Why Avataar's Varya AI Video Model Matters for NVDA and BABA

Avataar, a Bengaluru-based startup backed by Peak XV, launched an AI video generation model called Varya today. The company claims it can produce video at ₹0.50 per second, roughly $0.0057 at current rates. That is 17 times cheaper than Kling, the global market’s current cost champion, and more than 130 times cheaper than Google’s Veo 3.1 Standard, the company said.

The launch comes after OpenAI shut down Sora, its standalone video app, in March 2026. Sora’s economics had been brutal: each 10-second clip cost about $1.30 to produce, and lifetime revenue reached just $2.1 million against daily inference costs around $15 million, according to public data the company released. A reported $1 billion partnership with Disney fell apart. Active users dropped below 500,000 by February 2026.

The structural problem Sora exposed is simple: high-quality AI video is too expensive for most users. Google, OpenAI, Runway, and Kuaishou all charge between $0.10 and $0.75 per second. A 30-second video can run $3 to over $20. For small businesses, educators, and creators in price-sensitive markets like India, those numbers block adoption entirely.

Avataar’s cost advantage does not come from shrinking the model. CEO Sravanth Aluru told Inc42 that most distillation projects compress a 70-billion-parameter model to 7 billion, sacrificing quality. Varya retains the full 14-billion-parameter footprint of its base model, Alibaba’s open-source Wan 2.2. What Avataar changed is the inference process. Standard diffusion models take roughly 50 sequential denoising steps. Varya collapses that to four steps, with each step handling a distinct function: the first two shape trajectory and motion, the final two generate output frames. On an NVIDIA H200 GPU, Varya produces a five-second 720p clip in about 45 seconds. Wan 2.2 requires roughly 1,230 seconds for the same task, a 27x improvement in speed and cost, Aluru said.

The founder argued the model solves a demand problem, not a supply one. “There is a huge audience sitting with ideas but without affordable tools to express them through video,” he said. Avataar sees three user groups: enterprises that can fine-tune Varya on proprietary data for marketing and catalogues, creators and small businesses who can pay on a per-second basis, and India’s 1.5 million-plus schools. Aluru is especially bullish on education, where teachers could generate visual content without expensive production resources.

Peak XV managing director Ranjan Anandan framed the launch as part of a broader pattern. “India has never built leadership in any technology area by following what the West has done, simply because we can’t afford it,” he said. “When it comes to AI, that playbook of doing it differently, aiming for population scale, dramatically lower cost, making it culturally relevant and Indian contextual, is what’s going to be needed.”

Avataar was founded in 2014 focused on 3D augmented reality for e-commerce. It raised $45 million from Tiger Global and Sequoia Capital India in 2022, and counts brands like Sleep Number and Bajaj Auto among its customers. The company’s core competency – understanding how objects and spaces interact computationally at scale – became groundwork for the video generation problem, Aluru said.

The company has not yet published a technical paper detailing Varya’s architecture or third-party benchmarks. The quality claims remain self-reported. For traders watching the AI infrastructure space, the key question is whether a 14-billion-parameter distilled model can match the output quality of much larger systems at a fraction of the compute cost.

NVIDIA benefits from every GPU deployment, and Avataar’s use of H200 chips supports that narrative. Alibaba’s open-source Wan 2.2 gives BABA a foothold in the model distribution layer, even if the startup itself remains private. Alpha Score data shows NVDA at 67/100 and BABA at 46/100, both in moderate-to-mixed territory. The real test comes when independent reviewers run the benchmarks.

Avataar plans to release the technical paper in the coming weeks. Until then, the model wins on cost alone.