Voice-AI startup Gnani pays three Indian GPU cloud providers simultaneously, splitting its workloads. Immediate access to high-end chips like H100s, not price, dictates who gets the business. Startups like Gnani prioritize reliability for production and instant capacity for model training over any cost advantage.
How We Got Here
India's AI compute market historically saw Yotta, Neysa, and E2E Networks acquire chips from Nvidia to rent out. This changed as Nvidia expanded its role beyond a supplier, now providing architecture blueprints and orchestration software for these providers.
The Numbers
- Nvidia now provides the reference blueprint for cloud companies to rack, cool, and wire their GPUs.
- The American giant also develops the orchestration and diagnostic software monitoring these GPUs for problems.
- Yotta, Neysa, and E2E Networks initially differentiated on scale, custom software, and price respectively.
- Gnani runs production voice agents on Yotta, retraining jobs needing 8 GPUs on Neysa, and experiments on E2E Networks.
- Gnani's ML engineer requires another 256 H100s by Monday for a telecom customer, prioritizing availability over a 20% price saving.
What Happens Next
🇮🇳 Why This Matters for India
For AI founders and ML engineers in Bengaluru and Hyderabad, this shift means less choice and potentially higher, less predictable costs for crucial compute power.
The Take
Nvidia completely controls the Indian AI compute stack, from chips to orchestration software to the actual marketplace. This leaves Indian AI startups, and the providers themselves, with little leverage, effectively locking them into an expensive, single-vendor ecosystem.
Source:
The Ken ↗