ROI · Per petabyte, per year

Fewer bytes.
Every budget gets smaller.

Bytes removed at encode stay removed in storage, egress, memory and capacity. Here is what that is worth.

30–80%
of the bytes removed
Measured
$38k–$202k
saved per PB a year, storage alone
Modeled
2.0 yrs
of growth headroom at 40% a year
Modeled
22 days
to pay back converting the archive
Modeled

See your number

Your estate, your bill.

Start from
Bytes removed
What you pay for storage
How often the archive is read
Saved every year
$1.48M
49% of a $3.02M storage and egress bill
Net, year one
$1.26M
ROI over ten years
9.4×
Conversion paid back in
22 days
Net over ten years
$13.2M

The annual bill, before and after

StorageEgressBitruvius licence
View as table
StorageEgressBitruvius licenceTotal
Today$2.52M$500k$0$3.02M
With Bitruvius$1.29M$255k$150k$1.69M

Year one, step by step

SavedPaidNet
View as table
Year oneAmount
Storage saved$1.23M
Egress saved$245k
Licence paid−$150k
Conversion (once)−$73k
Net, year one$1.26M

Modeled Storage and egress only, from AWS us-east-1 list ($252,000 per PB-year on S3 Standard, $50,000 per PB read out). Licence $15,000 per PB a year; conversion $7,300 per PB. Compute and deferred capacity are not counted, and both are larger.

What a saved byte is worth

A byte kept out of RAM is worth 137× one kept off object storage.

Bitruvius formats are read compressed, so the saving follows the byte into memory, not just onto disk.

Cost of one GB for a year, by where it lives · log scale

View as table
Where the byte lives$ per GB-year
Glacier Deep Archive$0.0119
Glacier Instant Retrieval$0.0480
Egress, read once a year$0.0500
S3 Standard-IA$0.150
Egress, read 5× a year$0.250
S3 Standard$0.252
RAM, resident$34.44

AWS us-east-1 list, October 2026. RAM is marginal r7i over m7i at equal vCPU. Reading an archive five times a year costs about what storing it does, so for served data the egress saving roughly doubles the storage saving.

Bytes on disk
Measured
$38k–$202k
per PB a year
120 of 120 wins vs Esri LERC · −49% on NAIP
Egress
Measured
$15k–$200k
per PB a year, by read rate
Falls by the same 30–80%
RAM / working set
Modeled
137×
the value of a storage byte
Read compressed; resident set shrinks with the file
Decode
Measured
1.6×
Esri LERC’s median decode speed
Up to 18.6 GB/s on basemap terrain · 17–25× faster point-cloud reads · ~32× in-browser
Encode
Modeled
<$8k
per PB, one time, all-in
Embarrassingly parallel, on spot compute
Energy & facility
Modeled
Largest
and hardest to bank
Capacity not provisioned, racks not filled

Three customers, end to end

Millions a year, from storage and egress alone.

Hyperscaler tier

100 PB · 1 read a year · 50% off list

$8.62M/ yr
saved at −49%, of $17.6M spend
$5.28Mrange at −30% to −80%$14.1M
One-time conversion $730k

Imagery provider

10 PB · 3 reads a year serving tiles · 30% off list

$1.6M/ yr
saved at −49%, of $3.26M spend
$979krange at −30% to −80%$2.61M
One-time conversion $73k

Sovereign / defence

5 PB · 1 read a year · list as on-prem TCO proxy

$740k/ yr
saved at −49%, of $1.51M spend
$453krange at −30% to −80%$1.21M
One-time conversion $37k

Return on what they pay us, by year

Hyperscaler tierImagery providerSovereign / defence
View as table
HorizonHyperscaler tierImagery providerSovereign / defence
1 year3.9×7.2×6.6×
2 years4.6×8.6×7.9×
5 years5.2×9.7×9.0×
10 years5.5×10.2×9.4×

Modeled The hyperscaler's lower multiple is a pricing artifact: it already buys storage at half of list and pays the same licence. Priced off its real cost basis, it lines up with the others.

Net dollars by year
HorizonHyperscaler tierImagery providerSovereign / defence
1 year$6.39M$1.38M$628k
2 years$13.5M$2.83M$1.29M
5 years$34.9M$7.17M$3.29M
10 years$70.5M$14.4M$6.61M

ROI over time

The return improves every year.

One petabyte, ROI multiple

View as table
HorizonROI
1 yr6.6×
2 yr7.9×
5 yr9.0×
10 yr9.4×
Saving, cost and net
HorizonSavedCostNet
1 yr$147,980$22,300$125,680
2 yr$295,960$37,300$258,660
5 yr$739,900$82,300$657,600
10 yr$1,479,800$157,300$1,322,500

One petabyte, cumulative net benefit as the estate grows

Estate flat25% a year40% a year
View as table
YearEstate flat25% a year40% a year
Yr 1$126k$126k$126k
Yr 2$259k$292k$312k
Yr 5$658k$1.08M$1.45M
Yr 10$1.32M$4.41M$9.28M

Modeled Per-petabyte pricing rides the estate's growth; a flat licence would capture none of it. Discounted at 10%, the ten-year case is still 9.1× ($909k saved for $99k).

Capacity

Two more years before the next build.

Megawatts, not money, are the scarce resource. Halving the bytes turns a capacity problem into a software upgrade.

100 PB estate growing 40% a year against 200 PB provisioned

As isWith Bitruvius Provisioned, 200 PB
View as table
DateAs isWith Bitruvius
2026100 PB51 PB
2027140 PB71 PB
2028196 PB100 PB
2029274 PB140 PB
2030384 PB196 PB

Headroom bought by −49%, by growth rate

View as table
GrowthHeadroom
25% a year3.0 yrs
40% a year2.0 yrs
60% a year1.4 yrs
Deferring a 10 MW build two years
$17M–$21M
present value at 10%, before servers and power
Modeled

Scope: Bitruvius removes 30–80% of the bytes inside the spatial estate. Spatial data is a single-digit share of what hyperscalers host; this is not a claim on the data-centre power problem.

Switching

Paid back in weeks. The first step needs no migration.

Converting one petabyte

Encode, requests, verifyEngineeringSecond copy during cutover
View as table
Encode, requests, verifyEngineeringSecond copy during cutoverTotal
Convert in slices$141$4k$3.1k$7.3k
Full dual-run for a quarter$141$4k$63k$67k

Modeled About 1,389 core-hours at an assumed 200 MB/s per core on spot. The cost is how long two copies are held.

Slices
22 days
payback
Full dual-run
6.5 mo
payback
Or skip conversion: 8 drop-ins
TurboLERCTurboWebPTurboJXLTurboLZWTurboLZ4TurboZstdTurboLEPCCTurboSPZ

Same container, catalogue and pipeline. A configuration change, measured on your own data inside a week.

Payback against the per-PB storage saving at list ($123,480 a year).

For investors

Revenue rides the byte count.

ARR at $15,000 per PB a year, by petabytes under management

View as table
Under managementARR
500 PB$7.5M
1,000 PB$15M
5,000 PB$75M
10,000 PB$150M
25,000 PB$375M

One large imagery provider runs 10–30 PB; one hyperscaler's spatial tier plausibly 50–200 PB. The upper rungs come from signed estates growing 30–50% a year, not from new logos.

8.2×
what the customer keeps for every $1 of licence, storage alone
85–95%
gross margin: software on the customer’s own infrastructure
30–50%
annual growth in the byte count the licence is priced on

What we do not claim

Not the power problem
30–80% of the bytes inside the spatial estate. Nothing wider.
List price overstates a hyperscaler
Its cost basis is far below list. Its reason to care is workload capture.
The range is content
30–80% spans data types. Quote −49%, then measure your corpus.
Decode pays only where it binds
Network- or I/O-bound pipelines see the byte saving, not a compute saving.

Method

Measureda reproducible harness on a named corpus Modeledderived from the measurements and the inputs below

Measured: the byte reductions, decode throughput, 120 of 120 wins vs Esri LERC, −49% on NAIP, 17–25× point-cloud reads, ~32× in-browser decode. Every dollar, year of headroom and present value on this page is modeled.

Inputs
InputValueBasis
Petabyte1,000,000 GBDecimal, as cloud providers bill
S3 Standard, over 500 TB$0.021/GB-monthAWS S3 pricing, us-east-1
Egress, over 150 TB/month$0.05/GBAWS, same region
Marginal RAM$2.87/GiB-monthr7i vs m7i at equal vCPU
Spot vCPU$0.0447/vCPU-hourc7i.large
Encode throughput200 MB/s per coreAssumption, not measured
Licence$15,000 per PB-yearPricing model
Data-centre construction$10–12M per MW2026 US standard builds
Cost of capital10%For deferral and discounting
Spatial data growth25–60% a yearA bracket, not a forecast

Measure it on your own data.

A drop-in pilot gets a measured result on your corpus inside a week.