Production Cloud Cost · 1 of 2

Platform Cost — the shared floor Production-grade (HA / SLA) fixed infrastructure · present whether you run 1 store or 100 · scales step-wise, not linearly

Production HA / SLA Event spine all data sources Pilot floor ~$3.0K/mo

What the ~$3.0K/mo production floor buys (pilot scale)

Component	$ / mo	What it covers
Compute — AKS / Container Apps	~$600	Hosts the ~12 .NET services; orchestration in-process over the event spine
Azure Data Explorer (ADX)	~$1,200	Production time-series cluster — every event + the LLM / decision audit trail
Event Hubs (Standard)	~$100	The event spine — POS, traffic, inventory, sensors, video metadata
Cosmos DB (provisioned)	~$250	Operational records, agent memory, tenant / store context
Cache for Redis (Standard C2)	~$115	Hot state, cooldowns, the statistical anomaly baselines
Azure AI Search (S1)	~$250	Vector index for RAG — store & enterprise knowledge base
Blob storage	~$80	Canonical audit — full LLM prompts / responses retained (SOC 2)
Platform overhead	~$400	Monitoring, Key Vault, identity, networking, egress, contingency
Production floor	~$3.0K	≈ $36K / yr at pilot

What the floor carries — at near-zero marginal cost

The event spine already ingests POS, inventory, people-counters, sensors, weather, supply-chain & staffing. These structured streams are a rounding error next to video and agent reasoning — they ride essentially free on the same infrastructure.

Scales step-wise, not linearly

$3.0K→$4.0K→$18K1 · 10 · 100 stores

Floor grows by tier bumps (AI Search, ADX, Event Hubs Premium) as you add stores — so the per-store share of the floor keeps falling.

Not in this estimate — budget separately

Non-production envs (dev / test / staging)
Integration engineering (POS / inventory / counters)
Third-party data & API fees
Edge hardware (cameras / NVR)
One-time development & the team
Support & SOC 2 / compliance

Production Cloud Cost · 2 of 2 · Run cost & scaling

1-Year Run Cost — scaling with stores Azure-hosted · per-store LLM agent · frame-sampled video (1 fps) to a hosted vision-LLM — pilot basis

Video 8 cams · 1 fps Vision hosted LLM / frame Reasoning tiered Haiku→Opus Detection gates + minutely scan

Total cost of ownership — scaling with store count

	1 storepilot	5 storesexpansion	10 storespilot target	→ 100fleet · curve
Platform floor /mo (slide 1)	$3.0K	$3.3K	$4.0K	$18K
Per-store variable (~$685/store·mo)	$0.7K	$3.4K	$6.8K	$68K
Total / month	~$3.7K	~$6.7K	~$10.8K	~$86K
Pilot basis / year hosted LLM / frame	≈ $44K	≈ $81K	≈ $130K	≈ $1.0M
↓ At scale, cheap CV / year	≈ $40K	≈ $58K	≈ $85K	≈ $580K

Pilot basis: each sampled frame (1 fps) goes to a hosted vision-LLM for triage, with the LLM cascade only on the ~3% of flagged events — simplest to ship, nothing to operate. ~$130K/yr at 10 stores. At ~10+ stores, continuous monitoring moves to self-hosted GPU / edge CV → video ~4× lower (~$85K) — a scale lever, not baked into the pilot.

Per-store cost stack — pilot (~$685/mo)

Vision 1 — hosted LLM triage, every frame~$409

Vision 2 — Gemini Flash clips (escalation)~$56

Vision 3 — frontier (rare)~$45

Agent reasoning + minutely scan~$88

Infra + audit substrate (per store)~$80

RAG embeddings~$5

Video cost — pilot vs. scale ($/store·mo)

Per-frame frontier VLM (over-engineered)~$1,500

Pilot — hosted LLM per frame$510

At scale — cheap CV (GPU / edge)$132

…with edge CV continuous (+ HW capex)$101

A hosted LLM on each frame is the simplest pilot. At ~10+ stores, continuous detection moves to computer vision (one GPU fills up across the fleet) → video ~4× lower.