Neocloud Unit Economics: What Bitcoin Miners Already Know About Running GPU-Only Clouds

A neocloud's profit-and-loss statement is a race between two clocks. The GPU starts depreciating the day it ships. Revenue starts the day it's energized. Everything between those two dates is pure loss, and the gap between them is where most of the money in this business is won or lost.

That framing matters because the napkin math on a GPU cloud almost always looks great and the real math frequently doesn't. Model a neocloud honestly and most greenfield H100 deployments come out razor-thin. Model the same cluster from a Bitcoin miner's position and it holds up. The difference isn't cheap electricity, and in today's sold-out market it isn't cheap GPUs either. It's the facility a miner already has in the ground, and how fast they can turn it on.

TLDR

• Neocloud margins live or die on one race: utilization times rental rate, against the depreciation clock.
• The error that flatters nearly every estimate isn't assuming rates fall. It's assuming they move in one direction at all. H100 rates crashed from their 2023 peak, then reversed hard in 2026 on a Hopper supply crunch.
• A miner's durable edge isn't cheap power (worth only ~3–4 margin points), and in today's tight market it isn't cheap hardware either. It's the facility capital, above all the grid interconnect, and the ability to energize in months, not years.
• We're publishing the full formula and an interactive calculator so you can run the numbers, and your own rate curve, yourself.

The neocloud business model, in one equation

A neocloud is a GPU-first cloud provider that rents compute to AI workloads: the GPU-as-a-Service business model, stripped of the hyperscaler's other product lines. The economics reduce to one equation. Revenue equals rental rate times utilization times hours, minus power, maintenance, fixed operating costs, depreciation, and financing.

The number operators like to quote is gross margin, revenue minus power and maintenance, often cited in the 55–65% range. It sounds wonderful. The number that decides whether the business survives is net, after depreciation and interest, and that's where neoclouds routinely go red. The lever between them is utilization, but what counts as "good" depends on how the build is financed. Most debt-financed builds are underwritten against an offtake contract, so utilization there is effectively 100%: the capacity is sold before a GPU is racked. The ~70% break-even figure applies to a different animal, the self-funded operator putting GPUs on a marketplace and taking demand risk directly. For that operator, the gap between a half-idle cluster and a full one is the gap between losing six figures a month and making them.

Idle GPUs don't pause their depreciation while they wait for customers; they decay on schedule whether or not they earn a dollar. The most sophisticated operators claw some of that back, using automation to relist a customer's contracted-but-unused hours on the spot market at a higher rate, but the underlying decay never stops.

The two things every neocloud model gets wrong

Before running any numbers, two errors are worth naming, because almost every public estimate makes at least one of them.

The first is treating the rental rate as a number you set and forget, usually flat, sometimes gently declining. Rates are neither stable nor one-directional. H100 rates fell 64–75% from their 2023 peak as supply caught up, and everyone modeling this business internalized "rates always fall." Then, in 2026, they reversed. With Blackwell deployments backed up and Hopper capacity sold out, H100 rental rates climbed roughly 20% year-over-year, one-year contract rates jumped nearly 40% off their late-2025 low, and NVIDIA confirmed Hopper prices are still rising. A model that bakes in a fixed decline is fragile; so is one that assumes the current firming lasts. The deeper point is that the rental curve and the depreciation schedule are the same bet viewed twice: a GPU's falling economic value is the rate curve on the revenue line and depreciation on the cost line. Set a five-year depreciation while assuming rates hold, and you've made that optimistic bet twice. An honest model makes it once, explicitly, and lets you set the direction.

The second error is pretending the depreciation question is settled. In late 2025, Michael Burry argued the economic life of an AI GPU is two to three years, and that the major AI spenders were understating depreciation by roughly $176 billion across 2026–2028. Most operators, and Meta explicitly, book five to six years. Both camps have a case, and the current Hopper shortage has, for now, pushed resale values in the opposite direction from what a short-life assumption predicts. That's why we won't pick a number for you: the calculator lets you toggle a two-, three-, or five-year life and watch the outcome move. Depreciation on this hardware is genuinely hard to model, and any tool that pretends otherwise is selling you something.

Where the miner's edge actually comes from

The popular story is that miners win at AI because they have cheap power and cheap hardware. In 2026, both halves are shakier than they sound. Power is only about 3–4% of revenue at current H100 rates, so even a big difference in electricity cost, four cents versus eight, barely moves the outcome. And the cheap-hardware edge has largely evaporated: with Hopper sold out, secondary-market H100 servers trade well above where they sat six months ago, and new-old-stock changes hands near original pricing. A miner counting on discounted used silicon is counting on a market that doesn't currently exist. The acquisition-cost advantage is real in some cycles, but it isn't something to build a thesis on.

The durable edge, the one that survives whatever rental rates and hardware prices do, is capital and time. The largest line item after the servers is the facility: interconnect, land and shell, electrical build, cooling, and power conditioning. A miner has already paid for the slowest and most expensive of those. The interconnect, the single hardest thing to acquire in the entire AI buildout and gated by multi-year utility queues, is already energized; the land, shell, and base electrical are sunk. What the miner still funds is the delta: upgrading a site built for ASIC hardware to handle the far higher rack densities packed GPU deployments require, plus the power conditioning and redundancy AI customers expect and mining never needed. That delta is real money, but a fraction of building from zero.

The second edge is speed. A greenfield AI data center takes 18 to 30 months to bring online; a miner with power and space can energize in months, starting the revenue clock before construction would even finish. In a market where Hopper is sold out and commanding rising rates, being first to energize beats being cheapest to build. It's why CoreWeave and others have signed multi-billion-dollar deals to take over converted mining sites: an old mine comes with the one thing nobody can buy quickly, a live grid connection at scale.

One tension is worth stating plainly. Curtailment and demand-response revenue are genuine for miners, but interruptible power conflicts directly with the uptime AI customers demand. It's an edge for part of a converted site, not a free lunch across all of it.

The numbers: greenfield vs. miner-converted H100 cluster

Consider a 1,000-GPU H100 SXM cluster, built two ways. The all-in capital cost per GPU (server, networking, shared storage, and facility) comes to roughly $50,000 greenfield and $42,000 for a miner converting an existing site. Server, networking, and storage cost the same for both; the entire gap sits in the facility line.

One line worth pausing on is shared storage. It rarely makes the napkin math, but a high-performance parallel storage tier feeding a training cluster runs into the millions for a site this size, and both operators pay it in full. It doesn't shift the miner-versus-greenfield gap, but leaving it out understates what either build actually costs.

Now apply the corrections. Both operators pay the same for hardware, because in today's sold-out market they do. Hold rates flat and depreciate over five years, and both look fine: healthy margins, reasonable payback. That's the version most pitch decks show. Now stress it with rates softening again over a three-year life. The greenfield cluster compresses toward break-even or below, carrying $9,000 more per GPU of facility capital to depreciate against uncertain revenue. The miner-converted cluster stays positive, not because it bought cheaper GPUs, but because it committed far less capital to the facility and started earning while the greenfield build was still in the interconnect queue.

The point isn't that one scenario is the truth. It's that the gap between the two operators is almost entirely the facility line, the one part of the cost stack a miner has already paid and the one part that doesn't move with the GPU market. Whatever rates and hardware prices do, that gap persists. (Final figures to be locked against the live Ornn OCPI-H100 index at publish; all values above are estimates.)

Run it yourself: the neocloud unit economics calculator

Numbers in an article are an argument; numbers you can change are a tool. Below, model both deployments against your own inputs (GPU count, rental rate, annual rate change, utilization, acquisition cost, storage, power cost, depreciation life, and cost of capital) and watch the two economics update side by side.

The one thing this calculator does that most free tools don't is let the rental rate move over the asset's life, in either direction: set it to decline if you think the 2024–2025 softening resumes, flat or rising if you think the Hopper crunch holds. Relative to an institutional model, it simplifies two things: one blended rental rate rather than a per-SKU forward curve, and one utilization figure rather than a training/inference/spot blend. When the numbers work, the next question is sourcing and allocation, and that's the point to start a conversation with Luxor about hardware.

Who wins the neocloud race

The neocloud category is going to consolidate, and the dividing line won't be who has the newest silicon or the cleverest software. It will be who starts the revenue-versus-depreciation race with the assets already in the ground. Interconnects, shells, and energized power don't appear on a napkin, but they're most of what separates a cluster that clears from one that doesn't, and they're exactly what a Bitcoin miner spent years acquiring for a different reason.

That's a structural advantage, not a guaranteed win. It only converts if the miner models the economics honestly: the real facility delta included, no fantasy discount on hardware, and a rate assumption they actually believe. In a market where Hopper is sold out and rates are climbing, the window favors whoever can energize first. The miners who run the numbers now, and act on them, will be operating neoclouds while everyone else is still waiting in the interconnect queue.