Skip to content

Bandwidth Formula

The bandwidth provisioning formula calculates the minimum and recommended bandwidth for every AI-carrying link, accounting for U_eff, AIW, burst behaviour, utilisation targets, and a safety margin.

The formula set

B_minimum    = (U_eff × AIW × Peak_factor) / Link_utilisation_target

B_recommended = B_minimum × 1.5         [safety margin for AI burst]

B_per_link   = B_recommended / Number_of_parallel_uplinks

B_total      = B_recommended + Model_sync_BW

Formula variables

Variable Recommended value Rationale
Peak_factor 1.5–2.5 AI responses burst at start; model calls spike; concurrent requests spike during peak hours
Link_utilisation_target 65–70% Never exceed 70% on AI links — burst headroom is critical
Safety margin 1.5× Absorbs unforeseen AI workloads, user growth, and model expansion
Parallel uplinks 2 minimum Resilience: AI service must survive a single-link failure

Design input table

Input variable Reference Your value Result
U_eff From Chapter 2 Step 1 ___
AIW (Mbps/load unit) From Chapter 2 Step 2 ___
Peak_factor 1.5–2.5 ___
Link_utilisation_target 65–70% ___%
B_minimum (Mbps) = U_eff × AIW × Peak / Util = ___
B_recommended (Mbps) = B_min × 1.5 = ___
Number of parallel uplinks 2 minimum ___
B_per_link (Mbps) = B_rec / Links = ___
Model sync BW (Mbps) From model sync calculation ___
Total to provision (Mbps) = B_rec + Model_sync = ___

Worked example

Site: Mumbai contact centre
Inputs: U_eff = 1,390, AIW = 12 Mbps, Peak_factor = 1.8, Utilisation = 65%, Links = 2

B_minimum = (1,390 × 12 × 1.8) / 0.65
          = 30,024 / 0.65
          = 46,191 Mbps
          ≈ 46.2 Gbps

B_recommended = 46,191 × 1.5
              = 69,287 Mbps
              ≈ 69.3 Gbps

B_per_link = 69,287 / 2
           = 34,643 Mbps
           ≈ 34.6 Gbps per uplink

This is a naive all-cloud calculation

This example shows what happens if all 1,390 load units send their full 12 Mbps AIW to the cloud. A 69 Gbps WAN requirement is economically prohibitive. This is why MCP tier routing (Chapter 6) is critical: routing LL = 4 workloads to campus-edge AI reduces the WAN-bound AIW from 12 Mbps to approximately 2 Mbps per load unit, bringing B_recommended to ~11.5 Gbps — a practical 10G or dual-10G provisioning target.

Revised example with MCP tiering

After routing LL = 4 workloads to campus-edge AI:

WAN-bound AIW (cloud traffic only): 2.0 Mbps
B_minimum = (1,390 × 2.0 × 1.8) / 0.65 = 7,700 Mbps
B_recommended = 7,700 × 1.5 = 11,550 Mbps ≈ 11.6 Gbps
B_per_link = 5,775 Mbps — a dual 10G SD-WAN path is sufficient

WAN upgrade options comparison

Option Capacity Latency Cost (India) Best fit
Dedicated MPLS 10G 10 Gbps 5–15 ms High Large enterprise, SLA required
SD-WAN dual 10G internet 20 Gbps aggregate 8–25 ms Medium Most enterprise deployments
SD-WAN 10G internet + 1G MPLS 11 Gbps 5–25 ms Medium Hybrid resilience
Dark fiber / leased line 100 Gbps+ < 1 ms Very high LL = 5 sites, campus backbone

Selecting the right WAN tier

Use the B_recommended value from your calculation to select a WAN tier:

B_recommended ≤ 1 Gbps    → 1G MPLS or 1G internet is sufficient
B_recommended 1–10 Gbps   → 10G SD-WAN or MPLS
B_recommended 10–50 Gbps  → Dual 10G or single 100G SD-WAN
B_recommended > 50 Gbps   → Dark fiber or leased line mandatory

The B formula applies to every AI-carrying link, not just WAN. Campus core and distribution links must also be sized:

AI-dense deployments push significant east-west traffic between access and core switches (inference pods, AI cameras, telemetry). Campus core uplinks should be sized at:

Core_BW = B_recommended × 1.2        [20% additional for east-west AI traffic]

Each access switch serving AI-dense areas (agent floors, inference server rooms) needs uplinks sized at:

Access_uplink = (Agents_on_switch × AIW × 1.8) / 0.70

For a 48-port access switch serving 40 AI agents at AIW = 12 Mbps:

Access_uplink = (40 × 12 × 1.8) / 0.70 = 864 / 0.70 = 1,234 Mbps ≈ 2 Gbps

A single 1G uplink is insufficient. Dual 10G uplinks in port-channel recommended for any access switch serving AI agent floors.


Redundancy and growth in B calculations

Redundancy

Active-active dual-path design requires each path to carry 100% of B_recommended in failover scenarios:

B_per_path = B_recommended        [not B_recommended / 2]

Dual 10G paths in active-active carry 20 Gbps combined normally and 10 Gbps on the surviving path during failover. Plan for the failover scenario.

Growth planning

Apply a growth buffer to B_recommended based on your expected AI expansion timeline:

Growth scenario Buffer
Stable deployment, no planned expansion 1.0× (no buffer)
20% user growth expected in 12 months 1.2×
Active AI expansion, doubling in 18 months 2.0×
Greenfield — full capacity unknown 2.5×
B_with_growth = B_recommended × Growth_buffer