"I Built My Own AI Video Model, Now Make $3M a Month," Akool CEO Jeff Lu

Jeff Lu quit Google Cloud in 2022 to build his own AI video models at Akool, a generative AI platform for avatars, face swap, and video translation. He skipped the giant foundation models and trained small ones for specific tasks on rented servers, then watched revenue go from $100K in 2022 to $3M in 2023, $13M in 2024, $25M in 2025, and over $3M in June 2026 alone. He breaks down why Coca-Cola picked him over the foundation models and now pays him $4m per year. He'll share how he's driven huge usage (500,000 images and videos created daily) and what he'd say if Google offered $500M cash. Akool revenue data: https://getlatka.com/companies/akool.com

• 0:00Speaker 00

  I left Google Cloud in two thousand twenty-two. I have been working on the generative AI for images and videos. I decided to full-time create a job at Google and start a company.

• 0:14Speaker 01

  So how many videos are you are your users altogether free and paid creating per day on a cool today?

• 0:19Speaker 00

  About half a million, I would say, per day, images, videos and all the assets. For twenty-five, our recognized revenue is about twenty-five million dollars. And we are more than double this year.

• 0:36Speaker 01

  Hey folks, my guest today is Jeff Liu. He's the founder and CEO of Akool, a generative AI video platform for avatars, video translation and real-time video. He's an AI and computer vision veteran with over ten years of experience in this space, having previously worked at Microsoft, Apple, specifically on Face ID, and Google Cloud. Jeff, you ready to take us to the top?

• 0:57Speaker 00

  Yeah, thank you for inviting me here today.

• 0:59Speaker 01

  You bet. Okay, what year did you leave Google Cloud or Apple to launch Akool?

• 1:03Speaker 00

  So I left Google Cloud in two thousand twenty-two to launch Akool. So it's kind of at the beginning of this generative AI wave.

• 1:14Speaker 01

  And what were you seeing at Google Cloud at the time that sort of gave you the idea for this? What was the initial product you launched?

• 1:20Speaker 00

  I have been working on the generative AI for images and videos for a pretty long time. And what we see is that the technology is becoming mature. At that time, Google is pretty conservative, like generative AI, especially on videos and so on. They think they cannot control the results and they want everything to be one hundred percent under control. So I decided to full-time create a job at Google and start a company. But before I was kind of part-time work a little bit, it's during the COVID and so on, yeah.

• 1:57Speaker 01

  There's so many people building AI tools. They say it's the best. I asked you ahead of time, prepare a screen so you can share and show us the product. Show us what you're building.

• 2:05Speaker 00

  So the thing that we initially get popular is the swap features. So we have a whole swap series. You can swap the face, swap the character in the video, swap the motions, swap the head and swap the hair and so on. Then we get our avatar features, very popular, especially around the on the B two B side. Our most popular feature is actually avatars. And you can create static avatars and you can create live interaction avatars. You can create your own version of the avatars and so on. This is the most popular feature on the B two B side.

• 2:45Speaker 00

  The foundation models, which we do some fun tunes, are also very popular. And more recently, we launched something called Agentic Canvas. And it's a new feature and I think it's also getting popular. And basically you can use AI agents to help you to create a long format of the videos. And you can just type and it will create the video by itself.

• 3:14Speaker 01

  My question is, a lot of people are gonna go, Jeff's not doing anything special. There's a bunch of tools I can use to do a face swap. But you've actually, you're not just using, you know, other foundation models, you're actually fine-tuning. Now, you're a technologist. Explain to the marketer watching this, why that's valuable to them? Why your fine-tuning adds value to their life?

• 3:34Speaker 00

  Yeah, yeah. For the swap series of the models, we're actually not just fine-tuning. We are designing the model from scratch and doing the models ourselves. So it's a pretty different model architecture and training data size and so on. So the result quality and how it handles the corner cases and the boundary cases are very different.

• 3:54Speaker 01

  Can you just educate us? When you say you're building your own foundation models, most people think you have to have billions of dollars to do that. How are you doing it?

• 4:02Speaker 00

  We are not building the largest video foundation models. We are building like video models for specific tasks. So for example, the avatar, you generate humans in the video, human generation for only that task. And then the model is much smaller. And also we create a swap series. You make changes to humans in the videos and these models are much smaller than the large foundation models that generate everything. So that's how we control the cost and control the budget and make things move.

• 4:31Speaker 01

  You've been pretty transparent about your revenue growth. Are you comfortable sharing what revenue is today?

• 4:35Speaker 00

  For twenty-five, our recognized revenue is about twenty-five million dollars. And we are more than double this year. So we are expecting to get sixty or something, yeah.

• 4:50Speaker 01

  So you anticipate finishing in December twenty twenty-six with about sixty million of ARR?

• 4:54Speaker 00

  Revenue, not ARR.

• 4:56Speaker 01

  Tell me how you measure.

• 4:57Speaker 00

  Yeah, so revenue, you recognize the revenues and add the trailing twelve month fees. So that's how it works. And ARR, you probably use the last month's multiply twelve, that's revenue run rate. So it's a different way of measuring it, so, yeah.

• 5:15Speaker 01

  Jeff, I love that about you, right? That's the quickest way to tell if I'm interviewing a marketing founder or a technical founder. All the marketing founders take last month's revenue multiplied by twelve. You're conservative, you're looking back, which I think is awesome. What does that mean in terms of, we're recording this in July. If we just look at last month's revenue, how much did you do just last month?

• 5:34Speaker 00

  Last month's revenue, I think it's over three million dollars.

• 5:37Speaker 01

  That's incredible. How did, what was that like? I mean, did you ever expect when you launched this thing three, four years ago, you'd be doing three million dollars a month?

• 5:44Speaker 00

  It's actually pretty different when we just get started. We're trying to focus on B two B purely and we cannot get into mix of B two B and B two C and so on. But our goal is still to bring this to a very big company. So I think we are still on the way going there.

• 6:02Speaker 01

  What does your actual tech stack look like? I mean, you have a bunch of Mac Studios running in a closet or H two hundreds somewhere? Are you renting from one of the inference providers? Like, how do you do your training?

• 6:11Speaker 00

  We rent servers from the cloud providers and then we do training on these servers. We use multiple worms, but Lambda is used a lot and some others that will depend on the workload and the needs and so on.

• 6:26Speaker 01

  Since you're a user of those platforms, I'm trying to figure out who's gonna win that space. Who do you think is gonna win that space?

• 6:32Speaker 00

  I think on the infrastructure side, there will be multiple winners. It's, it's very hard for a company to dominate. So that's how the B two B side works. And if you are doing B two C, then probably there will be a bigger player, so, yeah.

• 6:50Speaker 01

  Other folks listening right now are trying to like build their own version of sort of foundation models. And I think that my audience generally, they know what a prompt in Claude is, right? Or ChatGPT. And then they say, well, I want my prompt out of memory. So then they upload a bunch of MD files, but then they hit context window issues and hallucinations in Claude and ChatGPT. So then they say, maybe I should host my own files, do my own training. And then it gets really complex really, really quickly. You've obviously pushed that pretty far the other direction. Do you have, do you have these same issues like context window issues and hallucination issues? When you're building your own foundation models for specific use cases?

• 7:23Speaker 00

  In the video space, I think that is a little bit less concerned. The video space, the main thing in how to make it more realistic, making it realistic is the biggest challenge. And also make it accurate to the prompt. In our case, since we are doing more specialized video models and we generally don't experience these issues.

• 7:47Speaker 01

  I published my book in twenty nineteen, sixty thousand words. We sold a lot of copies, Washington Bestseller List. How close am I to be able to use Akool to update, upload the PDF of my book? Plus maybe the audiobooks, you have my voice, plus an image or two of me and you make it a video?

• 8:03Speaker 00

  That depends on what kind of video you want to make. If it's an avatar video, talking about reading the books, that's pretty simple. It's like a story and you want to see a scenery of the what the book tells. I think, I think we can already do quite many of these works, piece by piece and should not be far to get the whole book to work, yeah.

• 8:24Speaker 01

  When you launched this business back in twenty twenty-two, how did you get the first hundred customers?

• 8:28Speaker 00

  When we first launched, we tried to close some B two B customers and some larger ones and the process is very slow. And there are lots of users come using it by itself. So we just put a payment gateway on it and then people begin to pay. We get some organic traffic and people come by themselves and just pay by themselves online.

• 8:53Speaker 01

  But really try to remember, I mean, when you launched, you didn't have traffic. Now today, you're getting two hundred and ten thousand organic clicks a month based off AHREF's data that I'm showing here on my screen right now. How did you get the traffic in the early days? How did people know about you?

• 9:07Speaker 00

  In the early days, we mainly get the data from the, get the user from the community. So quite many communities like Reddit and Facebook groups and Discord and so on. And we kind of show lots of the product demos.

• 9:28Speaker 01

  So I see you have your own Reddit channel, r forward slash Akool underscore official. Was this the major driver of your first hundred customers or no?

• 9:36Speaker 00

  No, this subreddit was created much more recently.

• 9:40Speaker 01

  So you are just posting in random other subreddits?

• 9:43Speaker 00

  In related communities for showing demos, without demos and so on.

• 9:49Speaker 01

  I see. Okay. What else did you do to get your first hundred customers? So Reddit communities, what else?

• 9:54Speaker 00

  There are some user groups already using video creation tools. And we try to talk with these customers and frequently schedule like one-on-one meetings and listen what they need and recommend them tools and so on.

• 10:12Speaker 01

  Can you name one or two of those communities?

• 10:14Speaker 00

  So Fiverr had lots of their people using, helping others to create videos. And we also talk with the people on the Fiverr.

• 10:23Speaker 01

  So you told them, hey, look, you're on Fiverr giving people video services like face swap. You should use Akool to get that work done faster.

• 10:31Speaker 00

  Uh, yeah, we did that work as well. So we did that before.

• 10:36Speaker 01

  Did you write the first line of code in 2020? Are my notes right?

• 10:39Speaker 00

  Uh, yeah, about right, 2021 or 2020 or something like that. Yeah.

• 10:43Speaker 01

  Okay. What did you finish 2021 at with total revenue?

• 10:46Speaker 00

  It's very small, around a hundred K or something.

• 10:48Speaker 01

  Okay. And in 2022, what did you finish with?

• 10:51Speaker 00

  Also about a hundred K.

• 10:53Speaker 01

  So pretty flat.

• 10:54Speaker 00

  Because the first 2021 is the part time and 2022 is I just get out. I think at that time, we spent most of the time pursuing large contracts. But was not having difficulties working on the clothing type of the large contracts. And it didn't grow much.

• 11:17Speaker 01

  And what did you finish with in 2023?

• 11:19Speaker 00

  Two to three million dollars.

• 11:21Speaker 01

  My notes tell me that that was your first big enterprise account. You landed Coca-Cola and they ran a six-month campaign swapping several million faces. Tell me how you closed that deal.

• 11:31Speaker 00

  So that team bombed and they found us and they see our technology is the best on the market. And so they benchmarked a lot, so they decided to work with us.

• 11:45Speaker 01

  Did you have AEs on your team that are looking for all the prosumers that sign up for a cool for 50 bucks a month? And if you see someone at coca-cola.com, you proactively reach out to them? Or how does that actually work?

• 11:56Speaker 00

  For this case, at that time, we don't have. And it's fully they reach out to us and they even didn't disclose their identity at the beginning. Until like from the first conversation to get engaged is about three months in the middle. So they spent three months benchmark testing different stuff and the devining and so on. Yeah. So, but we get our first A account, A E in 2024. And then begin to do more of the sales and the management work and so on.

• 12:31Speaker 01

  And so what did you finish 2024 with in terms of revenue?

• 12:34Speaker 00

  Around 12, 13 million dollar.

• 12:37Speaker 01

  You said in 2025, you finished with about 25 million of revenue. And you're going to more than double that here in 2026, right?

• 12:44Speaker 00

  Yeah, yeah, yeah, that's right.

• 12:45Speaker 01

  Walk me through the team size today. How have you structured it and how many folks are full-time?

• 12:49Speaker 00

  Our team is pretty dynamic. It depends on the needs and so on. So we have 60 to 70 people on board.

• 13:00Speaker 01

  How many are engineers, like 40?

• 13:02Speaker 00

  Yeah, around that.

• 13:03Speaker 01

  Okay. And do you have any quota-carrying sales reps or no?

• 13:06Speaker 00

  Yeah, so we have about four, five quota-carrying sales reps. And it's pretty standard to give one million dollar quota per person.

• 13:16Speaker 01

  So I guess how many total customers are you serving today? And what's the average customer paying per month?

• 13:21Speaker 00

  On the prosumer side, we have a lot. We have 10,000, 1000s of customers. And on the enterprise and SMB side, we have several hundred.

• 13:32Speaker 01

  So is it fair to say all of your paying customers added up together? Or you have more than 30,000 paying?

• 13:37Speaker 00

  I think so, like 20 to 30,000, around 30,000 maybe.

• 13:41Speaker 01

  But most of your growth is coming from the enterprise sales?

• 13:44Speaker 00

  More recently, yeah.

• 13:45Speaker 01

  Like, how much value can you add to one enterprise account if they start off paying 50 bucks a month? Are you comfortable, don't share the name, but are you comfortable sharing your largest contract value today?

• 13:55Speaker 00

  It's actually usage-based, but they pay us like around four million dollar a year. And

• 14:04Speaker 01

  You have individual customers paying four million a year?

• 14:08Speaker 00

  No, no, no. We are contract on the B2B side.

• 14:10Speaker 01

  That's still pretty incredible. What makes them scale so quickly? When you say usage-based, is it number of credits? How do you, what value do you assign to each credit? How do they consume the credits?

• 14:21Speaker 00

  We provide credits and to the users that's using their platform. And it's usage-based and some of them actually integrate into their software. And when they scale up their usage, they also scale up the API course. It's all through API usage.

• 14:41Speaker 01

  Jeff, I have to tell you this respectfully. I asked this question because when I look at your pricing page, I mean, this thing looks like shit. I'm so confused. I'm like, where do I go? There are so many sliders. What do I pay for? But this is working, huh?

• 14:55Speaker 00

  So, yeah, I think we provide one more tier. So before, we only provide Pro, Pro Max and the Business. And we just recently added a starter plan. So we kind of give people more options to get started. Yeah.

• 15:12Speaker 01

  So how many videos are your users altogether, free and paid, creating per day on a cool today?

• 15:18Speaker 00

  About half a million, I would say, per day. Images, videos and all the assets created per day on the cool.

• 15:24Speaker 01

  Is this bootstrapped or have you raised capital?

• 15:26Speaker 00

  Uh, we raised capital, but it's a small amount.

• 15:28Speaker 01

  When did you raise capital and how little was it? What year?

• 15:31Speaker 00

  We raised in total about 10 million dollars across our last four-plus years. And the last price round is about two years ago. And we are working a new price round and so on.

• 15:46Speaker 01

  Interesting. Are you comfortable sharing what that valuation was in 2024, the last price round?

• 15:51Speaker 00

  Around a hundred million dollars, two years ago. We haven't really raised the price round, but we take some safe notes at a much higher valuation. And we are thinking about a price round.

• 16:04Speaker 01

  I mean, what's the market looking like today for valuations on companies growing as fast as yours? You're having conversations. I mean, do you think you can raise it a billion-dollar valuation?

• 16:13Speaker 00

  So the market is constantly changing on the valuation side for the company. A billion, maybe a stretch, but I think probably like half of it or something like that. But a billion could be a little bit stretch on the current market. So, yeah.

• 16:29Speaker 01

  If your old bosses at Google come back and say, Jeff, we want you back. Here's 500 million all cash upfront. We want to acquire a cool. Do you take the deal?

• 16:40Speaker 00

  Uh, likely, I will.

• 16:45Speaker 01

  I love that you're honest. Most founders go, no, Nathan, never. I want to build this forever. And you're like, half a billion cash upfront, I'll take it. All right. Well, Jeff, as we wrap up here, you're a technologist by training. When you look at the technical side of pushing video creation forward, the next leap, what's the current bottleneck on the technical side?

• 17:05Speaker 00

  The videos are still pretty hard to create a long format and realistic videos and so on. How can we create one hour movie easily? And how can we like make everyone able to create it? There's still some gaps over there. The foundation model technology is growing pretty fast. So I think pretty soon, people will be able to create very good movie quality videos by themselves.

• 17:34Speaker 01

  Guys, scientist from Google in 2022, he saw where video was going and said, I'm quitting Google. I'm going all in on my side project. Last month, which was June 2026, did $3 million of total sales. And users are now creating over half of million videos and images per day on his platform. Jeff, thank you for taking us to the top. If you enjoyed that interview, you should click here to subscribe to the channel. So you don't miss new interviews every Tuesday. If you've got some free time and want to listen to the next interview right now, click right here. Lastly, subscribe wherever you get your podcasts. We're on Spotify, Apple Podcasts and here on YouTube.