Why Avataar Is Bullish About Cracking AI Video And Outdoing Global Giants
Avataar launched Varya, an India-built AI video model it says can make clips far cheaper. The startup is betting low inference cost can beat bigger, pricier rivals.
Intelligence analysis by GPT-5.4 Mini
Bengaluru startup Avataar is pitching Varya as a cheaper way to generate AI video, claiming major cost and speed gains over its teacher model and global competitors. The company says the model was built for practical use cases in India, where high video-generation costs still limit adoption.
Avataar made a video-making machine that it says works like a shortcut instead of a long slow road. That could make it much cheaper, like using a scooter instead of a truck to deliver the same package.
Analysis
What Avataar launched
Avataar has introduced Varya, an AI video generation model that it says was built in India under the government’s IndiaAI Mission. The company is positioning it as an India-first alternative to global models from OpenAI, Google, and Chinese startups.
The core bet: cheaper inference
The startup’s argument is not that Varya is smaller in the usual sense. Co-founder and CEO Sravanth Aluru says the company did not simply shrink the model to cut costs. Instead, Varya keeps a 14-billion-parameter footprint, the same size as its teacher model, but changes how the system reasons through video generation.
Traditional diffusion-style video models refine output over many repeated steps. Avataar says Varya reduces that process to four steps. The first two steps, according to the company, focus on shaping motion and structure, while the last two produce the final frames. Under the hood, Avataar says it uses role-aware supervision, distribution matching, and classifier-free guidance augmentation.
Speed and cost claims
On an NVIDIA H200 GPU, Avataar says Varya can generate a five-second 720p video in about 45 seconds. It says the same task on Wan 2.2, the base model it uses, takes about 1,230 seconds. That is the basis for its claimed 27x improvement in speed and cost versus the teacher model.
The company also says Varya can generate video for as low as ₹0.50 per second, or ₹0.48 in one part of the article, which it frames as far cheaper than current market options. The article compares this with models such as Google’s Veo 3.1 Standard, OpenAI’s Sora 2, Runway Gen-4.5, and Kuaishou’s Kling 3.0, all of which are described as materially more expensive per second.
Why this matters now
The story uses OpenAI’s Sora as a cautionary example: impressive output, but weak economics. The broader point is that AI video may fail as a mass market if generation costs stay too high. Avataar is betting that the winner will be the model that can do enough with less compute, not the one that burns the most.
For India, that matters because cheaper video generation could make the category more accessible to smaller businesses, teachers, and independent creators.
Key points
- Avataar launched Varya, an India-built AI video model under the IndiaAI Mission.
- The company says Varya can generate video for about ₹0.50 per second, far below major rivals.
- Varya keeps a 14-billion-parameter footprint but cuts the generation process to four steps.
- Avataar says the model can make a five-second 720p video in about 45 seconds on an NVIDIA H200 GPU.
- The article frames AI video as a market where economics, not just model quality, may decide the winners.
If Avataar’s cost claims hold up, AI video could become practical for more Indian users instead of just large companies with huge budgets. That would give creators, educators, and businesses a cheaper way to make video content. A strong result here would also show that Indian teams can compete by making AI more efficient, not just by building bigger models.
The model’s promise depends on whether its quality stays strong at lower cost. If the faster four-step approach produces weaker videos, the efficiency gains may not translate into real adoption. The article also shows that AI video remains a difficult business even for major global players. If users do not find enough value, cheaper inference alone may not be enough to build a lasting market.
