Insights
Market notes on GPU infrastructure credit: contract structures, rental rates, hardware values, and how GPU Capital underwrites them.
Written by Marek Omilian, CFA, Managing Partner.
LatestMarket
Fiber in 1999, GPUs in 2026
GPU servers are only half like 1999 fiber. The half that differs, asset life, decides whether a lender gets repaid.
In this note
- The shape of the build matches 1999: capex ahead of revenue, debt and equity paying for it, overbuilding, then consolidation.
- Asset life breaks the analogy. Fiber plant lasts 25 to 40 years; GPU servers are depreciated over about five to six. A fiber glut could wait for demand; a GPU glut expires.
- Today’s buyers are better: the stronger operators sign three to five year contracts, and GPUs trade on a live secondary market with measurable prices.
All insights
7 insightsShowing 4 of 7 insights for operatorsShowing 0 of 7 insights for investorsShowing 3 of 7 insights for market
Fiber in 1999, GPUs in 2026
Fiber lasted decades; GPU servers are depreciated over about five to six years. A GPU loan has to be repaid by today’s customers, inside that life.
Market5 min read
The free-for-all is over: AI infrastructure now has to pencil out
Meta’s free cash flow went negative in Q3 and Gartner sees $3.5 trillion of AI capex exposed to impairment. AI projects now have to pencil out first.
Market6 min read
What it takes to qualify for a GPU loan: the lender’s checklist
What lenders check before approving a GPU loan: the customer contract, node, sponsor, power and structure, plus the three mistakes that kill most deals.
Operators5 min read
Where GPU infrastructure fits in asset allocation
How GPU server financing differs from AI stocks: where the return comes from, how cash comes back, what can go wrong, and how much it diversifies.
Market10 min read
B300 vs H200 vs RTX PRO 6000: which GPU server pays back fastest?
We estimate H200 and RTX PRO 6000 nodes recover their capex in about 22 months on a 3-year contract; B300 takes 27. The payback math, and what breaks each case.
Operators9 min read
The real cost of financing AI compute: WACC from hyperscalers to tier 3 neoclouds
The same GPU costs about 11% a year to fund at a hyperscaler and roughly 23% at a small neocloud. Debt, equity and asset WACC, and when debt stops helping.
Operators9 min read
Locking in compute: why long-tenor, prepaid GPU contracts work, and when they stop
A long-tenor GPU contract with a customer prepayment is a trade, not a concession. Why it works in 2026, and the signals that say when it stops.
Operators7 min read