⚡ EBS Volume Cost

An EBS bill has three meters that never appear on the same screen: gigabytes provisioned, IOPS provisioned, and throughput provisioned. The first is linear in size and easy. The other two are step functions that start charging only after a free allowance you already paid for — 3,000 IOPS and 125 MiB/s on every gp3 volume, whatever its size. Get the allowance wrong and the same 3,000 IOPS prices at $8 or at $100 depending on which volume type you chose.

📐 Provisioning

Enter the numbers as they appear in the volume configuration, not the numbers the application achieved. EBS bills what is provisioned, so an idle volume costs exactly as much as a busy one of the same shape.

gp3, io2 and gp2 are billed per second with a 60-second minimum, which a monthly hour count approximates closely enough. The IOPS and throughput fields are ignored on st1 and sc1, where both are fixed by volume size.

The free allowance is per volume, not per gigabyte

Every gp3 volume includes 3,000 provisioned IOPS and 125 MiB/s of throughput at no charge, and that allowance does not scale with size. A 10 GiB volume gets 3,000 IOPS. A 4,000 GiB volume also gets 3,000 IOPS. This is the single most useful fact in EBS pricing, and it is the reason volume type matters far more than volume size for anything performance-sensitive.

storage = GiB × $0.08 IOPS = max(0, provisioned − 3,000) × $0.005 throughput = max(0, MiB/s − 125) × $0.04 — per volume, per month, us-east-1 3,000 IOPS × $0.005 = $15.00 125 MiB/s × $0.04 = $5.00 included performance = $20.00 / month, every gp3 volume

That $20.00 is not a discount. It is the price you would pay for the same performance if it were metered, and it is included in a volume that can cost as little as $0.80 a month at 10 GiB. Small volumes are therefore where gp3 looks almost free, and where a naive cost model that multiplies gigabytes by a rate goes most wrong.

gp2 requires a thousand gigabytes to buy what gp3 gives away

gp2 has no independent IOPS dial. It grants three IOPS per GiB, so the only way to buy performance is to buy storage you may not need. Reaching 3,000 IOPS — the gp3 baseline — requires a 1,000 GiB gp2 volume.

ConfigurationIOPS deliveredMonthlyMonthly per 1,000 IOPS
gp3, 10 GiB, baseline3,000$0.80$0.27
gp3, 100 GiB, baseline3,000$8.00$2.67
gp3, 1,000 GiB, baseline3,000$80.00$26.67
gp2, 1,000 GiB3,000$100.00$33.33
gp2, 100 GiB300, bursting to 3,000$10.00$33.33
gp3, 1,000 GiB, 16,000 IOPS16,000$145.00$9.06
io2, 1,000 GiB, 16,000 IOPS16,000$1,165.00$72.81

Read the fourth row against the third: 1,000 GiB of gp2 costs $100.00 a month for exactly the IOPS that 1,000 GiB of gp3 includes for $80.00. gp3 is cheaper at the same size, and the one throughput advantage gp2 retains — 250 MiB/s baseline against 125 — is worth $5.00 a month at the gp3 marginal rate. It is difficult to construct a case for gp2 on a new volume in 2026, which is why the migration is one of the few cloud changes with no downside beyond the maintenance window.

The fifth row is the trap that catches people who compare on list price alone. A 100 GiB gp2 volume advertises 300 baseline IOPS and bursts to 3,000 using a credit balance — it is genuinely fast in a benchmark and genuinely slow after ten minutes of sustained load, because the credits are exhausted and the burst stops. gp3 does not burst. It provisions 3,000 IOPS and holds them.

gp2 is not slower than gp3 at small sizes. It is slower later, which is worse, because the regression appears in production after a load test that passed.

Marginal IOPS is $5 per thousand, and it is a cliff

Above the baseline, IOPS costs $0.005 per provisioned IOPS-month. There is no taper, no volume discount, and no relationship to how much of the volume is used. The first IOPS beyond 3,000 costs $5 per thousand and the thirteen-thousandth costs the same $5 per thousand. Running a 500 GiB volume to the gp3 ceiling therefore looks like this:

Provisioned IOPSStorage (500 GiB)IOPS chargeTotal / month
3,000 (baseline)$40.00$0.00$40.00
6,000$40.00$15.00$55.00
10,000$40.00$35.00$75.00
16,000$40.00$65.00$105.00
40,000$40.00$185.00$225.00
80,000 (ceiling)$40.00$385.00$425.00

Provisioning the maximum multiplies the storage bill by 10.6× before throughput is considered, and adds $4,620 a year to a single volume. The ceiling is worth understanding rather than avoiding: gp3 supports up to 80,000 IOPS and 2,000 MiB/s per volume, and caps IOPS at 500 per GiB of size, so the full 80,000 becomes available only from 160 GiB upward and a 16 GiB volume is limited to 8,000. Throughput has its own floor: the 2,000 MiB/s maximum is provisionable from 8,000 IOPS and 16 GiB, because the two dials move together at 0.25 MiB/s per provisioned IOPS.

Volume count is an IOPS lever nobody prices

Because the baseline is per volume and the ceiling is per volume, splitting storage across volumes multiplies the total IOPS available without changing the storage cost by a cent. This is arithmetic that contradicts the usual instinct that fewer, larger volumes are simpler and therefore cheaper.

Shape, 2,000 GiB totalStorage / monthBaseline IOPSCost to reach 60,000 IOPSTotal / month
1 × 2,000 GiB$160.003,000+$285.00 for 57,000$445.00
4 × 500 GiB$160.0012,000+$240.00 for 48,000$400.00
8 × 250 GiB$160.0024,000+$180.00 for 36,000$340.00
20 × 100 GiB$160.0060,000+$0.00$160.00

Same 2,000 GiB, same 60,000 IOPS, $160 or $445 depending only on how the bytes are divided. The $285 difference is precisely the marginal IOPS bill that twenty free baselines replaced. To spend it, the volumes have to be striped with LVM or RAID 0, because a filesystem on one volume is still capped at 80,000 IOPS however many volumes exist beside it. Striping trades independent failure domains for aggregate throughput — one lost volume takes the whole fileset — which is why the pattern fits replicas, caches and scratch space better than primary data.

The caveat that keeps this honest: a workload already using 80,000 IOPS on one volume has no reason to be split. The saving only materialises when the target exceeds what a single volume can deliver, or when the fleet is small enough that its baselines are idle. Twenty volumes each running at 200 IOPS is $160 a month spent to acquire throughput nobody uses.

io2 lost the ceiling argument as well as the price one

The obvious question after seeing these numbers is whether io2 is ever worth it, since its IOPS rate is thirteen times the gp3 marginal rate. The old answer was that io2 buys a ceiling gp3 cannot reach. That answer expired: gp3 now provisions up to 80,000 IOPS, which is above the 64,000 IOPS that a standard io2 volume tops out at, so io2 loses on cost and on throughput at the same time.

Put the two side by side on the same 500 GiB volume. gp3 at 80,000 IOPS costs $425.00 a month. io2 at 64,000 IOPS — fewer operations, on a smaller ceiling — costs $3,598.50. The premium is 8.5× for a slower maximum, which means io2 is no longer a performance purchase at all.

io2 provisioned IOPSTier rateIOPS charge / monthPlus 500 GiB storage
3,000$0.065$195.00$257.50
16,000$0.065$1,040.00$1,102.50
32,000$0.065$2,080.00$2,142.50
64,000$0.065 / $0.0455$3,536.00$3,598.50
128,000$0.065 / $0.0455 / $0.03185$5,574.40$5,636.90

The tiers are the only place in EBS pricing where the marginal rate falls with volume, and they only engage above 32,000 IOPS. They buy less than they look like they buy: at 128,000 IOPS the average rate has fallen to $0.0436 per IOPS-month, still nearly nine times gp3's $0.005, for capacity that requires io2 Block Express and a volume of at least 256 GiB.

Three things still justify the premium, and none of them is throughput. Durability: io2 is designed for 99.999% annual durability against 99.8–99.9% for gp3, a hundredfold difference in volume-affecting failure rate. Latency: Block Express is designed for an average under 500 microseconds on 16 KiB I/O, against single-digit milliseconds for gp3, and it reduces the frequency of I/Os exceeding 800 microseconds by more than tenfold. A ceiling above 80,000: Block Express reaches 256,000 IOPS and 4,000 MiB/s, which is the only EBS configuration that serves a workload gp3 cannot.

For a database with replayable logs and hourly snapshots, the durability difference is a rounding error on the risk register. For a volume holding the only copy of something, it is the entire argument. That is a narrower case than it was two years ago, and it is the case io2's price actually reflects.

Block size decides which ceiling you hit

IOPS and throughput are provisioned independently on gp3, which means the workload's block size determines which of the two limits binds first. Multiplying 80,000 IOPS by the block size gives the throughput that ceiling implies, and comparing it against 2,000 MiB/s tells you which number to stop paying for.

Block sizeThroughput implied at 80,000 IOPSBinding limitIOPS actually delivered
4 KiB312.5 MiB/sIOPS80,000
8 KiB625 MiB/sIOPS80,000
16 KiB1,250 MiB/sIOPS80,000
25.6 KiB2,000 MiB/sBoth, exactly80,000
64 KiB5,000 MiB/s, 2.5× over the capThroughput32,000
256 KiB20,000 MiB/s, 10× over the capThroughput8,000

25.6 KiB is the exact crossover: 80,000 × 25.6 KiB is 2,000 MiB/s with nothing left over, which is why AWS quotes that I/O size beside the gp3 maximum. Below it, buying throughput is waste, because the IOPS cap arrives first. Above it, buying IOPS is waste, because the throughput cap arrives first. A 256 KiB workload with 80,000 provisioned IOPS delivers 8,000, so nine tenths of the $385.00 monthly IOPS charge — $346.50 — buys capacity the volume physically cannot deliver.

This is also the reason the two dials cannot be sized independently in practice. Throughput is provisionable at 0.25 MiB/s per provisioned IOPS, so the pair moves together unless you deliberately leave one side idle, and the ratio is the I/O size the volume is designed around.

What this page does not price

Snapshots. Snapshot storage is a separate meter, priced per GiB-month of changed data rather than anything provisioned. It is modelled on the snapshot cost page, because the growth curve is nothing like the volume cost curve.

Orphaned volumes. A volume is billed until it is released, not until its instance is terminated. A DeleteOnTermination flag set to false — the default for volumes created separately and attached afterwards — leaves a full-price volume behind after the instance is gone. This is the most common way a fleet of forty volumes becomes a fleet of sixty.

Fast Snapshot Restore. $0.75 per DSU-hour, billed for as long as the restore option stays enabled, not for the restores performed. Enabling it on a volume that snapshots daily is a recurring charge that reads like a one-off.

Volume initialisation. Restoring from a snapshot in the background costs nothing extra but delivers first-touch latency until the blocks are pulled. Paying the provisioned-rate fee to pre-warm is a performance decision with a price, and it is not included here.

Free tier. 30 GiB of storage is included per account per month, which is deliberately not subtracted, for the same reason as everywhere else on this site: a projection that assumes the free tier stops being a projection the moment the fleet grows.

Pricing Basis, Sources & Assumptions

Every rate on this page is a published vendor list price — no negotiated discounts, private pricing, savings plans, credits or enterprise agreements. Totals are in USD and exclude tax. Rates were last checked against the sources below on .

Pricing inputBasis used on this page
SourceAmazon EBS pricing for the volume type definitions and billing rules, cross-checked against the AWS price list feed for EBS for the exact us-east-1 rates: $0.08 per gp3 GiB-month, $0.005 per provisioned IOPS-month, $40.96 per GiBps-month of provisioned throughput (equivalent to $0.04 per MiB/s-month), $0.10 per gp2 GiB-month, $0.125 per io2 GiB-month with tiered IOPS at $0.065, $0.0455 and $0.03185, $0.045 per st1 GiB-month and $0.015 per sc1 GiB-month.
RegionUS East (N. Virginia), us-east-1. Other regions are priced higher and the difference is not modelled here.
CurrencyUSD — list price, tax excluded
Last checked — all seven rates plus the gp3 ceilings (3,000 IOPS and 125 MiB/s included, 80,000 IOPS and 2,000 MiB/s maximum), the 500 IOPS-per-GiB ratio and the 0.25 MiB/s per IOPS throughput ratio were re-read from the gp3 volume type documentation and the price feed on this date.

Modelling assumptions