Redundancy & Growth Planning¶

AI workloads create new resilience requirements that traditional enterprise networks are not designed for. A legacy network survives a link failure with a 30-second failover and some degraded performance. An AI contact centre with 500 agents loses ₹2–5 lakhs per minute of AI downtime during peak hours. Redundancy requirements are fundamentally different.

Redundancy design principles for AI networks¶

Principle 1 — Size each path for 100%, not 50%¶

In an active-active dual-path design, each path must carry the full AI load independently:

B_per_path = B_recommended    [NOT B_recommended / 2]

If B_recommended = 10 Gbps and you have two 10G paths, you normally carry 5 Gbps on each. If one path fails, the surviving path must handle 10 Gbps. Both paths must be sized for this scenario.

50% sizing causes AI outage on failover

A design that sizes each path at 50% of B_recommended will saturate the surviving path the moment failover occurs. All AI calls will fail. This is the most common enterprise AI network resilience mistake.

Principle 2 — Failover must complete in under 10 seconds¶

Traditional routing failover (OSPF reconvergence) takes 30–60 seconds. During that window, all AI inference calls timeout. For AI contact centres, 30-second failover is unacceptable.

Required failover technologies:

Technology	Failover time	Suitable for
BFD + OSPF	1–3 seconds	Campus core failover
SD-WAN active-active	< 1 second	WAN path failover
BGP with BFD	2–5 seconds	Multi-homed internet
FHRP (HSRP/VRRP)	3–10 seconds	Default gateway failover
Campus edge AI failover	5–15 seconds	Tier 3 pod redundancy

Principle 3 — AI inference pods need N+1 redundancy¶

A single GPU inference pod serving 500 agents is a single point of failure. If the pod fails, all 500 agents lose AI assist simultaneously. Design for N+1:

Minimum pods = CEIL(U_eff / Pod_capacity) + 1

Where Pod_capacity is the number of effective load units each pod can serve.

Principle 4 — Model weights must be cached locally¶

If the model registry is cloud-hosted and the internet fails, campus-edge pods cannot load new models or restart from a cold state. Cache model weights on local NVMe storage at each site. Model cache = model size × number of production models × 2 (primary + previous version).

Redundancy design worksheet¶

Complete for each site:

Redundancy element	Current design	Target design	Gap
WAN paths	___	Active-active dual paths	___
Per-path bandwidth	___	B_recommended (full load)	___
WAN failover time	___	< 1 second (SD-WAN)	___
Campus core redundancy	___	Dual core switches, ECMP	___
Core failover time	___	< 3 seconds (BFD+OSPF)	___
Inference pod count	___	N+1 minimum	___
Model weight cache	___	Local NVMe, all models	___
NGFW redundancy	___	Active-active HA pair	___
DNS redundancy	___	Dual resolvers, anycast	___

Growth planning¶

The 18-month rule¶

Enterprise AI deployments typically expand from pilot (10% of agents) to full deployment (100%) within 12–18 months. Infrastructure procured for the pilot is almost never adequate for full deployment. Plan for full deployment from day one.

Growth-adjusted bandwidth:

B_growth = B_recommended × Growth_buffer

| Timeline to full deployment | Buffer |
|-----------------------------|--------|
| No planned expansion        | 1.0×   |
| 20% growth in 12 months     | 1.2×   |
| 2× in 18 months             | 2.0×   |
| Full greenfield, unknown    | 2.5×   |

AI model size growth¶

Models are growing in size rapidly. A deployment sized for a 7B model today may need to serve a 13B or 30B model in 18 months. Plan storage and network accordingly:

Model_storage_per_site = Model_size_max × Num_models × 2 (versions)
                       + 50% growth buffer

For a site expecting to run a 13B model (26 GB FP16) with 3 models, keeping 2 versions each, with 50% growth buffer:

Storage = 26 × 3 × 2 × 1.5 = 234 GB local NVMe per site

U_eff growth projection¶

Calculate U_eff at each deployment milestone and verify IS remains below 3 at each stage:

Milestone	H	AI	IoT	Pod	U_eff	IS at current B	Action
Today	___	___	___	___	___	___	___
6 months	___	___	___	___	___	___	___
12 months	___	___	___	___	___	___	___
18 months	___	___	___	___	___	___	___
Full deployment	___	___	___	___	___	___	___

Scalability checklist¶

Before approving an AI network design, verify these scalability properties: