The fundamental appeal of the neocloud model—specialized cloud providers that focus exclusively on high-performance compute for AI workloads—often relies on "napkin math" that suggests overwhelming profitability. However, detailed economic modeling reveals that for many greenfield deployments of flagship hardware like the Nvidia H100, margins are razor-thin. Interestingly, when the same hardware cluster is modeled from the perspective of an established Bitcoin miner, the economics become significantly more robust. This advantage is not merely a result of access to cheap electricity or discounted hardware; rather, it stems from the "sunk" infrastructure already in the ground and the speed with which a miner can bring that capacity online.
The Neocloud Equation and the Reality of Gross Margins
A neocloud is a streamlined cloud provider that has stripped away the ancillary services offered by hyperscalers like Amazon Web Services or Microsoft Azure to focus solely on renting compute power for AI training and inference. The business model can be reduced to a single, critical equation: total revenue equals the rental rate multiplied by utilization and total hours of operation, minus the costs of power, maintenance, fixed operating expenses, depreciation, and financing.
In industry circles, operators frequently highlight gross margins, which typically range between 55% and 65%. While these figures are impressive, they only account for revenue minus direct power and maintenance costs. The metric that dictates long-term survival is the net margin after depreciation and interest. It is in this area that many neoclouds struggle to remain solvent. The most significant lever in this equation is utilization. For debt-financed builds, which are often backed by offtake contracts, utilization is effectively 100% because the capacity is sold before the hardware is even racked. However, for self-funded operators who list GPUs on open marketplaces and assume demand risk, the break-even point typically hovers around 70% utilization. For these operators, the difference between a half-idle cluster and a fully utilized one is the difference between losing six figures monthly and achieving profitability.
Crucially, idle GPUs do not pause their depreciation. They lose value on a fixed schedule regardless of whether they are earning revenue. Sophisticated operators attempt to mitigate this by using automation to relist contracted but unused hours on the spot market at premium rates, yet the underlying economic decay of the hardware remains a constant pressure.
Historical Volatility and the Depreciation Debate
Accurately modeling a neocloud requires addressing two common errors in financial forecasting: the assumption of stable rental rates and the lack of consensus on depreciation schedules.
Historically, rental rates for high-end GPUs have been far from stable. Following the 2023 peak of the AI boom, H100 rental rates plummeted by 64% to 75% as global supply began to catch up with initial demand. This led many analysts to assume a permanent downward trajectory for rental pricing. However, by 2026, the market saw a dramatic reversal. Delays in the deployment of Nvidia’s Blackwell architecture, combined with a total sell-out of Hopper (H100/H200) capacity, caused H100 rental rates to climb approximately 20% year-over-year. One-year contract rates jumped nearly 40% from their late-2025 lows. These fluctuations prove that a model assuming a fixed decline is just as fragile as one assuming permanent growth. The rental curve and the depreciation schedule are essentially two sides of the same coin; an honest financial model must treat them as a singular, explicit bet on the asset’s economic life.
The second major point of contention is the estimated lifespan of the hardware. In late 2025, prominent investor Michael Burry argued that the economic life of an AI GPU is only two to three years, suggesting that major AI spenders were understating depreciation by as much as $176 billion for the period between 2026 and 2028. In contrast, many operators, including Meta, utilize a five-to-six-year depreciation schedule. The current shortage of Hopper chips has complicated this further, as resale values have moved in the opposite direction of what a short-life assumption would predict. Because the economic life of this hardware is so difficult to pin down, viable financial models must allow for toggling between two-, three-, and five-year lifespans to understand the potential range of outcomes.
The Bitcoin Miner’s Structural Edge
The prevailing narrative suggests that Bitcoin miners have an advantage in AI because of their access to low-cost power. While power is a factor, its impact is often overstated in the context of high-end GPU compute. At current H100 rental rates, electricity costs represent only about 3% to 4% of total revenue. A significant difference in power pricing—for instance, four cents per kilowatt-hour versus eight cents—does not fundamentally shift the profitability of the operation. Similarly, the "cheap hardware" advantage has largely vanished; as of 2026, secondary-market H100 servers trade at prices well above their levels from six months prior, and new-old-stock often sells at original MSRP.
The true, durable edge for Bitcoin miners lies in capital and time. The most expensive component of an AI data center, after the servers themselves, is the facility infrastructure. This includes the land, the building shell, the electrical build-out, cooling systems, and, most importantly, the utility interconnect. A Bitcoin miner has already paid for these elements. The interconnect is the single most difficult asset to acquire in the current AI landscape, often gated by utility queues that can last several years.
A miner transitioning to AI compute does not need to build from scratch. Instead, they fund the "delta"—the cost of upgrading an existing site designed for ASIC mining to handle the much higher rack densities and redundant power requirements of GPU deployments. While this conversion requires substantial capital for power conditioning and advanced cooling, it represents a fraction of the cost of a greenfield build.
Furthermore, the speed of deployment provides a massive competitive advantage. While a greenfield AI data center can take 18 to 30 months to become operational, a miner with existing power and space can energize a cluster in a matter of months. In a market where hardware is sold out and rental rates are rising, being the first to provide live compute is more valuable than being the cheapest builder. This reality is reflected in the multi-billion-dollar deals signed by companies like CoreWeave to take over and convert existing mining sites.
Comparative Unit Economics: Greenfield vs. Conversion
To illustrate the disparity, consider a 1,000-GPU H100 SXM cluster. The all-in capital expenditure (CapEx) per GPU, including the server, networking, shared storage, and facility costs, is estimated at $50,000 for a greenfield project. For a Bitcoin miner converting an existing site, that cost drops to approximately $42,000.
The breakdown of these costs reveals that the server (GPU and chassis), networking, and shared storage costs are identical for both operators, at approximately $30,000, $4,000, and $2,500 respectively. The entire $8,000-per-GPU gap is found in the facility line. For a greenfield build, the interconnect, shell, and electrical work cost roughly $14,000 per GPU, whereas a miner-converted site requires only about $5,000 per GPU in additional infrastructure investment.
When these figures are stress-tested against market volatility, the miner’s advantage becomes even clearer. If rental rates soften and a three-year depreciation life is applied, a greenfield cluster often compresses toward or below the break-even point because it carries significantly more facility capital to depreciate. The miner-converted cluster, however, remains profitable. This resilience is not due to cheaper silicon, but because the miner committed less new capital to the facility and began generating revenue long before the greenfield competitor could even clear the utility interconnect queue.
Operational Tensions and the Future of the Industry
Despite these advantages, the transition from mining to AI compute is not without friction. Bitcoin mining traditionally thrives on "curtailment" and demand-response programs, where the miner earns revenue by shutting down during periods of high grid stress. AI customers, conversely, demand near-100% uptime and high-reliability power. This creates a fundamental conflict: the interruptible power that makes mining profitable is incompatible with the service-level agreements (SLAs) required by AI workloads. Consequently, while a miner can convert part of a site for AI, they cannot necessarily apply the same power-saving strategies across the entire facility.
As the neocloud category matures, consolidation is inevitable. The dividing line between winners and losers will not be defined by who possesses the latest chip or the most advanced software stack. Instead, the market will favor those who started the race with infrastructure already in the ground. Energized power, existing shells, and secured interconnects are assets that cannot be conjured quickly, regardless of how much capital is available.
The structural advantage held by Bitcoin miners is significant, but it is not a guarantee of success. To capitalize on this window of opportunity, miners must model their economics with brutal honesty, accounting for the real costs of facility upgrades and avoiding overly optimistic assumptions about hardware prices or rental stability. In the current environment, characterized by a "Hopper crunch" and rising rates, the market favors the swift. The miners who can navigate the technical and financial transition today will likely become the dominant neocloud operators of tomorrow, while others remain stalled in the years-long queue for grid connectivity.
