Used to be, this article was three sentences long. "Renting GPUs beats owning: someone else handles the failures, the depreciation is their problem, and capital is better spent on people." That advice was correct from roughly 2016 to 2024. Then the ground moved.
What changed
1. The rental floor stopped falling
Rental rates for a given card class bottomed out and stabilized. H100-class hours settled into the $1.40-$2.20 range on open markets — cheap versus hyperscalers, but no longer collapsing year over year, because power, real estate and hardware costs set a floor that competition can't dig through. Meanwhile list prices for the cards themselves kept behaving normally.
2. Used GPUs became liquid
This is the underrated one. A three-year-old accelerator used to be landfill with a warranty. Today there's a real secondary market — inference shops, sovereign AI programs, regional clouds and yes, DePIN supplier networks all buy used enterprise cards. Recovering 40-55% of purchase price after three years is now a reasonable planning assumption, not a fantasy. Depreciation is no longer a cliff; it's a ramp.
3. Spare capacity became sellable
Here's the twist that connects this whole column: if you own GPUs, you can now list idle hours on decentralized marketplaces. Your hardware's "unused" state has a market price. That changes the utilization math for everyone who owns any compute at all.
The break-even, roughly
| Your sustained utilization | Sensible play | Why |
|---|---|---|
| < 30% | Rent — preferably open market | Ownership overhead (power, cooling, failures, ops) exceeds savings. Idle capital. |
| 30-70% | Hybrid | Own the baseline, rent the peaks. Sell your off-peak idle hours back to the market if you can. |
| > 70%, sustained | Own | At this point you're paying rental margins on nearly every hour anyway. Buy, and treat resale as your exit liquidity. |
Sketch of the arithmetic for a data-center-grade card at ~$25K: owning costs (power, cooling, networking, a slice of an ops person) run maybe $150-250/month. Full ownership cost over 36 months, net of 45% resale recovery, lands near $19-21K — about $7.5/hr if run flat-out, before you sell a single idle hour. Every hour you rent instead at $1.80 only makes sense below roughly 35-40% utilization. Above that, the spreadsheet flips. Your numbers will differ; the shape won't.
The risks nobody prices
- Architecture risk. If inference demand shifts to hardware features your card lacks, resale value craters faster than any depreciation schedule. This is the big one, and it's a bet on software ecosystems, not silicon.
- Operational drag. A dead node at 2am is now your problem. Renting outsources on-call. If you don't have hardware people, add real money for them.
- Capital opportunity cost. $25K in GPUs is $25K not in payroll or GPUs-next-year. Buying early is a bet that your utilization forecast is right. It usually isn't, in the direction of optimism.
What we actually do
We rent what's bursty, own what's constant, and treat everything idle as inventory rather than overhead. Ten years ago that sentence would have been heretical in this industry. Today it's just the spreadsheet. The rule didn't get reversed — it got conditional. Which, coincidentally, is what happens to every "always rent" market the moment a liquid secondary market shows up. Ask anyone who ever said "nobody should own a car."
Comments