AI Video for Half a Cent: Varya Is 20x Cheaper

Every time I generate a video clip for a course lesson or a post, I glance at the meter — AI video is the one modality where the bill still stings. So this caught my eye: Indian startup Avataar AI, backed by Peak XV, just released Varya, a video generation model priced at ₹0.48 — half a US cent — per second of generated video. For comparison, Google's Veo, Kling, Luma, and Runway all charge from $0.10 per second and up. That's roughly a 20x price gap for the same basic job: type a prompt or drop in a reference image, get a video.

The speed numbers are just as loud. Varya produces a 5-second 720p clip in 45 seconds on a single Nvidia H200 GPU. The model it was distilled from — Alibaba's open Wan 2.2 — needs about 1,230 seconds for the same output. Ten times faster, because Varya runs generation in 4 steps instead of Wan's 50.

Distillation is eating the price list

Here's the part I find genuinely important. Avataar didn't train a frontier model from scratch — nobody hands a startup that kind of GPU budget. They took an existing open model and distilled it: compressed the big "teacher" into a small, fast "student" that keeps most of the quality at a fraction of the cost. It's the same playbook that made cheap chat models possible, now arriving in video. Avataar is one of 12 startups in the India AI Mission, a roughly $1.2 billion government program that subsidizes compute in exchange for releasing models publicly — so Varya's weights are open on the AI Kosh portal, and a public beta is live on the company's site. The model is also deliberately trained on Indian festivals, food, clothing, and architecture, because global models tend to produce stereotyped, tourist-brochure versions of all of it. Rajan Anandan of Peak XV put the thesis in one line: "Cost is the biggest unlock for AI adoption in India."

Why should you care if you're not in India? Because prices don't stay regional. We watched this exact movie with chat subscriptions: budget tiers launched in emerging markets first, then the price war came home to everyone. Video is the most expensive thing you can generate today, which is precisely why it's the next thing to get crushed by distillation. When a 720p clip costs half a cent to make, video generation stops being a premium feature and becomes a default button in every app you already use — presentations, marketplaces, family photo apps, all of it.

My take: the interesting race in AI video is no longer only about who makes the prettiest pixels — it's also about who makes them a hundred times cheaper. Distillation just proved it can compress the price by 20x in one release. The pretty pixels will follow.