Household Intelligence Platform
A private context orchestrator for cable's edge infrastructure. Memory-augmented inference inside a hardware-attested enclave. Operator-blind by architecture, not by promise.
Raw model intelligence is becoming a commodity. The gap between the best model and the one a tier below it is collapsing. What lasts is the harness around the model: the context it draws on, the memory it keeps, the systems it can reach.
For the consumer, the heart of that harness is context. Whoever owns it for the home owns the consumer AI relationship for a generation.
The Problem
The frontier labs are losing billions on consumers to win the enterprise. OpenAI lost $9 billion in 2025, just filed a confidential S-1, and its CFO Sarah Friar says the strategy plainly: free users average 7 queries a day, paid users hit 15, power users run 7x the average. But 95% of their 900 million weekly users are free. The path to profitability runs through enterprise seats at $30–50/month and advertising against the free tier — not through building differentiated household products.
The consumer is the onramp, not the customer. Every model improvement, every new capability, every infrastructure dollar is optimized for the token economics that work: enterprise contracts and API revenue. The household gets what's left over.
On June 4, 2026, OpenAI shipped Dreaming — a background process that reads across years of your conversations, synthesizes what it knows about you, and rewrites that knowledge as your life changes. It is longitudinal memory with automatic supersession, built on your most personal data, running on OpenAI's servers, in a data layer you don't control.
Dreaming validates the concept: memory is the moat, not the model. But it also demonstrates the problem. Every frontier lab's business model requires data access, not data custody. Your household context — health decisions, financial plans, family dynamics — becomes platform intelligence. There is no hardware enclave, no attestation, no guarantee that your facts aren't training signal. The household needs memory that belongs to the family, not the platform. No frontier lab will build that. Their incentive structure won't allow it.
Cable's history is a sequence of infrastructure kills. Coax buried network TV antennas. Broadband killed AOL and dial-up. VoIP displaced local voice. Bundled long distance collapsed interexchange carriers. Speed tiers made DSL irrelevant. Every time, the pattern was the same: cable leveraged superior local infrastructure to deliver a product the incumbent couldn't match, then captured the revenue stream permanently.
That pattern is broken. Comcast and Charter together shed more than a million broadband subscribers in 2025. Fixed wireless passed nine million. Starlink crossed ten million. Connectivity is commoditizing because the pipe alone is no longer a defensible advantage.
Cable needs a compounding value play that rests on its local infrastructure — not another speed tier, but a service that gets more valuable every month the subscriber stays. A service that creates switching cost through accumulated context, not through contract terms. A service that only works because cable has GPU compute within 10ms of 65 million homes, a billing relationship already in place, and a device already on the wall. The memory graph is that play. It appreciates. It compounds. And it can't be replicated by a satellite 340 miles overhead or a fixed wireless antenna on a water tower.
Why Cable
Filling the vacuum takes cheap inference close to the home, a private place for context to live, an interface that reaches the whole household, and a trusted billing relationship. Cable has all four. No other player has more than two.
GPU inference within 10ms of 65 million homes. Comcast and Charter are deploying NVIDIA AI Grid hardware today.
NVIDIA Confidential Computing on Blackwell GPUs. Hardware-attested enclaves the operator cannot read into. Operator-blind by silicon, not by policy.
A device already in the home. A billing relationship already established. Trust already earned through decades of service delivery.
Shared AI Grid infrastructure with marginal cost under $1.50/household/month. Operators allocate excess GPU capacity to external compute markets while ramping the consumer base — revenue from day one, not after scale.
Architecture
The router escalates by reason, not generically. Each tier has a distinct privacy posture. Sensitive queries never leave the trust boundary. The routing architecture is deployment-agnostic — it works whether inference runs at the cable edge hub, on an in-home device, or in a hybrid model. That optionality is deliberate. The privacy guarantee holds regardless of where the compute lives.
Household memory queries. Facts never leave the cable trust boundary.
OPERATOR-BLINDCurrent information via web fetch. Only search strings leave.
SEARCH STRING ONLYCC enclave inference. Larger model inside hardware TEE.
ENCRYPTED IN ENCLAVEUser-initiated frontier model. Household context stripped.
USER OPT-OUTEvery routing decision reads from a configuration plane. Toggle escalation on or off. Enable or disable axes independently. Change tier targets. All live, no code change, no restart.
This is more than application configuration — it's the control surface of a household AI operating environment. The same plane that routes queries today becomes the surface that third-party applications, operator management GUIs, and subscriber preferences all write to. Applications inherit the routing logic, the privacy posture, and the sensitivity gating without rebuilding any of it.
The demo toggle is the proof: flip escalation off, ask a current-events question, watch it fail. Flip it on, watch it route to the right tier and succeed.
Memory
HIP maintains a persistent, longitudinal memory graph per household. Facts are extracted, tagged, and organized across four properties. Supersession is built in: the system knows what is currently true versus what used to be. This graph is not just HIP's memory — it's a permissioned context substrate. Applications built on the platform query against it with scoped, permission-gated access. The sensitivity classifier enforces boundaries at the platform level, not inside each application.
Scoped retrieval: only pull facts from relevant domains per query
Memory ownership belongs to the speaker, not the household
High-sensitivity facts never leave the trust boundary
Ephemeral facts expire; permanent facts persist until superseded
The Edge
This is not a concept architecture. Comcast and Charter announced active NVIDIA AI Grid deployments at GTC 2026. The hardware HIP targets is being installed in edge data centers today. The AI Grid is the compute substrate of the platform — low-cost inference within 10ms of 65 million homes, shared across every application running on HIP. Applications don't provision their own models or negotiate their own GPU capacity. They call the platform, and the edge infrastructure handles the rest.
~200 edge data center locations. Field trials with Personal AI's memory-based SLMs on RTX PRO 6000 GPUs. Sub-500ms latency at P99. 65 million homes within reach.
1,000+ Edge Compute Infrastructure sites. Distributed GPU deployment based on NVIDIA AI Grid reference design.
The Foundation
Each application listed below requires most or all of these properties to exist. Any one of them can be built independently. Only this platform provides all six as infrastructure.
The system knows you over months and years, not just this session. Context accumulates. Supersession tracks what's currently true versus what used to be. No other consumer platform builds a temporal knowledge graph per household.
Speaker verification means the system knows who is talking and scopes everything accordingly. Age-appropriate for the teenager, permission-gated for the guest, personality-aware for each adult. Identity is a platform service, not an application feature.
Sensitivity classification at the platform level means health, financial, and personal data is gated automatically. Applications don't implement privacy — they inherit it. A health app never sees financial data. Not because of policy, but because the platform structurally prevents it.
Inference cost low enough to make always-on, ambient AI economically viable for a consumer household. Cloud-dependent architectures can't match this cost structure. The margin exists at $10/month because the infrastructure is shared and local.
Accessible to every household member regardless of technical ability, age, or physical capability. No screen required. No app to open. The interface disappears, and the platform is just there.
The household's context is encrypted with keys only the consumer holds — derived from biometric identity, custodied in hardware. The operator stores ciphertext it cannot read. The consumer can export it, delete it, or migrate it. This is data custody, not data access. If you switch operators, your context goes with you. Your memory graph belongs to your household, not to a platform.
The platform does not lock you to one model. Four tiers cascade by reason: a private edge model for household queries, a CC enclave for heavy reasoning, and your own frontier model subscription for everything else — with your household context layered in or stripped out, your choice. Bring your Claude account, your GPT subscription, your Gemini key. The platform orchestrates across all of them, routing each query to the right tier based on sensitivity, capability, and your explicit preferences.
Platform Applications
The memory graph, the privacy architecture, and the edge compute substrate are not specific to a single application. They are infrastructure. Every application built on the platform inherits the four-tier routing, the sensitivity classifier, and the privacy guarantee without rebuilding any of it.
Wellness check-ins, medication tracking, and family alerts for elderly household members. The privacy architecture matters most here: memory-impaired individuals need a system that remembers for them, protected by hardware, not by promise.
A coaching layer that meets each household member where they are. Age-appropriate, personality-aware, cognitively matched. Knowing when to guide, when to challenge, and when to step back. Every dimension of adaptation compounds over time.
Prosumer robotics in the household need identity, context, and safety gating. The platform provides all three — the same trust boundary that protects a voice query protects a robot's actions.
The house itself has a context graph. When was the roof done, what paint color is the living room, which contractor fixed the furnace, when does the warranty expire. No one remembers this. The house should.
Multi-member scheduling, task delegation, shared context across family members. Queries that only work with shared memory: "Did Sarah say what time she'd be home?"
Symptom logging, medication awareness, doctor visit preparation, longitudinal health context. Sensitivity-gated at the platform level — health facts never leave the trust boundary.
The system knows who's home, medical conditions, medications, allergies, emergency contacts. In a crisis it can brief first responders with information the family forgets under stress.
Spending patterns, bill tracking, budget conversations grounded in household history. Privacy architecture ensures financial facts stay inside the trust boundary.
When did I last call my mother. What did we decide about Thanksgiving. What does my nephew like. The extended social graph of the household. Not a CRM — a relationship memory.
Tutoring, homework support, and learning tracking with age-appropriate sensitivity gating. The platform adapts to the learner, not the other way around.
Dietary restrictions per member, allergies, preferences that evolve, what's been eaten recently. Plans for the household, not the individual.
Context-aware automation driven by who is speaking, what they've said before, and what the household needs right now. Identity-aware, not just device-aware.
Voice-first by design means natively accessible to visually impaired, motor-impaired, and cognitively impaired household members. The identity layer adapts the interface to whoever is speaking.
Knows which member speaks which language, mediates between family members, helps children maintain heritage language skills.
When a household member dies, the memory graph holds their context. The system can surface estate tasks, important contacts, unfinished business. Deeply sensitive, uniquely enabled by longitudinal memory with hardware-enforced privacy.
Separates business context from personal context within the same household, with different sensitivity rules and routing profiles per domain.
Contractor history, service provider preferences, neighborhood context. The cable operator already serves the neighborhood — the platform extends that relationship.
Economics
Edge-hosted, shared infrastructure. Marginal cost structure that improves with scale. Full model available under NDA.
Full storage BOM, inference BOM, model assumptions, and capacity planning available under NDA. Contact directly for access.
The Builder
Bill Brewster spent 12 years as SVP and GM of Canoe Ventures, the $1B cable industry joint venture spanning Comcast, Charter, Cox, and 100+ operators. He built the telemetry, analytics, and ML infrastructure. Achieved MRC accreditation. Managed the multi-stakeholder operating model that kept competing operators aligned.
Before Canoe, he designed Comcast's first centralized National Video Operations organization and worked on the X1 platform. Earlier: AT&T/TCI, where he drove a $55M annual revenue initiative through a 380,000-subscriber conversion program.
HIP is the convergence of 25 years of cable operations experience with a systems-builder's approach to AI infrastructure.
Olinda Solutions
Canoe Ventures, SVP & GM
Comcast, Transformational Lead
AT&T / TCI
UW MBA, Finance
CU Boulder BA
Full trust model. CC enclave design. Storage architecture. Four-tier routing internals. Edge infrastructure reference targets.
No spam. This is a technical document, not a newsletter.
Project Status
Last updated: June 8, 2026
If you're an operator, an investor, or building in this space.
bbrewster@olindasolutions.comLakewood, CO