Tenant archetype dictates isolation posture, SLA envelope, and commercial model — an ISV running a shared-tenancy vision model has nothing in common operationally with a FedRAMP-regulated enterprise tenant requiring dedicated silicon. Mixing these on a single isolation design is the fastest way to be unable to sell to either. STL Partners' 2024 operator survey found that 61% of tier-1 operators with active edge-AI-as-a-service offerings segment tenants by isolation class from day one.

AWS Wavelength documentation cites sub-10ms round-trip inference inside Verizon's 5G edge versus 60–80ms to the nearest central region — the gap that makes a carrier-hosted offer structurally differentiated. Setting an overly permissive SLA makes the offer indistinguishable from hyperscaler central cloud; setting an overly tight SLA across every tenant forces the most expensive hardware posture on every site. The SLA must be tiered by commercial SKU before deployment topology is fixed.

Billing-grade metering is a first-order architectural concern, not a late-stage instrumentation task — if per-request counts or per-GPU-hour attribution drift by more than a few percent, tenant disputes consume more operations time than the revenue is worth. Carrier finance, revenue assurance, and enterprise procurement all have existing standards for metering error (typically sub-5%) inherited from connectivity billing. An AI-as-a-Service offer that cannot pass revenue assurance review cannot be commercially launched regardless of how good the inference is.

SKU structure decides the platform's addressable market more than the technology does — a shared on-demand SKU can serve thousands of small developers, while a dedicated-capacity SKU serves a handful of regulated enterprises at 10x margin. Launching with only one SKU forecloses half of each market. Published pricing of $0.50–$2.00/GPU-hour shared and $5–$20/GPU-hour dedicated is consistent with hyperscaler edge benchmarks and is the reference for commercial plausibility.

Residency is typically a per-tenant contractual boundary, not a platform-wide one — the same platform must enforce EU residency for one tenant while serving a US tenant from US sites and a FedRAMP tenant from government-boundary-only sites. Architecting residency as a policy rather than a physical pinning invites drift — failover, backup, and logging paths all need to respect the boundary. State-level US requirements under CCPA/CPRA and sovereign-cloud contracts materially complicate US-only architectures.

3GPP Release 18 (5G-Advanced) and the in-progress Release 19 explicitly scope AI/ML functions in the 5G system, and O-RAN Alliance has published rApp / xApp frameworks that define how third-party AI plugs into the SMO and RICs. An edge AI-as-a-Service offer that does not map to these references becomes an integration project for every tenant and every carrier partner. TMF Open APIs specifically drive BSS / ordering / SLA interoperability that determines how fast a tenant can actually buy and consume the service.

Tower and MEC sites have finite power and cooling envelopes that were designed for radio equipment, not dense GPU inference — retrofitting a 10kW site to host a multi-tenant GPU pool is a civil and utility project, not a software deployment. Site survey data (power, cooling, HVAC, floor loading) must drive the hardware catalog rather than the reverse. A platform design that assumes hyperscaler-class power per site will never deploy on the actual footprint.