4.5 Million Videos a Day: How Higgsfield Went from $0 to $300M ARR without Building a Foundation Model with Higgsfield's CEO and Co-Founder

Every AI company is racing to train their own foundation model. Alex Mashrabov did the opposite. Higgsfield orchestrates existing models, Sora 2, GPT-4.1, and others, through a workflow engine that turns a product link or text prompt into a cinematic video ad in seconds. The result: 4.5 million videos produced daily, $300M ARR in roughly a year, a $1.3B valuation, and 15 million users. All without training a single model. Before Higgsfield, Alex launched MyAI at Snap, which hit 150 million users. He's now scaled a consumer AI company twice, and this time he did it by betting that orchestration beats ownership. In this session, Alex will cover: Why he chose to build on top of other people's models instead of training his own, and why that decision was the key to hitting $300M ARR in a year The orchestration architecture behind Higgsfield: how they connect multiple foundation models into a single workflow engine that produces 4.5M videos a day How Higgsfield ships 300+ releases a year (6 per week) and why that velocity is only possible because they don't own the model layer The economics of building on someone else's infrastructure: what the margins actually look like and how they manage cost at that volume How they reverse-engineer virality, analyzing short-form video patterns at scale and encoding them into reusable creative structures What he learned scaling MyAI to 150M users at Snap and how that shaped how he built Higgsfield This session is for you if: You're a founder or technical leader deciding whether to train your own model or build on top of existing ones, or you're trying to figure out how to scale an AI product to millions of users without burning through capital on GPU clusters.