating risk.
The reliance on leverage also means interest rates matter—most deals we've seen are at relatively high rates, reflecting lenders pricing in risk. Neoclouds must manage the risk that if AI demand falls or a contract gets canceled, they still owe money on those GPUs or facilities. Hence, we often see hedging strategies like having customer prepayments (Microsoft and others often pay some upfront) or staggered build-outs so they don't overshoot.
Capital Structure Comparison
| Company | Total Debt | Cash & ST Inv | Net Debt | D/E |
|---|---|---|---|---|
| CoreWeave | $18.8B | $1.4B | $17.4B | 4.8x |
| Nebius | $4.5B | $4.8B | -$0.3B | 0.9x |
| IREN | $965M | $1.0B | ~$0 | 0.3x |
| Cipher | $1.0B | $1.2B | -$0.2B | 1.3x |
The capital intensity is a barrier to entry for small players—raising billions is no small feat—so the ones who do have access to capital (and manage it wisely) will race ahead. Investors will want to track each player's burn rate vs. contracted revenue and their balance sheet health.
Company Deep Dives
CoreWeave's March 2025 IPO at $40/share has seen dramatic volatility, with the stock reaching $187 before correcting to the current $71-74 range. The company's Q3 2025 results demonstrated execution: revenue of $1.36 billion (+133% YoY) beat expectations, and the contracted backlog nearly doubled to $55.6 billion, providing roughly 10 years of visibility at current run rates.
The company's most distinctive feature is its financing architecture. CoreWeave pioneered GPU-collateralized debt, using Nvidia H100s and successor GPUs as collateral for facilities totaling $18.8 billion in debt. Special purpose vehicles (SPVs) ring-fence specific customer contracts like the $18.4 billion OpenAI relationship.
Customer concentration is the primary risk factor. Microsoft accounted for 62% of 2024 revenue and 71% in Q2 2025, though management expects this to decline below 50% as OpenAI and Meta contracts ramp. Interest expense tripled YoY to $311 million in Q3 2025, and off-balance sheet lease commitments total $34 billion through 2028.
View Company Profile
Nebius, spun off from Yandex in 2024 following the company's complete divestiture of Russian operations, has rapidly established itself as a credible neocloud competitor. The stock has appreciated +200% YTD and currently trades at ~$88-90 with a ~$22-24 billion market cap.
The company secured transformational contracts in Q3-Q4 2025: a $17.4-19.4 billion Microsoft deal over five years and a $3 billion Meta contract to support Llama model training. Combined with aggressive capacity expansion, management raised 2026 ARR guidance to a striking $7-9 billion.
Nebius's balance sheet is notably stronger than CoreWeave's. Following a September 2025 mega-raise of $4.2 billion (combining equity and convertible notes), the company holds an estimated $6+ billion in cash against ~$4 billion in convertible debt structured with high conversion premiums (40-50%) to minimize near-term dilution.
Nebius is moving up the stack with products like Token Factory (inference-as-a-service launched November 2025, supporting 60+ models with claimed 70% cost/latency improvements) and Aether (AI Cloud 3.0, October 2025), which adds enterprise-grade security certifications (SOC 2 Type II, HIPAA, ISO 27001) required for regulated industries like healthcare and financial services.
View Company Profile
IREN (formerly Iris Energy) has executed the most dramatic pivot in the sector, securing a $9.7 billion, 5-year Microsoft contract announced November 2025. The deal provides Microsoft with 200MW of critical IT load via liquid-cooled data centers at the Childress, Texas campus, deploying next-generation Nvidia GB300 GPUs.17
Financial results reflect the transition: Q1 FY26 revenue (calendar Q3 2025, reported November 6) reached $240.3 million (+355% YoY), though 97% still derives from Bitcoin mining (AI cloud just $7.3M). The company targets AI Cloud ARR of $3.4 billion by end of calendar 2026, contingent on deploying 140,000 GPUs.18
IREN's infrastructure advantage lies in its 2.9GW of fully contracted power across West Texas facilities. The balance sheet supports expansion: $1.8 billion cash following a zero-coupon $1 billion convertible raise, against ~$965 million debt for a 0.3x debt-to-equity ratio—far healthier than CoreWeave's leverage.
View Company Profile
Cipher Mining has secured $8.5 billion in AI/HPC contract value through deals with Google-backed Fluidstack (~$3B) and AWS (~$5.5B), transforming from a pure Bitcoin miner into a data center infrastructure provider.15
The Google relationship is particularly notable: Google provided a $1.4 billion payment guarantee backstopping Fluidstack's lease obligations and acquired a ~5.4% equity stake via warrants.15 The structure de-risks the 168MW Barber Lake development while providing Cipher with investment-grade credit support.
Unlike IREN's Microsoft deal, Cipher has not deployed GPUs directly—the contracts are colocation/hosting agreements where customers bring their own compute. Q3 2025 revenue of $72 million remains 100% Bitcoin mining; AI/HPC revenue commences August 2026 when AWS rent begins.
View Company ProfileCrusoe Energy represents the most differentiated model in the sector, having pivoted from its origins using flared natural gas to power Bitcoin mining into a pure-play AI infrastructure company with a claimed 45GW energy pipeline—roughly 10x the total capacity of Northern Virginia, the world's largest data center market.19
The October 2025 Series E round valued Crusoe at $10+ billion (up from $2.8B in December 2024), raising $1.375 billion from investors including Nvidia, Founders Fund, Fidelity, T. Rowe Price, and Mubadala.
Crusoe's flagship project is the OpenAI Abilene campus, part of the $500 billion "Stargate" infrastructure initiative. Phase 1 (200MW, two buildings) went live in September 2025—constructed in just one year versus the 3-4 year industry norm. The full campus is designed for up to 400,000 GPUs.
Key Risks
While the neocloud sector is booming, investors should be mindful of several risks and uncertainties that could impact these businesses. The sector's capital intensity and rapid growth create unique vulnerabilities.
CoreWeave's top three customers likely represent 70%+ of revenue; IREN and CIFR depend on single contracts (Microsoft, Google/AWS respectively) for their AI infrastructure thesis. Contract cancellation clauses tied to delivery schedules create execution risk. If a major customer delays, renegotiates, or cancels, the financial impact could be severe given fixed cost structures.
CoreWeave's 4.8x debt-to-equity and GPU-collateralized debt model is unprecedented in tech. A demand slowdown or GPU depreciation cycle could stress covenant compliance. Interest expense already tripled YoY to $311M in Q3 2025. Off-balance sheet lease commitments of $34B through 2028 add hidden leverage. Nebius and IREN maintain healthier leverage profiles, but rapid expansion could strain even conservative balance sheets.
IREN's 140,000 GPU deployment target by end-2026 represents a 75x expansion from current capacity in under two years. Crusoe's 45GW pipeline would require construction velocity never before achieved in the data center industry. History suggests large infrastructure projects routinely face delays, cost overruns, and supply chain bottlenecks. CoreWeave's Q3 guidance cut stemmed from exactly this—a third-party developer falling behind schedule.
Nvidia's Blackwell architecture is ramping, and future generations will follow. H100 values may decline faster than depreciation schedules assume, and providers with older GPU fleets could face margin compression. Companies depreciating GPUs over 6 years (like CoreWeave) may be understating the true economic cost if hardware becomes obsolete in 3-4 years. Custom silicon from hyperscalers (Google TPU, AWS Trainium) could also erode the Nvidia-centric value proposition.
Grid interconnection queues extend to 7 years in key markets. Even with "secured" power pipelines, energizing new capacity requires permits, transformer deliveries, and utility cooperation—all potential bottlenecks. Power costs can also be volatile; a spike in electricity prices directly compresses margins given the energy intensity of GPU operations.
H100 rental rates have declined 60-75% from peak ($8-10/hr to $2-4/hr). As more capacity comes online and Nvidia increases supply, further price compression is likely. Pure infrastructure players without software differentiation may see margins erode toward utility-like levels. The window for premium pricing may be closing faster than backlog would suggest.
Microsoft, Google, and Amazon are simultaneously customers and potential competitors. They're investing in custom AI chips and building their own capacity. The current outsourcing wave could reverse if hyperscalers decide to pull compute in-house once their own buildouts catch up. Strategic contracts today may not renew at the same terms tomorrow.
The entire thesis depends on sustained AI compute demand growth. If AI adoption slows, enterprise budgets tighten, or the AI investment cycle pauses (as happened with crypto), utilization rates could collapse. Unlike software with near-zero marginal costs, idle GPUs still incur depreciation, power, and financing costs. A recession or AI "winter" would hit these capital-intensive businesses harder than asset-light tech.
Catalysts & Timeline
The following calendar highlights key events and milestones that could drive sector performance through 2026:
Putting It All Together
The neocloud sector sits at the intersection of three structural tailwinds: hyperscaler AI capex acceleration, grid interconnection bottlenecks, and GPU supply constraints. The $100+ billion in contracted backlogs across major players provides unprecedented revenue visibility for growth equities, but the capital intensity and customer concentration create a risk profile more akin to infrastructure than software.
Business Model Divergence
CoreWeave (GPU Cloud)
Highest revenue and scale, but also highest leverage (7.5x D/E) and customer concentration (62% Microsoft). The GPU-collateralized debt model is innovative but untested through a downcycle. Stock volatility reflects this tension—trading at premium multiples when growth accelerates, correcting sharply on execution concerns.
Nebius (Full-Stack)
Cleaner balance sheet (0.7x D/E) and software differentiation attempt via Token Factory. The Meta + Microsoft contract diversification reduces concentration risk. Higher revenue per GPU potential if software layer gains traction, though execution remains to be proven at scale. Software differentiation could provide margin protection as GPU capacity commoditizes.
IREN (Bare Metal/Power)
Purest play on "power as moat" thesis. 2.9GW of renewable power contracts in Texas provide genuine competitive advantage. 75x GPU expansion requirement creates massive execution risk, but Microsoft contract de-risks demand. Lowest leverage (0.5x D/E). If they execute, IREN captures the highest ROI on infrastructure assets.
CIFR (Colocation)
Most conservative model—customers bring their own GPUs, reducing capex burden. Google payment guarantee de-risks Fluidstack contract. Lower revenue potential per MW, but also lower operational risk. Stock trades at discount to peers reflecting later revenue timing (August 2026 AWS commencement).
What to Watch
Execution on Deployment Timelines: IREN's 140K GPU target by end-2026 and Crusoe's 45GW pipeline are both extraordinary commitments. Delays or cost overruns would be significant negative catalysts.
Customer Diversification: CoreWeave reducing Microsoft concentration below 50%, Nebius ramping Meta workloads, and CIFR signing additional hosting agreements beyond AWS would all reduce single-customer risk.
GPU Pricing Trends: H100 rental rates declining 60-75% from peak signals commoditization risk. Companies with software differentiation strategies or long-term fixed-price contracts (IREN, CIFR) may be better positioned, though software layer monetization remains unproven across the sector.
Leverage Management: CoreWeave's interest expense tripling YoY to $311M while carrying $34B in off-balance sheet leases creates refinancing risk if growth slows. Nebius and IREN have room to add leverage; CoreWeave has limited flexibility.
The Bottom Line
Neoclouds are capturing a 3-5 year window where hyperscaler buildout timelines can't match AI demand growth. The $100B+ in contracted backlogs is real, and the power infrastructure moat is genuine. But this is not a software business—it's capital-intensive infrastructure with execution risk, technology transition risk, and customer concentration risk.
The sector will likely bifurcate: winners who execute on buildouts, maintain customer relationships, and manage leverage will compound returns; laggards who miss timelines, face contract renegotiations, or over-lever will see equity wiped out. The difference between a CoreWeave and a potential bankruptcy is measured in quarters, not years.
For investors with appropriate risk tolerance, the sector offers asymmetric upside if AI infrastructure spending continues at current pace. The key is matching risk profile to capital structure and business model: each player presents distinct trade-offs between leverage, customer concentration, and execution risk.
- 1 Combined contract backlogs: CoreWeave Q3 2025 earnings ($55.6B), Nebius investor presentation ($20.4B+), IREN Microsoft contract announcement ($9.7B), Cipher AWS/Fluidstack contracts ($8.5B). CoreWeave Q3 2025; Nebius Q3 2025; IREN Contract; Cipher Q3 2025
- 2 2025 hyperscaler AI infrastructure spending ($380B+, up 60%+ YoY): CNBC, October 31, 2025
- 3 Grid interconnection queues (5-7 year median wait times): Lawrence Berkeley National Laboratory Queued Up Hub, August 2025 update
- 4 AI infrastructure market to reach $758B by 2029 (accelerated servers 94%+ of spending): IDC Worldwide AI Infrastructure Tracker, October 28, 2025
- 5 CoreWeave debt-to-equity ($29.03B liabilities / $3.88B equity = 7.49x): Q3 2025 10-Q. IREN 75x GPU expansion (1,900 to 140,000 target) and margin guidance: Q1 FY26 earnings, Nov 6, 2025. Customer concentration (>50% from 1-2 customers): McKinsey neocloud report, Nov 2025
- 6 A100 40GB GPU pricing comparison: CoreWeave ($1.39/hr) vs. Azure and Google Cloud ($3.67/hr), November 2025
- 7 IREN Microsoft $9.7B contract (3 GW power portfolio, dedicated GPU capacity): IREN investor presentation, November 2025
- 8 Nebius Token Factory (production AI inference at scale): Nebius newsroom, November 2025
- 9 Connected vs contracted definitions: Nebius CRO Marc Boroditsky interview. Power capacity figures: CoreWeave Q3 2025 (590 MW active, 2.9 GW contracted); Nebius Q3 2025 (220 MW connected by end-2025, >2.5 GW contracted); IREN March 2025 (2.75 GW secured in West Texas)
- 10 BMaaS margin economics (55-65% gross margin before depreciation, 14-16% after): McKinsey, "The evolution of neoclouds and their next moves", November 2025
- 11 Comparative data: Company 10-Q/10-K filings, earnings releases, and investor presentations as of November 2025. Financial statements: CRWV, NBIS, IREN, CIFR
- 12 CoreWeave $7.5B debt facility: Blackstone press release, May 17, 2024
- 13 IREN GPU lease financing ($102M, 36-month lease, high single digit rate): IREN press release, August 25, 2025
- 14 Nebius $4.2B capital raise (equity + convertible notes): Nebius press release, September 15, 2025
- 15 Cipher-Fluidstack deal ($1.4B Google backstop, ~5.4% equity via warrants, ~$3B contract): Cipher Mining press release, September 25, 2025
- 16 IREN power sources: Mackenzie (BC Hydro), Childress (low-cost renewable energy)
- 17 IREN Microsoft $9.7B contract ($5.8B Dell GPU purchase, 200MW critical IT load at Childress): IREN press release, November 3, 2025
- 18 IREN Q1 FY26 results ($240.3M revenue, $7.3M AI Cloud, $3.4B ARR target, 140k GPU expansion): IREN Q1 FY26 earnings, November 6, 2025
- 19 Crusoe Series E ($1.375B, $10B+ valuation, 45GW pipeline): Crusoe press release, October 24, 2025; OpenAI Stargate announcement
- 20 Crusoe Abilene campus timeline and capacity: Data Center Dynamics