A reserved GPU is cheaper per hour than the same card on-demand. That fact tells you almost nothing about whether to reserve one, because reserved capacity bills you for the whole term whether you run it or not. The real question is not "is reserved cheaper" — it's "will I use it enough to beat the discount I'm paying for up front." That has an exact answer.

The rule: break-even utilization equals the discount ratio#

Set the two bills equal. Reserved charges you the reserved rate for the entire term, no matter what. On-demand charges only for the hours you actually run:

reserved_rate × term  =  ondemand_rate × (utilization × term)

The term cancels on both sides, and you're left with:

break-even utilization  =  reserved_rate ÷ ondemand_rate

That's the whole decision. If a 1-year reserved H100 runs about $1.89/hr against roughly $2.99/hr on-demand — a ~37% discount — the price ratio is 0.63, so your break-even utilization is 63%. Run that GPU more than 63% of the year and reserving wins. Run it less, and on-demand is cheaper even though its hourly sticker is higher, because you stop paying the moment you stop using it.

The counterintuitive part#

A deeper discount lowers your break-even: at 40% off, the ratio is 0.60, so you only need 60% utilization to justify it. At 20% off, you need 80%. So the bigger the discount, the easier reserved is to justify — on utilization alone.

But a longer contract raises your risk, and that's the axis the math doesn't show. Break-even utilization tells you the floor; your forecast confidence tells you how far out you dare commit. Committing to 63% utilization for one month is a safe bet if you've got a month of data. Committing to it for two years is a bet on your business's shape in 2028. Blend the two: reserve only when both the utilization clears the break-even and you'd stake the term on your forecast.

What the market will actually sell you#

For H100/H200-class cards, providers rarely sign for under six months — conversations realistically start there, and the sweet spot is one to two years. AWS offers shorter Capacity Blocks (1-day or 14-day) for H200, but on-demand H200 availability is limited. The practical menu is short reserved (1–3 months, ~20–30% off), medium (6–12 months, ~30–40%), and long (1–2 years, 40%+). Prices move monthly and by provider — price your own workload against a current map before you sign.

The savings are real when the utilization is real: at wholesale rates, moving 8×H200 from on-demand (~$4/hr) to reserved (~$2.70/hr) over six months of continuous use saves about $43,000. The word doing the work in that sentence is continuous.

The sequence for a team of one#

Don't reserve first. Reserve last, and only the part you can prove.

  1. Start on-demand or serverless. Pay the higher hourly rate to buy yourself flexibility and, more importantly, data. Measure four weeks of real utilization.
  2. Push batch to spot. Any fault-tolerant, checkpointable job — offline embedding, evals, fine-tuning runs — belongs on interruptible capacity at 50–70% off, not on a reservation.
  3. Reserve only the flat 24/7 floor. Once your utilization graph shows a steady base that clears the break-even, reserve that base and keep your bursts on on-demand. The classic, expensive mistake is reserving for your peak and paying around the clock for capacity you touch in spikes.

Reserved GPUs are a utilization instrument, not a discount you grab because the per-hour number looks better. Compute your break-even, check it against honest data, and commit to the floor — never the peak. If you're still deciding where to rent in the first place, start with where to serve an open model.