Sustainability Statement

1. About this statement

Arsenale Limited is below the size thresholds that trigger Streamlined Energy and Carbon Reporting (SECR) and the Climate-Related Financial Disclosures regime, and we are therefore not legally required to publish environmental disclosures. We publish this statement voluntarily because the architectural choice at the centre of our product — running AI inference on customer-owned hardware rather than in shared cloud infrastructure — has a substantive environmental story, and our intended customers in UK government, regulated finance, and critical national infrastructure increasingly look for an environmental position as part of supplier due diligence.

This statement covers the Arsenale Appliance product and the corporate operations of Arsenale Limited. It is structured around what is true today; quantitative measurement and externally-verified claims will be added as the company reaches the scale where they are meaningful.

2. Why edge inference is structurally lower-impact than cloud inference

The default deployment pattern for large-language-model inference is the cloud datacenter. A single query travels from the client through a CDN, an API gateway, a routing layer, and an orchestrator, before arriving at a GPU cluster that performs the inference and returns the response back through the same chain. The energy footprint of that round trip includes:

• The inference itself — typically performed on accelerators in the hundreds-of-watts class, often clustered across multiple devices to serve a single query for the largest models.
• Datacenter overhead — cooling, power conditioning, lighting, and physical security. Industry-standard Power Usage Effectiveness (PUE) ratios sit between 1.3 and 1.5, meaning every watt of compute requires 30–50% more for the facility around it.
• An idle baseline — cloud inference infrastructure runs whether you are using it or not. The fixed cost of having the service available is paid 24 hours a day.
• Network energy — the round trip between your premises and the datacenter, often crossing international links.

Edge inference on a customer-owned appliance removes most of those line items:

• Single-machine power draw — the appliance class we use draws on the order of tens of watts at idle and around a hundred watts under inference load. There is no separate accelerator cluster.
• No datacenter overhead — the appliance lives in a normal office environment. There is no PUE multiplier on top of compute energy.
• Sleeps when idle — outside of active use, the appliance can drop to a low-power state. There is no fleet of servers that must remain warm because someone might query in the next minute.
• No round-trip beyond the LAN — queries do not leave the building.

The exact energy delta depends on workload pattern, model size, and the specific cloud configuration being compared against. Honest framing: for typical enterprise patterns — bursty daytime usage, hundreds to low-thousands of inferences per day rather than continuous high-throughput serving — edge inference on a unified-memory appliance uses one to two orders of magnitude less energy per query than equivalent cloud inference. At very high sustained utilisation that gap closes; we are honest that for those workloads cloud may amortise better.

3. The Arsenale Appliance — operational profile

The appliance is built around consumer-grade unified-memory compute platforms, not server-class GPUs. This was a product decision made for sovereignty and TCO reasons; the energy profile is a downstream consequence.

These numbers are typical for the hardware class and not specific Arsenale measurements. We will publish measured values for our shipping configurations when we have a representative sample of in-field deployments to draw from.

4. Hardware lifecycle

Customer ownership of the hardware changes the lifecycle economics in ways that matter environmentally:

• No forced obsolescence cycle. The appliance does not depend on a vendor-side API that can be deprecated to push an upgrade. Software updates extend the useful life of the hardware rather than shortening it.
• Five years of supported software included in the base price. This is the explicit floor for the supported lifespan; in practice we expect units to be useful well beyond it, particularly for inference workloads against models that are themselves stable.
• Repairability where possible. Standard memory and storage modules where the platform allows, sourced through normal trade channels.
• UK WEEE compliance. End-of-life disposal is handled in line with the Waste Electrical and Electronic Equipment Regulations. As volume reaches a level where it is operationally meaningful, we will offer a take-back option to ensure responsible recovery and recycling rather than relying on the customer's own waste contractor.
• Repurposable substrate. The underlying compute platform is general-purpose. End-of-life appliances retain residual value as workstation or developer hardware, reducing the proportion that goes to disposal in the first place.

5. UK operational footprint

At the date of this statement Arsenale operates as a UK-domiciled solo-Director company. The corporate operational footprint is therefore small and specific:

• No corporate office — the registered office is a serviced address; engineering is performed on company-owned hardware in the UK. No commute energy, no commercial-real-estate footprint.
• No corporate vehicle fleet, and minimal business travel to date.
• Static-site delivery via Cloudflare — an edge CDN is among the more energy-efficient delivery mechanisms for a marketing site, because requests are served from a node geographically close to the visitor and content is cached aggressively.
• UK-hosted application infrastructure for the inbound enquiry and reservation APIs — on the UK grid, with the corresponding carbon intensity.
• Local LLM tooling for internal use — we run our own inference for internal drafting, code review, and documentation generation on the same class of hardware we ship to customers, rather than calling out to commercial cloud LLM APIs. This is consistent with eating our own dog food, and it is also lower-impact per query for the workload pattern.

6. Supply chain considerations

The supply chain for the appliance, the cloud subprocessors used for the website and email, and the open-source software ecosystem are described in detail in our Modern Slavery Statement and on the Trust page. The environmental considerations layered on top:

• Hardware is sourced from established suppliers in Tier-1 manufacturing jurisdictions, which generally operate under stronger environmental regulation than alternatives.
• The compact form factor uses materially less metal, plastic, and packaging per unit than rack-server deployments delivering comparable inference capacity.
• Subprocessors (Cloudflare, Google Workspace, Stripe) publish their own environmental statements; each is on a published path to renewable-energy-matched operations and we monitor those commitments as part of subprocessor review.

7. What we do not claim

An honest sustainability statement is as much about what is not claimed as what is. Items below are deliberate omissions:

• We do not claim "carbon neutral". We have not formally measured our operational carbon footprint, and we will not assert a status we have not measured.
• We do not purchase carbon offsets. Many available offsets are non-additional, unverified, or reverse-bought. Until we can identify a class of offsets we genuinely believe deliver atmospheric benefit equivalent to claimed, we prefer the honest position of not buying any.
• We do not market "AI for sustainability". Our product is general-purpose AI infrastructure. Customers may use it for sustainability work, but the claim that the product itself is a sustainability tool would be marketing rather than truth.
• We do not claim a renewable-energy contract at our scale. Our electricity is the UK grid mix as it stands, and we report that honestly.
• We do not claim Scope 3 supply-chain reductions we have not measured. Supply-chain choices are made for sovereignty and labour reasons (see /modern-slavery); environmental benefits where they exist are real but not yet quantified.

8. Targets and review

This statement will be reviewed annually, with the next review due no later than 14 May 2027. Quantitative measurement and externally-verified claims will be added as the company reaches the scale at which they become meaningful and credible.