While AI video generation has been dominated by deep-pocketed global companies like OpenAI, Google, and top Chinese firms, Indian startup Avataar is challenging this trend with its novel approach to reducing GPU compute needs and inference costs by an order of magnitude.
- Varya generates 5-second 720p videos 27X faster and cheaper than its base model
- Avataar retains a 14-billion parameter model size, emphasizing quality over downsizing
- OpenAI’s costly Sora model failed due to prohibitive inference costs and dwindling users
What happened
Avataar, a twelve-year-old Bengaluru-based startup backed by Peak XV, introduced Varya, an AI video generation model that drastically reduces the cost and time of creating AI-generated videos. Unlike bigger players relying on large compute budgets and massive models, Avataar focuses on optimizing the inference process. Varya reduces the standard 50-step diffusion video generation process to only four distinct steps, leading to significant speed and cost efficiencies.
Built on Alibaba’s open-source Wan 2.2 architecture, Varya maintains a 14-billion parameter size, matching its teacher model to preserve quality. This approach enables Varya to produce a five-second 720p video in roughly 45 seconds on an NVIDIA H200 GPU, compared to the base model's 1,230 seconds. The innovation delivers a 27-fold improvement in speed and associated cost, making AI video creation accessible at around ₹0.50 per second.
Why it matters
The AI video generation market has struggled with exorbitant costs, limiting widespread adoption despite technological advances. OpenAI’s Sora, once highly anticipated, collapsed after generating videos at around $1.30 per 10 seconds and failing to attract sustainable users and revenue. Similar cost barriers affect other global models from Google and Chinese startups.
Avataar’s model tackles a critical industry pain point by radically reducing inference expenses without compromising on video quality. This cost-efficiency positions the startup as a potential leader in a market currently dominated by capital-intensive giants. Moreover, as India pushes AI innovation through initiatives like IndiaAI, Varya exemplifies homegrown solutions competing on a global scale.
What to watch next
Key developments to monitor include Avataar’s ability to scale Varya’s adoption in both domestic and international markets, particularly among content creators, enterprises, and platforms seeking affordable AI video tools. The startup’s success may inspire further innovation around compute-efficient video models and potentially trigger shifts away from brute-force compute strategies.
Additionally, tracking competitive responses from industry giants like OpenAI, Google, and Chinese players is important, especially if they pivot towards cost-cutting model restructuring or new architectures. Partnerships, funding rounds, and integration of Varya’s technology into broader AI content ecosystems will be vital indicators of Avataar’s impact in the evolving AI video generation landscape.