The answer isn't always obvious, and the GPU cloud industry has muddied the waters by copying AWS's terminology without matching the economics. Here's the real math.
On io.net, on-demand pricing is already 50-70% below hyperscaler reserved rates. That means the traditional calculus — "commit for a year to save 30%" — doesn't really apply the same way. An H100 on io.net costs $2.20/hr on-demand. AWS charges $4.99/hr on-demand, and their 1-year reserved rate comes to roughly $3.50/hr. So io.net's on-demand is already cheaper than AWS's best committed pricing.
That said, io.net does offer enterprise reserved plans with additional discounts. Let's break down when they make sense.
The Economics, Plainly
On-demand (pay-per-second):
- No commitment. Spin up, use, spin down.
- Full flexibility to switch GPU types, scale up/down instantly
- io.net prices: $0.18/hr (4090), $1.49/hr (A100 80GB), $2.20/hr (H100)
- Best for: variable workloads, experiments, burst capacity
Reserved (enterprise commitment):
- 1-6 month commitment on io.net (shorter than AWS's 1-3 year terms)
- 10-20% additional discount on top of on-demand rates
- Guaranteed capacity — your GPUs won't be unavailable during demand spikes
- Best for: steady-state production, always-on inference, continuous training
Break-Even Analysis
The question is whether your usage pattern justifies the commitment:
H100 SXM example:
| Plan | Hourly rate | Monthly (24/7) | Monthly (12hr/day) | Monthly (8hr/day) |
|---|---|---|---|---|
| io.net on-demand | $2.20 | $1,584 | $792 | $528 |
| io.net reserved (3mo) | $1.98 | $1,426 | — | — |
| io.net reserved (6mo) | $1.76 | $1,267 | — | — |
| AWS on-demand | $6.98 | $5,026 | $2,513 | $1,675 |
| AWS 1yr reserved | ~$3.50 | $2,520 | — | — |
Reserved on io.net makes sense if you're running 24/7 or near it. At 12 hours/day, on-demand is smarter because you're not paying for idle hours.
The crossover point: Reserve if your average utilization exceeds 18 hours/day and you can predict it 3+ months out. Otherwise, on-demand on io.net is already so cheap that the flexibility premium is worth paying.
What Most Teams Get Wrong
Mistake 1: Reserving before you know your workload. Startups frequently over-commit based on optimistic projections. If your inference traffic is growing but unpredictable, stay on-demand and let the numbers stabilize for 2-3 months before committing.
Mistake 2: Comparing reserved-to-reserved across providers. The right comparison is io.net on-demand vs. competitor reserved. When io.net's pay-as-you-go H100 at $2.20/hr beats AWS's 1-year reserved at $3.50/hr, the commitment premium on AWS is just wasted money.
Mistake 3: Ignoring the option value of flexibility. If you lock in 8x A100s for 6 months and then discover that H100s would cut your training time in half (making them cheaper overall), you're stuck. On-demand lets you switch GPU types overnight.
Mistake 4: Not accounting for idle time. Reserved instances cost money 24/7 whether you use them or not. If your inference traffic drops on weekends or overnight, those idle hours add up. A workload that runs 16 hours/day burns 33% of reserved spend on empty time.
A Practical Framework
Ask yourself three questions:
1. Is the workload predictable?
If yes, and it runs 20+ hours/day → consider reserved.
If it's bursty, seasonal, or experimental → stay on-demand.
2. How long will I need this exact GPU type?
If 6+ months with confidence → reserved makes sense.
If the landscape is shifting (new model architectures, evolving requirements) → on-demand preserves optionality.
3. What's my total monthly GPU spend?
Under $5,000/month → on-demand simplicity saves ops time.
Over $10,000/month → talk to io.net enterprise about a custom reserved deal with volume discounts.
The io.net Advantage: No Wrong Choice
The dirty secret of GPU cloud pricing is that io.net's on-demand rates already beat most competitors' reserved pricing. So the penalty for "choosing wrong" is small — maybe 10-20% difference — instead of the 2-3x swings you see between AWS on-demand and reserved.
This means you can start on-demand, prove your workload, and switch to reserved later without feeling like you've been overpaying. It's the cheapest possible version of a low-risk decision.
Start on-demand, upgrade when ready — io.net GPUs from $0.18/hr with no commitment. View pricing
