Quick take

Savings Plans win on flexibility. Reserved Instances still win on raw discount for narrow, well-predicted workloads. The break-even is not a number, it is a question about your roadmap.

If you only have two minutes:

• 3-year No-Upfront Compute SP is the safe default for compute-heavy accounts in 2026.
• Standard EC2 RIs beat SPs by 1 to 4 percentage points on commitment discount, but lock you into instance family.
• Anything below 70% utilization turns the commitment into a cost, not a discount.

Why this question got harder in 2026

I get asked this every other week, and the answer keeps shifting. Three things changed since the last round of comparison posts:

Graviton4 spread. Most new workloads land on c8g or m8g by default. RIs are tied to instance family. If you commit to c7i and migrate to c8g six months later, the commitment goes to waste. SPs absorb that migration silently.

AI workloads are spiky. GPU instances run 80%+ utilization for a training week, then sit at 5% for two. A flat 1-year commitment turns that into a worse-than-on-demand bill. The new SageMaker SP helps, but only if your AI spend is concentrated in SageMaker, not raw EC2 GPU.

FOCUS billing. AWS finally exposes commitment-level utilization in the FOCUS export, so you can measure break-even after the fact. Three years ago you had to model it. Now you can read it.

The two commitment types, side-by-side

Convertible RIs sit between the two: flexible on family, but you give up 5 to 10 points of discount versus Standard RIs.

The break-even math, with real numbers

Here is the simplest case: you run one m7i.large 24/7 in us-east-1. On-demand list price is roughly $0.1008 per hour, or $883 per year.

On-demand baseline (1 year):    $883
1-year No-Upfront Compute SP:   $618   (30% off)
1-year No-Upfront Standard RI:  $556   (37% off)
3-year No-Upfront Compute SP:   $415   (53% off)
3-year No-Upfront Standard RI:  $336   (62% off)
3-year All-Upfront Standard RI: $309   (65% off)

The 3-year Standard RI looks like the obvious winner, until you ask what happens if you migrate to Graviton in month 14. The commitment is now stranded.

Break-even months for the 3-year SP (against on-demand) is roughly 5 months at 100% utilization, 7 months at 70% utilization, never at sub-50% utilization.

The number that matters is not the headline discount. It is the utilization floor where the commitment stops paying for itself.

The hidden modifiers

Every break-even calculator on the internet ignores three things that decide real outcomes:

1. Roadmap volatility

If your team has a quarterly habit of "let's try the new instance family," 3-year Standard RIs will burn you. Convertibles or Compute SPs hold up better. The cost of flexibility is real but smaller than the cost of a stranded commitment.

2. Payment terms

All-Upfront feels like a deal because the headline discount is highest. But the working capital cost of paying $1,500 today versus $50/month over 3 years is often higher than the 4-point discount you gain. Treasury's opinion matters here.

3. Spend growth rate

If your AWS bill is growing 30% year over year, a fixed-dollar Compute SP commitment becomes a smaller and smaller share of your bill. That makes the flexibility of SPs more valuable than the discount of RIs, because new spend lands on-demand anyway.

A decision framework in three questions

I run every client through these three before recommending a mix:

1. Will the instance family I commit to today still be optimal in 18 months? If yes, Standard RI. If no, SP.
2. Is my AWS spend growing more than 20% per year? If yes, SP-heavy. If no, RI is safe.
3. Is at least 70% of my baseline spend predictable for the next 12 months? If no, do not commit at all. Stay on-demand and revisit quarterly.

Common mistakes I see in 2026

• Buying SPs at peak usage, not steady-state. The high-water mark is not the floor.
• Mixing 1y SP + 3y RI on the same workload. The 1y SP applies first by AWS's rules, so the RI under-utilizes. You lose discount, not gain it.
• Ignoring SageMaker SP if AI workloads are 30%+ of spend. The discount stack is different from Compute SP.
• Forgetting Spot. For batch, training, and any retry-tolerant workload, Spot still beats every commitment by a wide margin. Commit only the workloads that genuinely cannot run on Spot.

How to actually measure it after committing

The FOCUS billing export in 2026 includes per-commitment utilization rate as a first-class field. Pull it monthly and group every commitment by its average utilization. Anything below 70% for the month is a commitment that lost money against on-demand.

Three checks worth running on every renewal:

• Average utilization for the month. Below 70% = renegotiate or shrink the commitment.
• Underused dollar amount per commitment. Multiply the unused capacity by the discounted hourly rate to see the dollars literally evaporating.
• Trend over the last 3 months. A commitment dropping from 92% to 78% to 65% utilization is heading underwater fast. Catch it before the next renewal locks you in for another year.

The honest move is to factor underutilized commitments into the next renewal, not to argue with the dashboard.

Treat commitments like reservations at a restaurant. If you book a table for 8 and 4 people show up, that is not a discount.

So which one in 2026?

For most teams with steady, compute-heavy workloads on a mix of x86 and Graviton, the answer is 3-year No-Upfront Compute SP at 60 to 70% of your baseline, on-demand for the rest, Spot for batch. This is the configuration that survives instance-family migrations, AI workload spikes, and a growing bill.

If your workload is genuinely fixed (a single instance family, no migration plans, 90%+ utilization), 3-year All-Upfront Standard RI still wins on raw discount.

What is your current commitment mix? If it is leaning hard one way, the comments are open. Happy to look at the trade-offs you are weighing.