What is the difference between Savings Plans and Reserved Instances?
A Reserved Instance (RI) is a billing discount tied to a specific instance configuration: instance type or family, region, platform (operating system) and tenancy. When running usage matches those attributes, the discounted rate applies. A zonal RI is scoped to a single Availability Zone and also reserves capacity there. A regional RI gives up that capacity guarantee in exchange for applying across zones and, for Linux, across sizes within the same family.
A Savings Plan is a commitment to spend a fixed dollar amount per hour on compute for a one- or three-year term. AWS applies that hourly amount to eligible usage automatically, so you never match a purchase to an instance. Savings Plans do not reserve capacity. If you need guaranteed capacity, pair a Savings Plan with an On-Demand Capacity Reservation.
The four options sit on a flexibility spectrum:
| Commitment | What can change without losing the discount | Covers Fargate/Lambda | Capacity reservation |
|---|---|---|---|
| Compute Savings Plan | Instance family, size, region, OS, tenancy | Yes | No |
| EC2 Instance Savings Plan | Size, OS, tenancy, AZ (family and region fixed) | No | No |
| Convertible RI | Can be exchanged for a different family, OS or tenancy | No | Zonal only |
| Standard RI | Limited modifications (AZ, size within family) | No | Zonal only |
The discount reflects how much forecasting risk you take on. A narrower scope means AWS can predict your usage more reliably, so it generally offers a deeper discount. An EC2 Instance Savings Plan is priced roughly like a Standard RI, and a Compute Savings Plan is priced roughly like a Convertible RI. The underlying choice is how much of the discount you are willing to give up so your fleet can change instance family or region without stranding the commitment.
How do AWS commitments apply to a backend fleet's usage?
A commitment is an hourly spend promise, so it only pays off on usage that runs every hour of the term. In a typical backend fleet that means the always-on floor: the minimum API replicas, queue workers and databases that never scale to zero. Anything above that floor, such as daytime traffic peaks, batch jobs or deploy surges, should stay on on-demand or Spot. If you buy commitment to cover a peak, you pay for idle hours whenever usage drops below it.
You do not assign a Savings Plan to instances. Each hour, AWS applies the commitment automatically to eligible usage, starting with the usage that gets the highest discount percentage. When the commitment runs out, the remaining usage is billed at on-demand rates. With consolidated billing, that discount can also spread across accounts in the organization. As a result, a team's bill can change even though its workload did not, because another team launched instance types that now absorb the discount first. Per-team showback in shared accounts therefore needs an explicit allocation rule. The raw billing data will not give you one.
Track two metrics:
- Coverage: the share of eligible usage paid at a committed rate. Low coverage means you are paying on-demand prices for steady load.
- Utilization: the share of the commitment actually consumed. Low utilization means you bought more commitment than you use.
If you optimize for only one of these, you make the other worse.
Do not apply commitments and Spot to the same workload. Savings Plans do not apply to Spot usage. If a team moves its stateless workers to Spot after you committed for them, the commitment goes unused and utilization drops.
+------------------------------------------------------------+
| Burst: daytime peaks, batch jobs, deploy surges |
| --> on-demand or Spot (Spot never draws on commitment) |
+------------------------------------------------------------+
| Always-on floor: min API replicas, queue workers, DBs |
| --> Savings Plan / Reserved Instance commitment |
+------------------------------------------------------------+
v
+------------------------------------------------------------+
| Each hour AWS applies the commitment automatically |
| (across accounts with consolidated billing): |
| 1. highest-discount eligible usage first |
| 2. next-highest, until commitment is spent |
| 3. leftover eligible usage --> on-demand rate |
+------------------------------------------------------------+
+-----------------+-----------------+
v v
+--------------------------+ +----------------------------+
| Team A account | | Team B account |
| bill shifts though its | | launches types that |
| workload did not change | | absorb the discount first |
+--------------------------+ +----------------------------+
+-----------------+-----------------+
v
+------------------------------------------------------------+
| Coverage = committed-rate usage / all eligible usage |
| Utilization = commitment used / commitment purchased |
| Watch both: fixing one alone worsens the other |
+------------------------------------------------------------+
The commitment covers the always-on floor while on-demand and Spot handle bursts; AWS applies the commitment to the highest-discount usage first, so it can shift between teams' accounts, and coverage and utilization show whether the commitment fits the actual usage.
Why should a Compute Savings Plan be the default for most backend fleets?
A narrower commitment buys a deeper discount only while your usage keeps matching its terms. Backend fleets rarely stay that still. A typical year might include several of these changes:
- Processor architecture migration, such as moving from x86 to Arm-based Graviton instances. This changes the instance family.
- Containerization, such as moving services from EC2 to Fargate.
- Serverless, such as moving queue consumers to Lambda.
- Rightsizing, such as moving to a newer generation or a different family.
- Region moves, such as relocating for latency or data residency.
Each of these changes which commitments your usage qualifies for.
How each commitment type handles change
- Standard RIs are tied to an instance family and region.
- EC2 Instance Savings Plans are flexible across size, OS and tenancy within one family in one region. A family or region change leaves them with nothing to apply to. You keep paying for that commitment until the term ends.
- Convertible RIs can be exchanged for different attributes. Someone has to notice the change and perform the exchange.
- Compute Savings Plans apply to EC2 usage across families, sizes, OS and regions, and also to Fargate and Lambda. The commitment follows the spend through all of the changes above.
Who should use it
The discount on a Compute Savings Plan is smaller than on the narrower options. The real comparison is a slightly lower rate against the risk of paying for a commitment that no longer matches anything you run.
A Compute Savings Plan is the right default for any team that has migrations on its roadmap or cannot predict its instance mix for the full term. That describes most backend teams. Size it to your baseline floor before those changes begin, not after they finish.
How it saves engineering time
Infrastructure teams can change architecture, platform or region without first checking whether a commitment would be stranded. Finance does not need to re-plan purchases around each roadmap change. Coverage is checked against total compute spend, not against each instance type.
How do you size a commitment from a measured baseline floor?
Step 1: Export hourly eligible spend
Pull hourly on-demand spend for the usage the plan would cover: EC2, Fargate and Lambda compute for a Compute Savings Plan. Use your billing export, such as the Cost and Usage Report. Exclude usage already covered by existing commitments, and exclude services the plan does not discount. Use a window of at least several weeks so that weekly cycles, deploy patterns and month-end jobs all appear. Daily totals are not enough because a Savings Plan is a per-hour commitment. Any hour where usage falls below the commitment wastes the difference.
Step 2: Find a low percentile, not the average
Sort the hourly values and read off a low percentile, such as the 5th or 10th. The average sits well above the overnight and weekend troughs. A commitment sized to the average will go unused during those hours. A low percentile is the level your fleet almost never drops below. The snippet that follows loads the export, groups by hour, and prints a set of percentiles alongside the mean, so you can see the gap between them.
Step 3: Convert and commit to part of the floor
The commitment is stated in discounted dollars per hour. Multiply your on-demand floor by your expected Savings Plan rate ratio before you buy. Then commit to a fraction of that figure, not all of it. The floor you measured reflects the past. Rightsizing, Graviton migrations or a service moving to containers can lower it.
Step 4: Layer purchases over time
Buy in tranches spread across weeks or months, and re-measure the floor before each one. Smaller, staggered commitments expire on different dates. That staggering gives you regular points to adjust coverage instead of one large renewal decision.
# Compute hourly spend percentiles from a billing export to find a commitment floor.
# Assumes a CSV already filtered to eligible, uncovered on-demand compute usage,
# with columns: usage_start (timestamp) and on_demand_cost (USD).
import sys
import pandas as pd
df = pd.read_csv(sys.argv[1], parse_dates=["usage_start"])
# Bucket line items into clock hours and sum, because the commitment applies per hour.
hourly = (
df.assign(hour=df["usage_start"].dt.floor("h"))
.groupby("hour")["on_demand_cost"]
.sum()
)
# Fill hours with no rows as zero spend so true troughs are not hidden.
full_range = pd.date_range(hourly.index.min(), hourly.index.max(), freq="h")
hourly = hourly.reindex(full_range, fill_value=0.0)
print(f"hours in window: {len(hourly)}")
print(f"mean: {hourly.mean():.2f} USD/h")
for p in (0.01, 0.05, 0.10, 0.25, 0.50):
print(f"p{int(p * 100):<3}: {hourly.quantile(p):.2f} USD/h")
# Placeholder inputs: replace with your own rate ratio and risk tolerance.
RATE_RATIO = 0.70 # assumed discounted / on-demand price ratio
COMMIT_FRACTION = 0.6 # commit to only part of the floor, add tranches later
floor_on_demand = hourly.quantile(0.10)
commitment = floor_on_demand * RATE_RATIO * COMMIT_FRACTION
print(f"first-tranche commitment (illustrative): {commitment:.2f} USD/h")
A pandas script that aggregates an hourly eligible-spend export, prints low percentiles next to the mean, and converts the 10th-percentile floor into a partial first-tranche commitment using placeholder rate and fraction values.
Illustrative estimate: committing at the floor, the average and the peak
Illustrative estimate (assumed numbers, not a benchmark). Assume a fleet whose on-demand-equivalent compute spend follows a daily profile of 8 hours at a $60/hour floor, 8 hours at $100/hour, and 8 hours at a $140/hour peak. That makes the average $100/hour and on-demand cost $2,400/day. Assume a flat 30% commitment discount. For simplicity, the commitment (C) is expressed in on-demand-equivalent dollars per hour.
For each hour with usage U:
- Cost = C × (1 − 0.30) + max(0, U − C)
- Unused commitment = max(0, C − U)
| Commitment level | Commitment cost/day | On-demand overflow/day | Net cost/day | Savings vs on-demand | Hours with unused commitment |
|---|---|---|---|---|---|
| Floor ($60/hr) | $1,008 | $960 | $1,968 | $432 (18%) | 0 |
| Average ($100/hr) | $1,680 | $320 | $2,000 | $400 (17%) | 8 |
| Peak ($140/hr) | $2,352 | $0 | $2,352 | $48 (2%) | 16 |
Committing at the peak eliminates on-demand spend but nearly eliminates the savings. You pay for $960/day of commitment that nothing consumes.
The rule behind this result: each extra dollar of commitment costs $0.70 every hour. It only returns $1.00 in the hours when usage actually reaches it. An increment pays for itself only if usage fills it in more than 70% of hours, which is 1 minus the discount. The band between the floor and the average is filled only 67% of the time in this profile, so it loses slightly.
The $960/day of on-demand overflow at the floor is not a mistake. It is what you pay to scale down, change instance families, or shrink the fleet without being locked into spend.
When should you avoid Savings Plans or choose Reserved Instances instead?
That flexibility is worth less in a few cases, and in some it does not apply at all.
A fleet that will not change for the full term. If a service has run on the same instance family in the same region for years and nothing on the roadmap changes that, a Standard RI or an EC2 Instance Savings Plan earns a deeper discount than a Compute Savings Plan. Both trade flexibility you would not use for a lower rate.
You need an exit. A Savings Plan cannot be cancelled, exchanged, or resold. Convertible RIs can be exchanged for different instance attributes, but only for equal or greater value, and their discount is smaller. Standard RIs can sometimes be sold on the RI Marketplace, subject to its rules.
You need guaranteed capacity. Savings Plans and regional RIs only discount usage. They do not reserve capacity. If a zone-pinned workload must be able to launch during a regional shortage, use zonal RIs or On-Demand Capacity Reservations. A Savings Plan can still discount the reserved hours.
The workload is not EC2, Fargate, or Lambda. RDS, ElastiCache, OpenSearch, and Redshift have their own reservation models. Check what current Savings Plans cover for those services before assuming they apply.
Your history is about to stop being true. After a migration, re-architecture, or large customer churn, the measured floor describes a fleet that no longer exists. Wait until the new baseline has settled before committing.
| Situation | Better choice | Why |
|---|---|---|
| Stable, single-family fleet | Standard RI / EC2 Instance SP | Deeper discount |
| Likely instance changes | Convertible RI | Exchangeable |
| Zonal capacity required | Zonal RI / Capacity Reservation | Reserves capacity |
| Databases, caches, search | Service-specific RIs | Check SP coverage |
| Recent migration or churn | Delay commitment | Baseline unreliable |
Term length, payment options and operational ownership
A one-year term gives a shallower discount but expires before most architectures drift far. A three-year term discounts more deeply, but it is a bet that the fleet will exist in roughly its current shape for three years: the same rough compute volume, no large move to managed services, and no major cut in spend. Payment options trade cash against rate.
| Option | Discount depth | Cash flow impact | Lock-in risk |
|---|---|---|---|
| 1-year, no upfront | Lowest | Spread monthly | Low |
| 1-year, all upfront | Moderate | Large payment now | Low |
| 3-year, partial upfront | Higher | Split between upfront and monthly | High |
| 3-year, all upfront | Highest | Largest payment now | Highest |
Paying upfront does not lower the commitment risk. It only moves cash earlier. Pick the term based on how confident you are in the architecture, and pick the payment option based on finance's cost of capital.
Ownership and renewal
Name one owner, usually the platform team. That owner reviews coverage and utilization reports monthly and before every purchase. Stagger purchases across quarters so expirations never land at once. Keep a renewal calendar that starts review weeks before each expiry. Without one, coverage silently drops back to on-demand, and the only signal is a larger bill.
Central purchasing under consolidated billing creates a second problem. Discounts get applied across accounts, so a team's bill can look cheaper than its actual usage. A team that never sees its real rate has little reason to rightsize. Allocate commitment costs back to the teams whose usage consumes them. The upkeep is a few hours of engineering time each month for review and renewal, which is small next to the rate difference it protects.
How OneEnv can help
Disclosure: OneEnv is the product behind this blog.
OneEnv (oneenv.io) is a workspace-based governance platform for the things services share — environment variables, database structure, API routes and validation rules. A change to a shared value is staged on a branch, reviewed once, and then every affected service explicitly approves or rejects it before it goes live, so you can see who a change touches before it ships instead of finding out in production.
Applied to savings plans vs reserved instances, this is where it helps:
Saves time
- Every service that inherits a shared value is listed in one place, so nobody searches repos and chats to find out who is affected.
- One review per branch replaces coordinating each team separately; each service still keeps its own approve/reject.
- Changing a shared variable is one edit and one review instead of editing every service's
.envby hand.
Saves cost
- Fewer production incidents and rollbacks caused by a shared value, route or column changing under a consumer unannounced.
- Less duplicated configuration to maintain: one shared definition inherited by many services.
- A wrong or stale value is caught at review time by the services it affects, not by a failed deploy.
- OneEnv has a free plan (1 workspace, unlimited services and members, full governance), so trying it costs nothing.
These are qualitative benefits, not measured results — how much you save depends on how many services share the value.
Want to try it? Start at oneenv.io.