The unit economics of debt-financed GPU clouds

Depreciation is usually the biggest single cost. An 8-GPU H100 server from an OEM like Dell or Supermicro runs roughly $250,000 to $400,000 as of early 2026.² That cost gets spread across the years the operator expects the hardware to earn, which is where accounting choices start to matter. CoreWeave depreciates its servers over 6 years, Lambda over 5, and Nebius over 4.²

The length of that schedule matters more than it first looks. A $300,000 server on a 6-year schedule carries about $50,000 a year in depreciation. The same server on a 4-year schedule carries about $75,000. Across a 10,000-GPU fleet that difference adds up to roughly $30 million a year in reported expense, on identical hardware.² A shorter schedule is more conservative and front-loads the cost, which is part of why depreciation assumptions get scrutinized closely when these companies raise money or report earnings.³

Power and space come next. GPU racks draw far more than traditional servers, roughly 40 to 80 kW per rack, three to eight times what a conventional enterprise rack draws.⁴ That density is why purpose-built AI colocation is scarce, and why power, more than chips, is now the main limit on growth. Each new GPU generation makes this harder: a B200 can draw up to 1,000W per GPU against about 700W for an H100, so a facility that comfortably ran one generation may not have the headroom for the next.²

Interest works differently from the other costs. Most large GPU fleets are bought with debt, secured against the hardware and the contracted revenue it's expected to generate. The repayments come due whether the GPUs are busy or idle. CoreWeave carried about $21.4 billion in debt at the end of 2025 and paid roughly $1.2 billion in interest over the year.¹ When utilization drops, that payment doesn't drop with it, which makes it the hardest cost to carry when demand is weak.

‍

The bridge from EBITDA to net loss

‍

CoreWeave's own 2025 figures show how a 60% margin becomes a net loss. Revenue was about $5.13 billion, up from $1.92 billion the year before, at an adjusted EBITDA margin around 60%.¹ That works out to roughly $3.1 billion of adjusted EBITDA, a measure taken before depreciation, interest, and stock-based compensation. Below it, about $2.45 billion of depreciation and amortization, $1.2 billion of interest, and stock-based compensation brought the company to a net loss of about $1.17 billion.¹ Most of those charges are non-cash. Depreciation spreads the earlier cost of the hardware across its useful life and lowers reported earnings without money leaving the business that year; interest is the part that is paid in cash. So the net loss reflects the cost of building and financing the fleet more than the economics of renting GPUs out.

The loss persists because the company keeps expanding. Each wave of new debt-financed hardware adds depreciation and interest, and during a heavy buildout those charges outrun the profit from the fleet already in service. CoreWeave is reinvesting far more than it earns, against a contracted backlog that reached $66.8 billion at the end of 2025.¹ Amazon looked similar in the early 2000s, reporting losses while it poured the cash from its existing operations back into growth. It isn't unique to CoreWeave. Nebius and Applied Digital, the other public GPU clouds, are smaller but run on the same cost structure.

At current rental rates a GPU is usually profitable month to month. The uncertainty is the full life of the hardware. The question is whether the hourly rate in year three still covers depreciation and financing, after prices have changed and with the debt still being repaid. On illustrative public assumptions, a debt-financed fleet breaks even somewhere in the high-$1 range per GPU-hour, before colocation, power, and operations are added.² That isn't a wide margin against rates that moved between about $1.70 and $2.35 for a one-year H100 contract over six months in late 2025 and early 2026.⁵