It costs $0.10 to generate one second of AI video. That's the global standard charged by Veo, Kling, Luma, and Runway. A 30-second product ad runs $3, and a 5-minute educational clip? Over $30.
Peak XV (formerly Sequoia India)'s managing director put it bluntly: "Current AI video models are too expensive for population-scale use in India."
So India built its own. Varya. ₹0.48 per second — roughly $0.005. That's 1/20th the price of global tools.
Why global AI video tools couldn't scale to India
The pricing structure of global AI video tools makes the problem clear.
| Model | Price ($/sec) | 30-sec video cost |
|---|---|---|
| Sora 2 (OpenAI) | $0.75 | $22.50 |
| Veo 3.1 (Google) | $0.15+ | $4.50+ |
| Kling 3.0 (Kuaishou) | ~$0.10 | ~$3.00 |
| Runway Gen-4.5 | $0.10+ | $3.00+ |
| Varya (Avataar AI) | $0.005 | $0.15 |
Apply these numbers to real scenarios and the gap becomes stark. A rural school in India making AI video lessons, or a small business creating a product ad — global tools were economically out of reach. The problem wasn't price — it was access.
The second barrier is culture. Global models are trained on Western-centric data. They struggle to accurately depict India's diverse festival attire, regional architecture, food, and public spaces. Type "Holi festival scene" and you'd get something stereotypical or off.
Distillation compressed 50 steps into 4 — and that changed the price
Avataar AI didn't build Varya from scratch. They started with Alibaba's open-source Wan 2.2 (14 billion parameters) and applied distillation — compressing a large model's capabilities into a leaner, faster version.
The result: Varya generates video in 4 steps instead of Wan 2.2's 50. On an NVIDIA H200 GPU, that's a 5-second 720p clip in 45 seconds. The same task takes Wan 2.2 1,230 seconds. That speed difference becomes the price difference.
The Indian government's role matters here too. India AI Mission provides access to 38,000+ GPUs (NVIDIA H100·H200 included) at a subsidized rate of ₹65/hour (~$0.78) for startups. Commercial cloud pricing on AWS or Azure runs ₹300–600/hour — so the compute cost barrier has effectively been eliminated. Avataar was one of 12 selected startups that used this subsidized infrastructure to build Varya.
CEO one-liner
Avataar CEO Sravanth Aluru defined Varya this way: "Affordable AI is inclusive AI." MeitY Secretary S. Krishnan called it "frugal innovation that is both world class and accessible at population scale."
Why cultural awareness is a moat
Varya's edge isn't just price. The training data was curated for Indian culture — enabling accurate generation of local festivals, clothing, architecture, food, and public spaces.
| Global AI Video Models | Varya | |
|---|---|---|
| Indian festivals | Stereotypical output | Accurate regional depiction |
| Architecture | Western-centric training data | Indian regional styles supported |
| Food and daily life | Inaccurate or awkward | Accurate local food and spaces |
| Price | $0.10+/sec | $0.005/sec (20x cheaper) |
| Access model | Closed API | Open-weight (self-hostable) |
The open-weight release on AIKosh matters too. Developers and businesses can download the model weights and training data, then self-host or fine-tune for their own domain. Partnerships with Higgsfield and Adobe Firefly have also been announced.
AI tool competitiveness is shifting from "best quality" to "who designs for a specific market first." This isn't just India's story. The same playbook could produce Korean-specific, Southeast Asia-specific, or Arabic-specific models — Varya just showed the formula works.
How to start using Varya
- Try it in your browser
Avataar AI's site (avataar.ai) lets you generate video from text prompts or reference images right now. General-purpose video generation works without Indian cultural context. - Download the open-weight model
AIKosh (indiaai.gov.in) provides Varya's model weights and training data for download. Self-host on your own GPU infrastructure or fine-tune for a specific domain. - Integrate via API
Avataar's hosted service lets you connect via API at ₹0.48 ($0.005) per second. At 20x less than global alternatives, it's especially attractive for high-volume video generation pipelines. - Access via partner tools
Higgsfield and Adobe Firefly integrations are confirmed. If you're already using those platforms, Varya will be available within them.
Known limitations
Varya currently maxes out at 720p. For 4K or 2K production work, Veo 3.1 or Seedance 2.0 remain better options. Direct quality benchmarks against the top global models are still limited in public releases.
