Compute behind the meter

A reference for the engineers, insurers, regulators and operators evaluating the BayesGrid system.

Paste the whitepaper draft here. Sections below are scaffolded with the IDs the sidebar TOC expects — replace each section body with the real copy and the layout will pick it up automatically.

01 · Executive summary

Demand for compute is surging, driven by AI, while the power infrastructure needed to supply it is increasingly constrained. In Great Britain the gap is unusually sharp: building centralised data centre capacity is slow and capital-intensive, and projects are gated less by silicon than by the grid. Connection offers have been pushed years into the future, and the queue to connect has, until recently, run to roughly four times what the country is expected to need by 2030.

BayesGrid takes a different path. Rather than waiting for new hyperscale sites and the transmission reinforcements they require, BayesGrid uses existing behind-the-meter electrical headroom on managed UK estates to deploy distributed AI infrastructure. The unit of deployment for the proof of concept is the estate: build-to-rent neighbourhoods and managed residential communities served by professionally operated landlord supplies, with some larger sites also operating dedicated on-site substations. These environments provide the practical combination of available capacity, centralised site control and a single counterparty needed for fast early deployment.

Each BayesNode pairs a high-performance compute module with circuit-level energy visibility and control, battery backup, and — where advantageous — on-site solar. Workloads are coordinated through BayesOS, which schedules and routes jobs across the fleet by latency need and each BayesNode's real-time energy and hardware capacity. When grid or site conditions require it, compute load can be reduced or shifted, helping the site operate within its available capacity and potentially supporting wider system needs where participation arrangements allow.

BayesGrid is not intended to replace hyperscale data centres, which remain essential for large-scale AI training. It complements them by adding dispatchable, geographically distributed capacity well suited to AI inference, cloud gaming, content streaming and rendering — workloads where rapid deployment, locality and latency matter. The plan is to prove the model on a small number of UK estates, beginning with a 50-node proof of concept, then scale toward an annual capacity in excess of 100 MW of AI inference compute — roughly 1.5% of the UK's stated 2030 AI-capable ambition, achieved without breaking ground on a single new hyperscale site.

02 · The challenge: compute vs power

2.1 Exploding compute demand

AI training models continue to grow rapidly, pushing the limits of even the most advanced data centre hardware. At the same time, inference workloads — including retrieval-augmented generation and cloud gaming — are consuming GPU capacity at record rates as services reach mass adoption. The UK government's own assessment is that demand for compute at the AI frontier could rise by orders of magnitude before the end of the decade, and that the country will need at least 6 GW of AI-capable data centre capacity by 2030 — roughly three times today's installed AI capacity.

As model architectures become more modular and more efficient, a growing share of inference is better suited to distributed capacity placed closer to users and data. That shift favours the kind of geographically distributed, inference-oriented capacity BayesNodes are designed to provide.

UK data centre demand

Connected data centre capacity today is about 2.4 GW (~7.6 TWh/yr), concentrated around London. The system operator's central forecast sees this reaching 5.2 GW (range 3.7–6.3 GW), roughly 20 TWh/yr, by 2030, and 30–71 TWh by 2050. Network operators have reported 72.8 GW of new data centre connection requests through 2039 — a step change the existing build-out cannot absorb on its own.

2.2 Grid capacity and connection limits

Despite planned generation build-out, GB compute demand is projected to outpace deliverable power for years. The constraint is rarely total generation — it is the network. New compute needs large connections, from tens to hundreds of megawatts, plus transmission upgrades, substations and transformers. Permitting, equipment lead times and the connection queue have been extending timelines by years.

Great Britain's connection queue is the clearest symptom. Before reform, the pipeline of generation and storage projects waiting to connect exceeded 700 GW — about four times what the country is projected to need by 2030. The demand side tells the same story: requests to connect new electrical load rose from 41 GW in late 2024 to roughly 125 GW by mid-2025, of which around 50 GW was data centres alone. For context, GB peak demand on a cold February day in 2026 was about 45 GW.

Reform is now under way. Under the new "first ready, first connected" regime, the system operator confirmed a prioritised pipeline of 381.5 GW in December 2025 — 283 GW of generation and storage plus 99 GW of transmission-connected demand. But two facts stand out for anyone trying to add compute quickly: the battery and solar queues are effectively full or oversubscribed in many regions, and large new demand connections are the next target for reform rather than a solved problem.

The build-out paradox — and why behind-the-meter is the fast lane.

Compute demand is scaling quickly. Even where data centre construction can keep pace, the power-delivery infrastructure to feed it can only scale at network build-out speed — and GB networks have been asked to roughly double transmission build rates and expand regional networks several-fold within this decade. The practical consequence for compute buyers is that capacity is not always available where and when it is needed. Headroom that is already connected is, for now, the fastest route to new compute.

03 · What is BayesGrid

BayesGrid is a distributed AI infrastructure platform built from a fleet of BayesNodes deployed on managed UK estates. It uses available electrical headroom at the edge to host high-demand compute capacity by pairing GPU capacity with real-time energy visibility and control. Workloads are coordinated through BayesOS, which connects compute offtake customers to the fleet and makes many independent Nodes behave like a coherent platform.

Each BayesNode integrates GPU compute modules with a high-efficiency cooling system and a local management gateway, paired with circuit-level energy control and battery storage at the estate level. BayesGrid is modular: BayesNodes are repeatable building blocks that can be added incrementally across a fleet, and capacity within each BayesNode scales through additional compute modules. The same core architecture and BayesOS are intended to support different deployment sizes, from a single residential block to small-business or campus-style configurations.

Energy control is what makes the headroom usable. It provides the real-time, circuit-level visibility needed to provision BayesNodes as always-on loads within the estate's verified capacity, while continuously monitoring site conditions and holding safe operating limits. Battery storage safeguards workloads during disturbances, buffers short-term demand spikes, and lets the site respond to system demand-response events. On-site solar can be integrated where it improves the economics or the site's resilience.

At the fleet level, BayesOS schedules and routes workloads across BayesNodes based on customer needs (including latency), GPU and system availability, and real-time energy constraints at each site. It also coordinates graceful failover, so service stays reliable even when individual BayesNodes must throttle or pause during a grid event.

04 · The Great Britain unlock

4.1 Unlocking existing electrical headroom

Electrical infrastructure is designed for rare peak conditions, while typical use sits far lower — leaving substantial capacity underused. In Britain, BayesGrid begins with the managed estate, where professionally operated landlord supplies and centralised site control make early deployment simpler at proof-of-concept stage. Estates provide a practical starting point for deploying meaningful capacity quickly, while leaving other settings open over time.

Why the estate is the unit of deployment

Managed UK estates typically have a landlord supply served by a three-phase connection, and some larger schemes also operate dedicated on-site substations. A three-phase supply rated to 80 A per phase provides about 55 kVA sustained (69 kVA cyclic), while larger estate connections can scale materially beyond that. For the proof of concept, that single, professionally managed connection point provides the practical combination of electrical capacity, site control and a single counterparty needed for deployment.

4.2 Making headroom usable for always-on compute

Headroom only becomes useful when it can be operated safely and predictably. BayesGrid achieves this by combining circuit-level visibility and control, battery-backed resilience and fleet orchestration. BayesNodes are provisioned to operate within site-assessed operating limits and agreed safety margins. If rare peaks occur, BayesGrid preserves BayesNode operation by first drawing on storage, and in extreme cases by temporarily trimming non-critical flexible loads. BayesNode interruption occurs only during defined power events — a grid outage, a demand-response event, or a safety-triggered shutdown — in which case BayesOS coordinates controlled failover to other BayesNodes in the fleet.

4.3 Assurance and standards in Great Britain

For a behind-the-meter compute load to operate as infrastructure, it must satisfy Great Britain's electrical assurance framework — a layered, well-established stack that installers, network operators and insurers already recognise:

• ENA Engineering Recommendations G98 / G99 / G100 — network operator approval for connecting generation and storage, and for export/import limitation.
• BS 7671:2018+A4:2026 (the IET Wiring Regulations, in force from April 2026) — including the chapters covering stationary battery installations and prosumer low-voltage installations.
• IEC 62477-1 and IEC 62933-5-2 — product-level safety for power-conversion equipment and grid-integrated energy storage.

All equipment is UKCA-marked. Together these provide the framework for establishing safe operating limits, controlled export, and battery and converter safety — the assurances a serious estate operator and its insurer will expect before an always-on load is energised.

4.4 Rapid deployment at scale

Because BayesGrid uses existing behind-the-meter capacity, capacity can be deployed incrementally — BayesNode by BayesNode and estate by estate, in parallel — without waiting for large centralised site development or a new high-capacity connection for each deployment. The modular design supports several configurations while keeping one operating model: standardised BayesNodes, consistent security controls, and the same grid-aware operating policies across every site, all coordinated through BayesOS as a single distributed infrastructure platform.

05 · Addressing the big questions

5.1 Up-time, reliability and continuity

BayesNodes are intended to operate as infrastructure rather than best-effort capacity. Each deployment includes energy storage to support ride-through, orderly shutdown and workload continuity across the fleet. When a grid power event occurs, the BayesOS coordinates failover of workloads where capacity is available elsewhere in the fleet, powers the affected Node down safely, and — on residential estates — can preserve battery capacity for site-critical needs. Storage materially improves continuity through short disturbances, although overall service levels will depend on fleet scale, site configuration and the nature of the event.

5.2 Sustainability and grid impact

The core constraint on new UK compute is power delivery, not silicon. BayesGrid expands compute by using existing grid infrastructure more effectively rather than waiting for new large-scale connections. Just as importantly, the fleet can be designed to reduce or shift compute load during periods of stress, while distributed storage provides local buffering. That creates a potentially useful fit with Great Britain's flexibility markets — including the Demand Flexibility Service, the Balancing Mechanism and network operators' regional flexibility services — though participation will depend on site configuration, aggregation model, metering and market rules. For the system operator and networks, a controllable load may be easier to integrate than inflexible new demand, particularly where it can respond to local constraints or time-varying system conditions.

5.3 Estate and resident value

BayesGrid aims to install the energy and compute system at no capital cost to the estate, while compensating the operator for hosting. Depending on the commercial structure and site configuration, that value could help offset communal energy or connectivity costs and improve resilience during outages. Residents may also benefit from a modern energy-management foundation and, where designed into the site, battery capacity that remains available to support essential services during an outage. The proposition is intended to be predictable and host-friendly, with resident energy use protected by design and by operating limits.

5.4 Security and trust

BayesGrid is being designed for enterprise-grade reliability and security in a residential setting, spanning customer workloads, fleet operations and physical hardware. The BayesOS is intended to enforce encrypted communications, strong identity and access controls, and workload isolation suitable for multi-tenant environments, with operational monitoring and audit logging across the fleet. Physical security is addressed through secure enclosures, monitoring and tamper-evident protections; if a device is removed or compromised, BayesGrid can revoke credentials and disable access to workloads and data. The detailed control framework, certifications and operating assurance model will need to be validated as the platform moves through pilot and production deployment.

06 · Target workloads

BayesGrid is best suited to workloads that can tolerate distribution across smaller hardware footprints, benefit from rapid incremental capacity, and derive value from proximity to users or data. That makes it a plausible complement to large data centres for selected inference-heavy and latency-sensitive workloads, rather than a universal substitute for centralised compute.

• AI inference + RAG: ≈19–20% CAGR (global inference market to 2030). The workload most aligned to modular, distributed capacity.
• Cloud gaming: ≈43% CAGR for the UK market to 2030. Proximity directly improves user experience.
• Rendering + streaming: ≈8% CAGR (global). Latency and locality reduce backbone congestion.
• HPC + simulation: ≈7% CAGR (global), alongside extraordinary growth in UK public AI compute.

As AI architectures become more modular and more efficient, an increasing share of inference can be served by distributed capacity placed closer to users. BayesGrid adds exactly that — geographically distributed, inference-oriented capacity — without requiring new large-scale connections. It is not intended to replace hyperscale training infrastructure: the largest training workloads need single-site cluster scale and tightly coupled networking best delivered in centralised facilities.

07 · Concept and architecture

7.1 BayesNode architecture

A BayesNode is a repeatable deployment unit: one or more GPU compute modules, a local management gateway, and thermal management designed for reliable, low-impact operation in a residential setting. Compute modules are purpose-built — each equipped with enterprise-grade GPUs plus ample CPU, memory, high-speed storage and networking — and are liquid-cooled for quiet operation and improved reliability. BayesNodes are installed alongside circuit-level energy control and an energy storage system, and can integrate on-site solar where advantageous and approved. The configuration supports always-on compute within the estate's verified capacity while providing resilience during outages and short disturbances.

7.2 Fleet orchestration and power events

At the fleet level, BayesNodes connect through a low-latency network fabric and are coordinated by BayesOS, which schedules and routes workloads based on customer needs, BayesNode availability and health, and real-time grid constraints at each site. During a power event — a grid outage, a demand-response instruction, or a safety-triggered shutdown — site constraints are surfaced to the cloud and BayesOS coordinates the response: workloads fail over to other BayesNodes and affected BayesNodes power down safely. A local safety interlock can trigger shutdown directly at the panel if the network or cloud is unavailable and a local safety issue arises.

7.3 Why this architecture matters

Together these elements create an infrastructure layer that improves how compute and power are delivered at the edge. Nodes scale incrementally across geographies, reduce reliance on single large sites, and increase resilience by distributing capacity across many independent deployments — all while using existing grid infrastructure more effectively.

08 · Core value propositions

BayesGrid uses underutilised behind-the-meter capacity to deploy a managed, geographically distributed compute layer, while improving resilience through local storage and BayesOS. The value lands across the ecosystem:

Compute offtake customers — rapidly deployable GPU capacity delivered as a resilient, secure, scalable fleet, with workloads placed closer to end users and a path to add capacity without long-lead site development or connection timelines.

Estate operators (Node hosts) — energy and compute infrastructure installed at zero capital cost, host compensation, improved outage resilience, and a modern energy-management foundation, with resident energy use protected.

System operator and networks — better use of existing infrastructure, distributed storage at the edge, and a controllable load that curtails during stress, easing peak-driven planning and improving return on grid investment.

Government and society — support for AI-driven compute demand while reducing reliance on slow, capital-intensive centralised build-out alone, and a more geographically balanced, resilient infrastructure base.

09 · Next steps

BayesGrid's near-term emphasis is not on scaling prematurely, but on maximising speed of learning through rapid, system-level iteration across hardware, installation workflows, orchestration software, reliability and customer experience.

First proof of concept: a 50-node deployment

The plan is to prove the model on a small number of managed UK estates with a 50-node proof of concept, partnering with a leading build-to-rent operator and, as a second setting, a managed retirement-living community. These estates are intended to provide the combination of landlord-supply capacity, operational control and host profile needed for a fast, informative early deployment. Following the proof of concept, BayesGrid will prioritise new-build and build-to-rent deployments as the path to broader GB rollout, while developing retrofit options for existing estates and larger configurations for small-business and campus sites.

Beyond the pilot, BayesGrid's ambition is to scale toward an annual capacity in excess of 100 MW of AI inference compute placed behind the meter on UK estates. In context, that is roughly 1.5% of the 6 GW of AI-capable capacity the UK says it will need by 2030, and a meaningful fraction of a single AI Growth Zone. The pace and shape of that expansion will depend on pilot results, customer demand, host adoption, economics and regulatory execution.

10 · Risks

BayesGrid is a new infrastructure model, approached with disciplined risk reduction through staged pilots and measurable milestones. Key risk areas:

• Product-market fit — will compute customers commit to capacity at pricing and terms that support a durable business?
• Technical performance and reliability — can the fleet consistently deliver the performance, security, uptime and operational safety expected of infrastructure?
• Deployment and operations — can Nodes be installed, serviced and monitored at scale with a host-friendly experience and sustainable unit economics?
• Regulatory and network variability — how do connection rules, flexibility programmes and network operator processes affect deployment, operating modes and economics across GB regions?
• Supply chain and manufacturing — can components and build capacity be secured predictably as demand scales?

These will be mitigated methodically through the proof-of-concept phases, iterating with real customers and partners and expanding scope only as reliability, economics and operational readiness are demonstrated.

11 · Conclusion and call to action

BayesGrid is a pragmatic response to the growing imbalance between compute demand and power-delivery capacity — an imbalance that is especially acute in Great Britain, where the grid, not the chip, is often the binding constraint. By deploying distributed compute Nodes on managed estates, paired with circuit-level energy control, storage and optional solar, BayesGrid seeks to convert underutilised, already-connected headroom into deployable, managed compute. It is intended to complement centralised data centres by enabling faster, more geographically distributed expansion that aligns more closely with grid reality.

Access to compute is becoming a strategic constraint across industries and countries. BayesGrid offers Britain one possible path to deploy selected capacity faster, closer to users, and in a way that is more responsive to grid constraints than a purely centralised model. The opportunity now is to test that thesis with the right ecosystem of investors, compute platforms, estate operators, network participants and policymakers.

The demand is real. The grid constraints are real. BayesGrid is designed to bridge the gap.

We are seeking partners to help test and build BayesGrid in Great Britain: investors to fund proof-of-concept execution and early scale-out; compute platforms interested in pilot and production offtake; estate operators willing to host early deployments; and network, flexibility and policy stakeholders interested in shaping how distributed compute can participate safely and effectively in the power system.

12 · Sources & assumptions

The figures in this paper are drawn from public 2024–2026 sources. Forecasts are presented as ranges where the underlying source does so, and should be refreshed as the next system-operator forecasts and the GB demand-connection reforms land in 2026.

Note on growth rates: where a Great Britain-specific market size exists (for example UK cloud gaming) it is used; where only global figures are available (cloud rendering and streaming, the broad inference market, HPC) they are labelled as global. Market-research growth rates vary by provider and should be treated as indicative.