AI video startup reached $25 million revenue on $10 million raised

1. Narrow models, lower cost — ▶ 3:54

   A Cool does not build the largest video foundation models. It builds models for specific tasks: generating humans for avatars, and modifying humans in existing footage through its swap series. Lou says these are much smaller than general models that generate everything, and that this is how the company controls its budget. For swap, he says the team designs models from scratch rather than fine-tuning, with different architecture and training data, which he claims changes results on corner cases. Training runs on rented cloud servers, often from Lambda. He also says context-window and hallucination problems matter less in video than realism and prompt accuracy.

2. Trailing revenue, not run rate — ▶ 4:54

   Lou measures revenue as recognized revenue over the trailing 12 months, not last month multiplied by 12, which he calls run rate. On that basis, 2025 revenue was about $25 million, and he expects to more than double it this year, to roughly $60 million, a figure he offers loosely. Last month's revenue was over $3 million. Editorial inference: $3 million a month annualizes to about $36 million, so the forecast implies much larger months ahead. The transcript does not explain how that acceleration will happen.

3. Payments before sales team — ▶ 8:28

   Early large B2B deals closed slowly, but the product already had users arriving on its own. Lou's response was to put a payment gateway on it, and people began paying online. Those early users came from posting product demos in communities such as Reddit, Facebook groups and Discord, and from one-on-one conversations with people already doing video work, including freelancers on Fiverr. The company's dedicated subreddit came much later. The pattern is demand discovered through use, not created by outbound selling.

4. Flat year, then Coca-Cola — ▶ 10:43

   Revenue was around $100,000 in both 2021, when Lou worked part-time, and 2022, when he left Google and spent most of his time pursuing large contracts that did not progress. It then reached $2-3 million in 2023, $12-13 million in 2024 and about $25 million in 2025. The 2023 jump included Coca-Cola, which Lou says came inbound without disclosing its identity, then benchmarked the company for about 3 months. The first account executive arrived in 2024, so the largest early deal was won without outbound sales.

5. Usage-based enterprise economics — ▶ 12:42

   Lou describes a team of 60-70 people, about 40 of them engineers, with four or five quota-carrying reps at $1 million quota each. Customers are tens of thousands of prosumers, roughly 20,000-30,000 of them paying, plus several hundred enterprise and SMB accounts. The largest B2B contract is usage-based, about $4 million a year. Some enterprise customers integrate the platform into their own software, so their API usage rises as their own usage scales. Growth is coming more recently from enterprise.

6. Small raise, candid exit math — ▶ 15:24

   A Cool has raised about $10 million over 4-plus years. Lou puts the last priced round, about 2 years ago, at around $100 million. The company has since taken SAFE notes at a much higher valuation and is considering a new priced round. He thinks about half a billion dollars is plausible and a billion a stretch in the current market. Asked whether he would accept $500 million in cash up front from Google, he said he likely would, which is unusually direct for a founder.

7. Long-form realism still unsolved — ▶ 17:02

   Lou names the current technical bottleneck as creating long-format, realistic video, such as a 1-hour movie that anyone can make easily. He says gaps remain, but foundation-model technology is improving quickly and he expects people to create movie-quality video by themselves fairly soon, without giving a timeline. A Cool's newer agent-based canvas, which creates long videos from typed instructions, is an early attempt at the problem. Lou also says a book-to-video workflow is doable piece by piece and should not be far from working in full.

Jeff Lou's central claim is that a generative video company does not need a frontier-scale model to build a large business. It needs a narrow task and customers who find it. A Cool trains models for specific jobs: generating humans for avatars, and modifying humans in existing footage through its swap series. Lou says these models are much smaller than general video models that generate everything, and that this is how the company controls cost. By his account, the result is about $25 million in recognized 2025 revenue on roughly $10 million raised over more than 4 years.

The origin story fits that logic. Lou says that when he left Google Cloud in 2022, Google was conservative about generative video because it believed it could not control the results and wanted everything under 100% control. He saw the technology maturing and acted. That is his characterization of an incumbent's risk posture, not an established fact. Narrow models carry a related appeal, since a bounded task is easier to make reliable. Lou says the hard problem in video is realism and accuracy to the prompt, not the context-window and hallucination issues familiar to text users, and that his specialized models generally avoid those.

How customers arrived is as notable as what was built. Lou says early enterprise sales were slow, so he put a payment gateway on a product that already had users, and people began paying online. Early users came from demos posted in communities such as Reddit, Facebook groups and Discord, and from conversations with people already doing video work, including freelancers on Fiverr. Revenue was roughly $100,000 in both 2021 and 2022, the latter spent chasing large contracts. Then came $2-3 million in 2023, $12-13 million in 2024 and about $25 million in 2025.

Enterprise business followed product pull, not the other way around. Lou says Coca-Cola approached the company without revealing who it was, then benchmarked it for about 3 months. The first account executive was hired only in 2024. Today the company has four or five quota-carrying reps at $1 million each, 60-70 staff and about 40 engineers. Its largest contract is usage-based, at around $4 million a year, and some customers embed the product in their own software, so their API usage grows with their own. Lou counts tens of thousands of prosumers, with roughly 20,000-30,000 paying, and several hundred enterprise and SMB accounts.

The evidence has limits. Every number here is self-reported, and some do not reconcile cleanly. Lou says last month's revenue was over $3 million, which would annualize to about $36 million, yet he forecasts roughly $60 million this year, a figure he offered loosely. His trailing-12-month method means the two need not conflict, but the path to that acceleration goes unexplained. The host also found the pricing page confusing, and Lou says a starter tier was recently added to the existing pro, pro max and business plans. Nothing in the discussion tests how durable the advantage of narrow models is against larger general ones, a gap that Lou's own comment that foundation-model technology is growing fast makes relevant.

Lou's financial posture is consistent with this realism. The last priced round, about 2 years ago, valued the company at around $100 million. Since then it has taken SAFE notes at a much higher valuation and is considering a priced round. He thinks about half a billion dollars is plausible and a billion a stretch. Asked whether he would sell to Google for $500 million in cash, he said he likely would. His stated goal is still to bring the product to a very large company. The two positions are compatible, and a founder who has raised only about $10 million faces a lower bar for a rational exit.

The larger issue is where video generation is heading. Lou names long-form, realistic video, such as a 1-hour movie anyone can make, as the open technical problem, and expects movie-quality results from individual creators fairly soon. A Cool's own answer is an agent-based canvas that builds long videos from typed instructions, plus specialized models used piece by piece. Whether long-form video emerges from that kind of assembly, or from general models that make assembly unnecessary, will determine if the narrow strategy is a durable position or only an efficient early one.