The profitability of a neocloud—a specialized, GPU-first cloud service provider—is fundamentally a race against two competing clocks. The first clock begins ticking the moment a Graphics Processing Unit (GPU) ships from the manufacturer, as the hardware immediately begins its steep depreciation curve. The second clock only starts when that hardware is energized and begins generating rental income. The period between these two events represents pure capital loss, and the efficiency with which a provider can bridge this gap determines the ultimate survival of the business.
As the demand for Artificial Intelligence (AI) compute reaches unprecedented levels, the business model of "GPU-as-a-Service" (GaaS) has come under intense scrutiny. While the "napkin math" used by many startups suggests high margins and rapid returns, the reality of the market is far more complex. Modern deployments, particularly those involving NVIDIA’s H100 and Blackwell architectures, face razor-thin net margins when modeled honestly. However, a significant divergence has emerged in the industry: Bitcoin miners, leveraging existing industrial infrastructure, are finding a structural advantage that greenfield data center developers struggle to match.
The Neocloud Business Model and the Core Economic Equation
A neocloud is a lean cloud provider that eschews the broad product suites of hyperscalers like Amazon Web Services (AWS) or Microsoft Azure to focus exclusively on high-performance compute for AI workloads. The economics of this model can be reduced to a single, critical equation: Total Net Profit equals (Rental Rate × Utilization × Hours) minus (Power + Maintenance + Fixed Operating Costs + Depreciation + Financing).
While operators frequently highlight gross margins in the 55% to 65% range—calculated as revenue minus direct power and maintenance—this figure is often misleading. The metric that dictates long-term viability is the net margin after accounting for the massive depreciation of silicon and the interest on the debt used to acquire it. In many cases, these costs push neoclouds into the red.
Utilization is the primary lever in this equation. For debt-financed builds, utilization is often effectively 100%, as the capacity is pre-sold through long-term offtake contracts. However, for self-funded operators or those placing hardware on open marketplaces to capture spot demand, the break-even utilization rate typically hovers around 70%. For these operators, the difference between a 50% utilized cluster and an 80% utilized cluster is the difference between losing six figures a month and achieving profitability.
The Volatility of GPU Rental Rates: 2023 to 2026
One of the most common errors in neocloud financial modeling is the assumption that rental rates are stable or follow a predictable, linear decline. Historical data from the current AI cycle proves otherwise. Following the initial surge of interest in Large Language Models (LLMs), H100 rental rates plummeted by 64% to 75% from their 2023 peaks as supply chains stabilized and more capacity entered the market.
However, by 2026, the market experienced a sharp reversal. As deployments of the next-generation Blackwell chips faced logistical delays and existing Hopper (H100) capacity became fully committed, rental rates for H100s climbed by approximately 20% year-over-year. One-year contract rates saw even more dramatic movement, jumping nearly 40% from their late-2025 lows.
This volatility highlights a critical risk: the rental curve and the depreciation schedule are essentially the same bet. A GPU’s declining economic value is reflected as a falling rate on the revenue line and as depreciation on the cost line. Operators who assume a five-year depreciation life while also assuming flat rental rates are effectively "double-counting" optimism, leaving them vulnerable to market corrections.
The Depreciation Debate: Economic Life vs. Accounting Standards
The question of how long an AI GPU remains economically viable is a subject of intense debate among investors and operators. In late 2025, noted investor Michael Burry argued that the true economic life of an AI GPU is only two to three years, suggesting that major AI spenders were understating their depreciation obligations by as much as $176 billion over a three-year period.
In contrast, many large-scale operators and hyperscalers, including Meta, have moved to book GPUs over a five-to-six-year lifespan. The current shortage of Hopper chips has temporarily supported the longer-life argument, as resale values for H100 servers have recently moved in the opposite direction of what a short-life assumption would predict. In early 2026, secondary market prices for H100 servers traded significantly higher than they had six months prior, with "new-old-stock" hardware changing hands near its original MSRP.
Because depreciation is the largest non-cash expense on the balance sheet, the choice between a three-year and five-year schedule can fundamentally alter a company’s reported earnings, even if the cash flow remains the same.
The Miner’s Edge: Why Infrastructure Trumps Power Costs
A common narrative suggests that Bitcoin miners are winning the AI race because of access to cheap electricity. However, a detailed analysis of H100 unit economics suggests this is a misconception. At current rental rates, power consumption accounts for only 3% to 4% of total revenue. Consequently, even a significant reduction in electricity costs—such as moving from $0.08 per kWh to $0.04 per kWh—has a negligible impact on the bottom line.
The true "miner’s edge" lies in two areas: capital expenditure (CapEx) avoidance and speed to market.
1. Facility CapEx and Sunk Costs
The largest expense after the servers themselves is the facility infrastructure, including the utility interconnect, land, building shell, and primary electrical systems. A Bitcoin miner has already paid for these components. While converting a site from ASIC mining to GPU hosting requires significant upgrades—specifically in rack density, cooling, and power redundancy—the cost is a fraction of a greenfield build.
In a comparison of 1,000-GPU H100 SXM clusters, the all-in capital cost per GPU is estimated at:
- Greenfield Build: ~$50,000 per GPU
- Miner-Converted Site: ~$42,000 per GPU
The $8,000 difference per unit is entirely situated in the facility line. While networking, storage, and server costs remain identical for both parties, the miner carries significantly less facility capital to depreciate against revenue.
2. The Speed to Energization
Time is the most expensive variable in the neocloud race. A greenfield AI data center typically requires 18 to 30 months to move from planning to energization, largely due to multi-year queues for utility interconnects. A Bitcoin miner with an existing, energized grid connection can often deploy hardware in a matter of months.
In a market where hardware is sold out and rental rates are rising, being first to energize is a decisive advantage. This explains why firms like CoreWeave have signed multi-billion-dollar agreements to take over and convert existing mining sites. An old mining facility provides the one thing that cannot be bought at any price: time.
Technical Considerations: The Hidden Costs of Storage and Networking
Beyond the headline cost of GPUs, neocloud operators must account for high-performance networking and storage, which are often underestimated in preliminary models. For a 1,000-GPU training cluster, a high-performance parallel storage tier can cost several million dollars.
Networking costs, primarily driven by InfiniBand or high-end Ethernet fabrics required for multi-node training, add approximately $4,000 per GPU to the initial CapEx. Unlike the facility costs, these expenses do not offer a "miner’s discount," as the networking requirements for AI are fundamentally different from the simple internet connectivity required for Bitcoin mining.
Furthermore, the tension between mining and AI operations must be managed. Many miners rely on "curtailment" programs—turning off power during peak grid demand to earn revenue from utilities. However, AI customers demand 99.9% or 100% uptime. This means that while a miner can use their grid connection for AI, they must often forgo the demand-response revenue they previously relied on, or partition their site into "interruptible" and "always-on" zones.
Market Implications and Consolidation
The neocloud sector is moving toward a period of consolidation. As the "Hopper crunch" of 2026 eventually eases and more Blackwell-based capacity comes online, the market will likely split into two tiers.
The first tier will consist of providers who own their infrastructure and have low debt-to-equity ratios. These operators can weather periods of low utilization or falling rental rates because their fixed costs are minimized. The second tier will comprise "asset-light" providers who rent space from third-party data centers and lease their hardware. These providers are highly sensitive to market fluctuations and may struggle to survive if rental rates fall below the cost of their lease obligations.
The winners in this race will not necessarily be the ones with the most advanced software stacks, but rather those who started the race with the most assets already in the ground. The structural advantage of the Bitcoin miner—having an energized, high-scale grid connection—remains the most formidable moat in the AI compute industry.
As the industry matures, the "neocloud" label may fade, but the requirement for specialized, high-density compute will not. The operators who model their economics honestly, account for the volatility of silicon value, and leverage existing infrastructure will be the ones operating the backbone of the AI economy, while those waiting in the interconnect queue may find the market has moved on without them.
