The Question That Determines Everything
If you are deploying voice AI at the scale of India — potentially hundreds of millions of calls per year — the unit economics of the voice pipeline are not an engineering concern. They are a viability concern.
A system that costs ₹5 per call cannot serve a farmer who earns ₹300 per day. A system that costs ₹0.28 per call can.
This is why we obsess over the cost model.
The Two Deployment Options and Their Full Costs
Option A: Bhashini Cloud (Variable Cost Model)
Bhashini operates at ₹330/hour for 50 simultaneous calls. At typical 3-minute average call duration, this means:
- Calls per hour at 50 concurrent: ~1,000 calls/hour
- Cost per call: ₹0.33
- Full variable cost at 1M calls/month: ₹3.3 lakhs/month
No CapEx. No hardware. Pay as you go.
Option B: On-Premises (Fixed Cost Model)
A JOHNAIC-80 server handles 50 simultaneous calls.
- Hardware: ~₹40-60 lakhs (one-time)
- Electricity: ~₹15,000/month
- Maintenance & ops: ~₹30,000/month (shared staff)
- Total recurring: ~₹45,000/month
At 1M calls/month (roughly 2,000 calls/hour continuous operation):
- You need: ~2 servers
- CapEx: ~₹1 crore
- Monthly recurring: ~₹90,000
- Effective cost per call (amortised over 3 years): ~₹0.04/call
The break-even against Bhashini Cloud is approximately 18 months at 1M calls/month.
At 10 Million Calls Per Month
This is where the economics become starkly different:
| Cloud | On-Prem | |
|---|---|---|
| Monthly cost | ₹33 lakhs | ₹3 lakhs |
| CapEx | None | ₹5 crores (10 servers) |
| Break-even | N/A | ~15 months |
| Cost per call (3yr) | ₹0.33 | ₹0.004 |
At scale, on-premises deployment is 80x cheaper per call.
What This Means for a State Government
A state government running a voice AI helpline for 10 million citizens, averaging 2 calls per citizen per year, needs to handle 20 million calls per year.
- On cloud: ₹66 lakhs per year
- On-prem (after break-even): ₹36 lakhs per year in recurring costs, with full data sovereignty
The on-prem CapEx (~₹10 crores for this scale) is a single line item in any state IT budget. The recurring savings fund the initial investment within 3 years.
The Hidden Cost Nobody Calculates
There is a cost that does not appear in any per-call analysis: the cost of vendor dependency.
If a state government builds its citizen services on a commercial voice AI API that changes pricing, discontinues Indic language support, or ceases operations — the rebuild cost is not zero. Procurement cycles, rebidding, integration work, and service downtime are real costs.
The fully-owned VoicERA stack eliminates this hidden cost entirely.