When to Rent, When to Reserve, When to Buy: A Decision Framework for GPU Procurement

◆ THE DECISION FRAMEWORK

◆ WHEN RENT WINS OUTRIGHT

Below 30% utilization, no close call

A GPU sitting mostly idle costs more to own than to rent by a wide margin, since ownership pays a fixed cost regardless of use while rental only bills for hours actually consumed. Teams under 40 GPU-hours a week fall well inside this zone.

◆ WHEN OWNING WINS

75-80% sustained, and it has to stay that way

Above roughly 75-80% sustained utilization against reserved cloud rates, owning starts to win. At that usage level, the fixed cost of ownership finally gets outrun by what the same capacity would cost rented hour by hour, for years. Teams running over 200 GPU-hours a week, consistently, sit in this range.

◆ THE MIDDLE ZONE: RESERVED

Where most teams get it wrong

Between those two lines is where the real decision-making happens, and where most teams get it wrong by defaulting to whichever option feels safer rather than running the actual math. Reserved cloud sits in exactly this middle zone. A 1-year reserved commitment extends the breakeven point to roughly 15 months. A 3-year reserved term extends it further, to nearly 22 months. Reserved pricing buys a real discount over on-demand, but it also locks in a specific configuration for a specific term, which only pays off if utilization actually holds at the level assumed going in.

Depreciation changes the buy side of this math faster than most teams price in. A GPU loses roughly 50% of its value by month 18. By month 30, that loss reaches 70 to 80%. A $25,000 H100 bought today can be worth $5,000 to $7,500 in under three years. Cloud rental carries none of that risk, since a rented GPU is never an asset sitting on a balance sheet losing value while it depreciates.

The utilization threshold moves depending on what a team is comparing against. A buyer escaping expensive hyperscaler on-demand pricing sees a strong return even at moderate utilization, because the baseline they're leaving is so much more expensive. A buyer already sitting on well-negotiated reserved capacity at $1.30 to $1.80 per GPU-hour needs utilization near 95% just to justify buying, because the rental baseline they're comparing against is already efficient. The same utilization number means something completely different depending on what it's being measured against.

Workload type matters as much as utilization level. Training runs benefit from dedicated, owned hardware, since the load is sustained and predictable for the duration of the run. Inference benefits from renting, since it's variable, needs geographic distribution close to users, and scales up and down in ways owned hardware can't match without sitting idle half the time. A team running both workload types often gets the split wrong by applying one procurement model to both.

An H100 lease runs $1,200 to $1,800 a month on 24 to 36 month terms, no upfront capital required. That's 30 to 50% more expensive than buying outright over the same term, but it shifts residual value risk onto the lessor instead of the buyer. Worth knowing as a fourth option, even for teams that end up choosing rent, reserve, or buy instead.

None of this requires picking one model forever. A generational stagger works for teams that can plan around it: buy the current generation for premium workloads now, and as the next generation launches and rental rates for the current generation fall, shift routine work to renting the newer tier while repurposing owned hardware for lower-priority jobs.

◆ FAQ

Frequently asked questions