Where the Intelligence Sits

On-device AI, sovereign data, and the quiet repositioning of the telco gateway.

There is an infrastructure decision facing telcos in the next capex cycle that almost never gets called by its name. It sits beneath the routine business of care, customer experience, and service assurance, and it amounts to a question of where intelligence is allowed to live in the network.

Operators that recognise the question and answer it correctly will own a position in the digital economy that nobody else can credibly assemble. Operators that miss it will be outflanked on cost, on trust, and on access to whatever comes next in the value-added services stack.

The gateway stopped being a router

The modern carrier gateway stopped being a router some time ago. What sits on the wall today is a multi-core ARM computer with several gigabytes of memory, packet-processing offload and, in some newer platforms, dedicated AI acceleration, planted inside a household with stable power, a regulated commercial relationship with the operator and continuous network presence.

The model classes this hardware can host are the ones household intelligence actually depends on. Small classifiers for traffic, device and application identification. Embedding models for similarity and clustering. Distilled task-specific models for QoE scoring, anomaly detection and intent recognition. And, increasingly, compact language models in the one to three billion parameter range that recent Phi, Gemma and Llama families have made genuinely capable for local diagnostics, natural-language interfaces and structured summarisation.

Models in these classes fit in modest memory budgets, run usefully on spare application CPU capacity and, where present, dedicated AI acceleration, while packet-processing offload preserves the gateway's core routing function.

Recent engineering work by Silpion and Presciense has benchmarked this configuration on a representative MediaTek MT7988-class platform, as found in current premium carrier CPE. A quantised Qwen 2.5 0.5B runtime, configured for short household diagnostic interactions, sits alongside a compact Seq2Seq non-intrusive load monitoring model for household energy disaggregation. Combined memory consumption is under 600 megabytes. Sustained throughput in the region of ten tokens per second is sufficient for the household-facing workload, which is event-driven rather than conversational. The MediaTek Network Processing Unit on the chipset preserves packet-processing performance and handles QoS and tunnelling offload, which is its actual function. The application CPU runs the inference and manages orchestration. Where the hardware platform includes a true neural accelerator, it can also be used. In the tested configuration, routing performance was preserved under event-driven inference load. The exact figures are workload- and configuration-dependent, but the resource envelope is small enough to sit on current premium gateway platforms and the next refresh cycle of managed CPE.

What this changes about data

Local processing significantly changes the data sovereignty calculation. The system can be designed so that raw personal data need not leave the home without explicit consumer consent.

The data classes that can be processed or protected at the household edge are bounded by what the gateway can actually see. Most application content is encrypted within TLS sessions and is not legible to a broadband gateway. What the gateway can see, and what useful household intelligence can be built on, includes network and device telemetry, DNS or destination metadata where available, application-category behaviour, gaming and streaming QoE signals, Wi-Fi performance, device health, security events and, where the customer has authorised specific integrations, smart-home, energy, voice, health or appliance data through those authorised endpoints.

These sit adjacent to the data classes the dominant platforms have built their business models around: household behaviour, device usage, service engagement, attention, identity and intent. The operator has provided the pipe and captured none of the upside. Gateway-resident processing changes that balance. The household keeps the data. The operator holds the consent fabric. The platform partners come in through a permissioned interface on terms set by the operator.

What was free becomes priced. What was extracted becomes licensed.

What leaves the gateway is the derivative output the customer has authorised, scoped to the purpose for which the partner has been granted access, and data-minimised, purpose-limited, and pseudonymised or aggregated where appropriate, so that only what the partner needs and is permitted to see leaves the gateway. Pseudonymised data can still be personal data under UK GDPR where re-identification is reasonably possible, and is governed accordingly.

The regulatory tailwind

The legal and regulatory direction of travel is moving towards this configuration faster than most operators have noticed.

The UK Data (Use and Access) Act 2025, which received Royal Assent on 19 June 2025, establishes a framework to support Smart Data schemes across sectors, including energy, transport and telecoms, subject to scheme design and secondary legislation. The EU AI Act is being phased in: prohibited practices and AI literacy obligations applied from February 2025, general-purpose AI obligations from August 2025, general applicability from August 2026, high-risk systems in specified areas (such as biometrics, critical infrastructure, education, employment, migration and border control) from December 2027, and high-risk AI embedded in regulated products from August 2028 following the AI omnibus political agreement. Adequacy decisions between the UK and EU remain sensitive to the perceived rigour of data protection enforcement at home. US state-level privacy legislation is converging on similar consent-and-purpose architectures.

Operators that build the on-gateway architecture end up positioned to satisfy these frameworks by design, with audit trails that map cleanly onto what regulators are increasingly asking to see. Operators that do not are accumulating compliance risk on a data architecture that is moving out of regulatory favour year by year.

The position no other actor can assemble

The telco position in this configuration is structurally unique.

Operators are already the entity in the value chain with the physical connectivity, the billing relationship, the consumer trust, the regulatory permissions and the installed device footprint. Adding sovereign on-device processing to that stack produces a configuration that no other actor in the digital economy can credibly assemble.

Hyperscalers have compute and, in some cases, consumer devices, but they generally do not hold the regulated broadband access network, the managed CPE estate, the connectivity billing relationship, and the operator consent fabric in one place. Device manufacturers have hardware presence and customer accounts, but not the regulated access-network position or managed broadband relationship. Content and service providers have the demand but not the underlying infrastructure. AI-native platforms have the model capability but no presence in the home and no established consent architecture.

Only the telco sits at the intersection of all of these.

The commercial expression is an ecosystem of knowledge rather than a pipe. Telcos that build the gateway as a local inference and consent platform become the broker of derivative insight to partners across health, energy, insurance, mobility and content, on terms that those partners cannot replicate by going around them. A federated query architecture lets specialist vertical operators ask questions of the installed base without seeing the underlying data, with the telco governing what is asked, what is answered, what compensates the household for the contribution and what is logged for regulatory audit.

A new revenue layer emerges above connectivity: governed insight per query. The underlying access network becomes the moat rather than the only product being sold.

The cost case that justifies the build

The investment is justified before the platform value is even considered.

Customer-recognised issues are where operating costs concentrate, and the prevailing cloud-centric telemetry architecture spends heavily to detect them slowly. Streaming raw data from millions of gateways into a centralised analytics platform consumes backhaul, ingest, storage and compute on a scale that has become disproportionate to the value extracted. By the time the cloud-side analytics flag a degradation, the customer is normally already on hold with care.

Pushing the inference into the gateway collapses telemetry volumes by orders of magnitude. Cloud cost falls. Resolution speed improves from minutes or hours to seconds. The gateway closes the loop locally, renegotiating Wi-Fi channels, restarting a misbehaving radio, throttling a runaway device or prompting the customer through an app before the call to care is ever made.

The incremental energy and compute costs on the gateway itself are negligible, ranging from low single-digit pence to low pounds per household per year, depending on workload, and are incurred on a unit that is already provisioned, powered, and running.

In well-designed pilots, high double-digit reductions in the affected telemetry-processing workload should be a credible target, with care-cost reduction measured separately through avoided contacts, repeat calls and engineer dispatches.

The capability built for care is reusable across the portfolio

The same inference fabric that detects Wi-Fi degradation can also detect a streaming session falling below the QoE threshold of a premium entertainment tier, monitor the health of a connected security camera estate, detect anomalous behaviour in a smart energy device fleet, and verify that a telehealth consultation meets clinical-grade reliability.

The classifier changes. The architecture does not.

A single investment in on-gateway inference, justified solely by care-cost returns, becomes the substrate for entertainment QoE, home security, energy management, and telehealth assurance, with marginal cost dominated by model development rather than infrastructure. By the time a fifth service is layered onto the same fabric, the unit economics approach pure software margins.

The consolidation trap

The customer authorised your relationship. They signed a contract with you. Ofcom regulates the communications relationship. The ICO and UK data protection law govern the personal data processing that sits on top of them.

When you install a vendor agent on the gateway that streams household data to the vendor's cloud, the lawful basis for that processing does not automatically extend. A vendor-controlled data path, especially one used for the vendor's own commercial purposes, needs a fresh controller/processor or controller/controller analysis. If the vendor is acting as a processor, Article 28 UK GDPR terms are required. If it is acting as an independent or joint controller, the operator needs an appropriate data-sharing basis, transparency, purpose limitation, security guarantees and an audit trail. What looked like a single regulated relationship becomes a chain of relationships, with the vendor inserted between you and the customer who originally consented to you.

That is disintermediation, dressed up as consolidation.

A version of the platform consolidation argument is being pitched hard by a handful of well-funded vendors. One agent. One cloud. One AI brain owned by the vendor. The diagnosis is right: a dozen-point tools in the home experience stack is an integration tax that eats the savings. The prescription is the part to watch.

It replaces a multi-vendor integration problem with a single-vendor concentration problem that is far harder to unwind once the contract is signed. The agent sits on the gateway, but the data flows to a cloud the operator does not own. The intelligence runs on infrastructure that the operator cannot audit. The operator becomes a tenant of the platform rather than the owner of its own household telemetry. The exit cost rises every quarter the relationship continues. The negotiating position erodes. The supplier captures a greater share of the value the operator generates with each release cycle.

The alternative is the ecosystem under operator control. Keep the data inside the perimeter that the customer consented to. Run the inference on the gateway estate the operator already owns. Open the consent fabric to specialist partners on terms set by the operator, not on terms a vendor imposes. The household keeps its data. The operator keeps the customer relationship. The partners pay for permissioned access. What was being extracted for free becomes licensed at scale.

The right consolidation move is architectural rather than commercial. Consolidate the data fabric, the consent architecture and the inference layer onto the operator's own gateway estate. Keep vendor relationships at the model and application layer, where they can be swapped, multi-sourced and competed.

Whoever holds the data architecture eventually holds the customer. An operator that has spent two decades watching that play out at the OTT layer should not repeat the mistake at the home experience layer.

The window is open now

The cleanest way to think about the next telco capex cycle is as a competition for where intelligence is allowed to live.

The cloud-centric model will keep paying for backhaul, storage, compute and the latency-induced churn that comes with detecting problems after the customer has already noticed them. The gateway-centric model pays for none of those at any meaningful scale, sees care-operation costs fall as a direct consequence, builds a service assurance fabric that is reusable across several revenue lines, and holds a sovereign data position that the wider digital economy can neither replicate nor route around.

Cost makes the decision urgent on its own. Platform value makes it strategic. The combined sovereignty and ecosystem position is what elevates it to a generational question.

The next telco capex cycle is not just a gateway refresh. It is a decision about where household intelligence lives, who controls the consent fabric, and who captures the value of the data generated inside the home. Operators that treat the gateway as a cheap router will become tenants in someone else's platform. Operators that treat it as a sovereign inference layer can rebuild the economics of care, assurance and household services around an asset they already own.

Jonathan Lishawa is Founder & Chairman of Presciense, writes on infrastructure under hard limits at lishawa.com, and contributes to illuminem.