Sustainability by architecture, not by promise.
At SkyeConnex, sustainability comes from the same architecture that delivers sovereignty. It isn’t a separate initiative. Data is compressed, encrypted and adaptively sharded across infrastructure that already exists. AI compute goes to the data instead of the data being copied to the compute. Each of those decisions reduces footprint, and none of them costs security or resilience.
Resilience that costs less than a single copy.
Default 5+2 sharding with compression delivers jurisdiction-resilient storage at roughly 0.7× the original data volume, versus about 3× for conventional replicated copies. The data survives the loss of two locations while using less raw capacity than a single uncompressed copy.
Overhead by profile, after 2:1 compression. Three of the four SkyeConnex profiles sit at or below the size of a single unprotected copy.
| Profile | Raw overhead | At 2:1 compression | Locations it can lose |
|---|---|---|---|
| 5+2 (default) | 1.40× | 0.70× | 2 |
| 8+3 | 1.38× | 0.69× | 3 |
| 4+4 (high assurance) | 2.00× | 1.00× | 4 |
| 3-way replication (industry baseline) | 3.00× | 3.00× | 2 |
Compression ratios vary by data class. Structured records, logs and databases compress well. Media, Office formats and pre-encrypted content are not compressed at all — we skip them deliberately, because compressing something already compressed spends CPU for no gain.
Where the savings actually come from.
None of these is an offset or a pledge. Each one is a consequence of how custody already works.
Right-sized redundancy.
Redundancy is spent where classification or regulation requires it, not applied uniformly to everything.
Data is compressed before encryption and erasure coding, so storage, network transfer and endpoint I/O all shrink before anything leaves the client boundary. The coding profile is chosen per upload, and every file carries its own.
Build nothing new.
SkyeConnex layers custody onto storage and compute customers already have. No new data centres, no dedicated hardware, no single provider.
Because placement is portable, you can shift data toward lower-carbon infrastructure as your options improve, without re-platforming. Endpoint carbon posture shows you where those options are — and shard migration is how you act on it.
Compute goes to data.
SkyeGXU brings models to governed data instead of copying datasets into central clusters.
- Runs on hardware already deployed — no new accelerator capex, no new embodied carbon.
- Outputs return into custody, preventing duplicate-dataset sprawl.
- Compression reduces transfer on every federated round.
- Heavy workloads run in confidential-computing enclaves — data-centre efficiency, inside the custody boundary.
Carbon you can audit.
SkyeConnex records the carbon-intensity band of the electricity grid behind every endpoint your data is placed on. The band appears beside that endpoint’s other posture attributes, such as jurisdiction and assurance level. Where we don’t yet know an endpoint’s band, we say so plainly rather than estimating.
Bands on one axis. “Low” is under a fifth of the scale; “high” is half of it. Endpoints with no known band sit off the scale entirely, by design.
| Band | gCO2eq/kWh | Typical grid | Examples |
|---|---|---|---|
| Low | < 150 | Hydro, nuclear, wind | Quebec, Ontario, France, Sweden |
| Moderate | 150–400 | Gas plus renewables | Virginia, Ireland, Germany, UK |
| High | > 400 | Coal or oil | Alberta, Ohio, Poland, South Africa |
| Not assessed | — | No band | Personal cloud, multi-region |
Bands are keyed on cloud region, not on sovereignty zone, because grids don’t follow borders. In Canada, ca-central-1 is low and ca-west-1 is high — roughly an order of magnitude apart inside one jurisdiction. An automated test pins that distinction so a future refactor can’t collapse the two.
The product states each of these itself, alongside the band.
A page that publishes the limits of its own metric is rare. These are the limits.
-
Average, not marginal
It reflects what the grid emitted overall, not what our added load caused. Marginal intensity is what you’d optimize against, and we don’t hold it.
That’s one reason the band is a reported attribute, not a scheduler. -
Annual, not real-time
Intensity swings hourly. An annual average describes an endpoint’s ordinary posture, not any particular moment.
-
Coarse boundaries, on purpose
A region drifting from 390 to 410 shouldn’t silently reclassify an endpoint in your report, and the published source figures are approximations anyway.
-
Not a measurement of the endpoint
It characterizes the electricity market the endpoint draws from, which is the finest resolution the region supports.
-
“Not assessed” is never a default
A region missing from the table stays empty rather than falling back to moderate.
It’s shown as a gap in what we know, not a finding that the grid is clean.
Five claims, and what each one rests on.
- 01
Sovereign resilience at less than the footprint of one uncompressed copy.
0.70× stored volume at default 5+2 with 2:1 compression, surviving the loss of two locations.
- 02
Redundancy where policy requires it, not everywhere by default.
The coding profile is chosen per upload and carried by each file, so overhead tracks classification rather than a global setting.
- 03
Sustainability by architecture, not by promise.
Every reduction above is a consequence of how custody works, not an offset purchased afterwards.
- 04
Governed AI without new infrastructure or new copies of your data.
Models travel to governed data on hardware already deployed; outputs return into custody.
- 05
Grid carbon you can audit, not carbon we assert.
A band per endpoint, its basis stated alongside it, and unknowns shown as unknown.
Every claim is measured, baselined and scoped to where it holds.
- Numbers
Workload-specific
Compression ratios are quoted per data class and backed by measurement. Structured records, logs and databases compress well; media, Office formats and pre-encrypted content are skipped rather than compressed poorly.
- Baselines
Stated
Footprint comparisons are against replicated resilient storage, not against a single unprotected copy.
- Trade-offs
Honest
High-assurance profiles use more capacity. 4+4 is 1.00× even after compression, and we present that as a deliberate policy choice rather than hiding it behind the headline.
- Small objects
Handled
Small files are handled so erasure-coding overhead doesn’t erode the efficiency claim.
- Security
First
Compress-before-encrypt is covered in our cryptographic review, including length-leakage mitigations for sensitive data classes.
- AI claims
Workload-appropriate
Federated efficiency is claimed for inference, embeddings and fine-tuning. Heavy training routes to enclave infrastructure.
- Unknowns
Visible
Endpoint carbon bands are recorded, not assumed. An endpoint without a known band is shown as unknown, never filled with an estimate.
Frequently asked questions
How can sovereign storage use less capacity than a single copy?
Compression happens before encryption and erasure coding. Default 5+2 sharding carries 1.40× raw overhead, so at 2:1 compression the stored volume is roughly 0.70× the original — while surviving the loss of two locations. Conventional three-way replication is 3.00× and survives the same two losses. Compression is applied only to data that benefits: already-compressed formats such as JPEG, MP4, ZIP and Office documents are skipped, because compressing them costs CPU for no gain.
What is a grid carbon intensity band?
A band describes the carbon intensity of the electricity grid serving a storage endpoint’s region, on annualized average intensity: low is under 150 gCO2eq/kWh, moderate is 150 to 400, high is over 400. Bands are shown beside each endpoint’s other posture attributes such as jurisdiction. An endpoint we have not classified is shown as “not assessed” rather than being given a default value.
Is the carbon band a real-time measurement?
No. It is an annualized average for the electricity market the endpoint draws from, not a live or marginal measurement of that endpoint. Grid intensity swings hourly; an annual average describes an endpoint’s ordinary posture rather than any particular moment. The product states this alongside every band it displays, and the qualification is embedded in exported signed reports.
Does the carbon band affect where my data is placed?
No. The band is a recorded attribute, not a placement rule. Sovereignty constraints gate placement because the failure is legal; carbon has no equivalent, and a control that could refuse an upload for being high-carbon is one operators would disable. The band is recorded, displayed and reported on so you can act on it — including by migrating shards toward lower-carbon endpoints.
Why is one country listed in two different bands?
Because grids don’t follow borders. In Canada, ca-central-1 runs on hydro and sits in the low band while ca-west-1 is gas-weighted and sits in high — roughly an order of magnitude apart inside one jurisdiction. Bands are therefore keyed on cloud region rather than on sovereignty zone, and an automated test pins that distinction so a future refactor cannot collapse the two.