Formula Overview¶
All formulas used in this workbook, presented in calculation order with variable definitions and enterprise adaptation notes.
Complete formula set¶
Step 1 U_eff = (H × 1.0) + (AI × 3.5) + (IoT × 0.2) + (Pod × 7.0)
Step 2 AIW = Σ(stream_Mbps) × Burst_factor
Step 3 CS = 1–5 → BW overhead %:
CS1=+8% CS2=+18% CS3=+28% CS4=+40% CS5=+60%
Step 4 LL = 1–5 → RTT budget (ms):
LL1=500 LL2=200 LL3=80 LL4=31 LL5=20
Step 5 B_min = (U_eff × AIW × Peak_factor) / Link_utilisation_target
B_rec = B_min × 1.5
Model_sync_BW (Mbps) = GB × Updates/day × 8000 / 86400
B_total = B_rec + Model_sync_BW
Step 6 IS = (U_eff × AIW × CS × LL) / (B × A)
Step 7 PUO = IS / U_eff
Step 8 Route = f(LL, CS, Data_residency, Cost)
Step 9 B_req = (U_eff × AIW × CS × LL) / (3.0 × A) [target IS = 3]
B_gap = B_req − B_current
Variable definitions¶
Primary formula variables¶
| Variable | Full name | Units | Range | Source |
|---|---|---|---|---|
| H | Human operators | Count | Any | Headcount |
| AI | AI software agents / bots | Count | Any | Platform inventory |
| IoT | IoT devices and edge sensors | Count | Any | Network scan |
| Pod | GPU inference pods | Count | Any | Infrastructure inventory |
| U_eff | Effective load units | Dimensionless | Any | Calculated (Step 1) |
| AIW | Avg AI workload per load unit | Mbps | 0.5–40 | Calculated (Step 2) |
| CS | Cybersecurity risk factor | Dimensionless | 1–5 | Scored (Step 3) |
| LL | Low-latency factor | Dimensionless | 1–5 | Assessed (Step 4) |
| B | Provisioned bandwidth | Mbps | Any | Procured / planned |
| A | Adjustment factor | Dimensionless | 0.5–1.0 | Infrastructure type |
| IS | Impact Score | Dimensionless | Any | Calculated (Step 6) |
| PUO | Per-User Output | Dimensionless | Any | Calculated (Step 7) |
Derived and supporting variables¶
| Variable | Formula | Typical value |
|---|---|---|
| Peak_factor | AI burst multiplier | 1.5–2.5 |
| Link_utilisation_target | Max safe link load | 65–70% |
| Model_sync_BW | GB × updates × 8000 / 86400 | Varies |
| B_gap | B_required − B_current | Must be ≤ 0 for readiness |
Entity multipliers¶
| Entity type | Multiplier | Rationale |
|---|---|---|
| Human operator | × 1.0 | Baseline. Think/pause cycles limit sustained load. |
| AI software agent / bot | × 3.5 | 24/7, no idle time, 3–4 parallel API calls continuously. |
| IoT / edge sensor | × 0.2 | Low per-device but high volume aggregates dangerously. |
| GPU inference pod | × 7.0 | Serves all users simultaneously, model weights, inter-pod comms. |
IS verdict thresholds¶
| IS range | Status | Action |
|---|---|---|
| IS ≤ 1.0 | Optimal | Monitor quarterly. AI can grow into headroom. |
| 1.0 < IS ≤ 3.0 | Monitor | Enforce 6-class QoS. Alert at IS > 3. |
| 3.0 < IS ≤ 10.0 | Upgrade required | Bandwidth upgrade or edge AI within 90 days. |
| IS > 10.0 | Deployment blocker | No AI go-live. Redesign first. |
PUO verdict thresholds¶
| PUO range | Status | Action |
|---|---|---|
| PUO < 0.3 | Over-provisioned | Review if B can reduce at next renewal. |
| 0.3–1.0 | Balanced | Maintain. Monitor trend quarterly. |
| 1.0–2.0 | Stressed | Add bandwidth or reduce AIW via model quantisation. |
| PUO > 2.0 | Overloaded | Immediate action: edge AI or bandwidth upgrade. |
A factor reference¶
| Infrastructure type | A value |
|---|---|
| Dark fiber / dedicated leased line | 0.95–1.00 |
| Dedicated MPLS with SLA | 0.90–0.95 |
| SD-WAN dual-path managed | 0.82–0.90 |
| Shared MPLS / business broadband | 0.70–0.82 |
| SD-WAN over shared internet | 0.65–0.75 |
| Legacy single-ISP internet only | 0.50–0.65 |
Enterprise adaptations¶
These variables are adapted from the base IS model to fit enterprise realities.
Why U_eff replaces raw headcount¶
Raw headcount treats every entity as a human with equal network demand. In an AI enterprise, this is dangerously wrong. A GPU inference pod consumes 7 times the sustained bandwidth of a human operator. An AI bot runs continuously with no idle time. The multiplier system converts all entity types to equivalent load units before any calculation.
Why AIW is a composite stream value¶
Enterprise AI is always multi-modal. A Webex CC agent using AI assist simultaneously generates: a speech-to-text stream (0.3 Mbps), an LLM inference stream (1.5 Mbps), a screen analytics stream (3 Mbps), and a RAG query stream (1.8 Mbps). These run concurrently, not alternately. AIW is the sum of all active streams multiplied by the burst factor, not the largest stream alone.
Why CS translates to bandwidth overhead¶
CS is not just a risk rating. Every CS point above 1 represents real network overhead from encryption inspection, DLP scanning, micro-segmentation headers, and audit logging. CS = 4 (DPDP-regulated PII workloads) adds 40% overhead to every byte traversing the network. If your provisioned bandwidth is 10 Gbps and CS = 4, your effective AI capacity is 7.14 Gbps, not 10 Gbps.
Why LL determines topology, not just SLA¶
LL is not a performance target — it is an architecture forcing function. LL = 4 (31ms budget) is physically impossible over a cloud WAN path from India to any non-Indian cloud region. A Mumbai-to-Singapore round trip alone is 45ms. LL = 4 mandates on-campus AI inference. LL = 5 (20ms) mandates on-premises GPU with zero WAN traversal. No QoS policy or bandwidth upgrade can overcome physics.