Session: 2026: The Year UA Could Become Good, Fast, & Cheap? - Alex Engel, Layer AI.

2026, the year UA could become good, fast, and cheap. And this is from um Alex Engel from our track sponsors, Leia AI. So, round of applause, please, for Alex. [applause]
You got it. Thanks so much. Appreciate it. So, yeah, we're going to we're going to solve UA here today. It's going to be great. Um we're never going to have to worry about UA again. You're going to have to worry about discovery or anything like that. Uh my name is Alex. I'm VP of product for layer.ai. Um, and I'm here to advance the slide.

And the slide isn't advancing. There we go. Perfect. Uh, layer is the AI operating system for modern games. Uh, what we do is we do a whole bunch of things. We make your videos, we make your images, we make your audio, we make your, uh, 3D models. We are used by a wide variety of people that are here, wonderful customers here today, as well as a number of other people in the industry. So we are bestass models, workflows and enterpriseg grade infrastructure. Does a whole bunch of things 1.65 million images generated every month. 9,000 hours of videos generated every month. Trusted by indies and enterprise alike from 1 to 300 plus seats. You can have up to infinite seats. All the latest AI models in one place. We're sock 2 compliant. That is the last time I will pitch all of our amazing sales team. I'll let them do the best of it after me because I just do the product stuff, man. and I just make things move. Uh, what studios are doing in 2025? This is really interesting. So, 2025 is a weird world. There's been some

significant changes in the way people are using AI in the video game space, especially in the mobile game space. Frankly, we're kind of just doing magic right now, which is really bizarre. So, we're seeing a lot of weird stuff. So, yeah, I was like, dude, I totally want to. So, I've secretly, not so secretly, I've been working on a custom model, which is all based around the worst of 1970s and '8s fantasy art. Uh, so lots of terrible stuff here. So, obviously the dinosaur on the right. That's what I want. Weird proportions, strange things going on. Teeth that don't quite work. Uh, generally fighting it with a big sword. That's that's my jam. I love that stuff. So, I want to go like rough sketch right there. Right. So, really rough sketch coming in right there. And then I'm going to render it with Gemini 3 uh with some style references. And then I'm going to go ahead and pump it through a custom uh model right there, which is something I created. Whoops. I'm just going to kick glasses around, too. uh 1970s fantasy style custom Laura 10 steps $180 2048 x 2048 pixel final 10 you know right there around $180 uh it

took me I think 40 seconds to make this image. Uh so that's where we're at today and pumping it through my own custom model. So that's pretty cool. Uh and if 2025 is about good and cheap, 2026 is all about going fast, right? We're going to do it, guys. We're going to make it to the center of the iron triangle. we may be on the verge of finally reaching it if we can build it correctly. That's the idea right here, right? Cuz we've all tried to do this before. Um we've all gone through I'm going to stay here. We've all gone through this and said, "Well, if only we could make a good, fast, and cheap, wouldn't that be great?" But of course, most of the times we just end up making something expensive, slowly, or poor quality. Could we actually solve that problem here today? Maybe. We'll see. uh custom Laura models right now on the good side and new training frameworks are creating style consistent results with as few as a dozen images in the training set right here. So what do we got right here? We got a whole bunch of these stylized 3Dish mobile gameish content coming in right here. We're pumping it through a

Koa training framework at 300 epics with 100 Laura rank here and then we're pushing it through Flux Swan Dev. Now, Flux1dev, not the most um uh cutting edge model, let's say. Very utilitarian, very cost effective, and frankly, pretty good. So, we're going to go ahead and we're going to dump 91 images in there. And then we are going to go ahead and generate a wizard. Of course, it wouldn't be a wizard if they weren't pondering an orb. Uh so, they're going to render the items in the scene right here. We're going to take this wizard pondering his orb in his I'd rather Spartan for a wizard uh library right here. And then we're actually going to use Gemini 3 to extract these assets. So we're going to go ahead and take this right here. We're going to say render the items in the scene. Furniture, clothing, books, accessories one by one. Place them in a 3x3 grid against a neutral background. Prioritize showing clothing and primary subjects first. Don't isolate facial hair, uh, hair or minor accessories like belts and shoes. The reason why we can do these these

days is because Google decided to roll out a bunch of reasoning systems, reasoning models that allow you to actually go ahead and have sort of a conversational uh discussion with your AI extraction model. So you could say, I really want to take this guy's hat and I want to take this guy's jacket and I I want his orb so I can ponder it myself and then I'm going to have all these other things here and we're going to extract those and we're going to keep them as style consistent assets that I can then integrate into stylesheets, sprite sheets, or anything else in my product. And you could do that across a wide variety of different models. On the top left, a 2D fantasy. On the top right, a more stylized fantasy. The bottom left an isometric Laura that we trained. and the bottom right more of a chibi anime style. And these are all just kind of standard things in here. And you get decent results for each one of these. Pretty solid. I mean, like I would play the bottom left game right there. I'd go ahead. That's cool. I have like a sticker game where you go ahead and you drag your wizard stickers onto a a blank room. That could be kind of fun going in right there. And each one of those, you're able to similarly extract

these items, bring them out, and you can do this at scale. But it's not just good, it's cheap, frankly. Uh it used to be back in 2023, 2024, V1, V2, you'd be looking around 50 seconds per sec, uh 50 cents per second of your generations. But now we're seeing even with VO3.1, which is I would consider one of the best video generation models out right now, 27 seconds or 27 cents per megapixel second of video. That's actually a 1080 by 1920 portrait uh cost right there. It's even cheaper if you use a model like Cling 2.5, which is about 10 cents a second. Uh Cling 2.1 VO31 fast 15 cents a second. Uh Seance Pro about 30 cents a second. Significantly cheaper even for the most expensive things you can do, which is generating HD video coming in right there. We've seen that come in extremely cheaply right here. So this on the right right here, a 4-se secondond video clip. It's not playing right now, but if it were to play, you would see a knife come

out of nowhere in the best hallucinatory AI fever dream fashion slice that car as if it was made of chocolate cake. 4 seconds right there. Total video cost about a$110. And that's of course converted uh from our unit cost on the platform to the USD cost right there. So things are getting cheap, right? So they're good. They're cheap. Are they actually going to get fast? Well, let's see right here. Is this a my video? Oh, that's too bad. Uh, going wide from the start is now possible. So, we think that this is what's happening in the near future. In fact, we could argue it's already here, which is why even bother having a sketch? Why even bother having storyboards? If the cost of going from a storyboard to a video is essentially the same, why have a storyboard? Why not just start with video? Why not generate four different assets for video right here? And then why even stop there? If cheap isn't just lower model prices,

it's scaling with automation to drive lower costs per iteration. Suddenly you realize that the cost of generating four images with GBT image 1.5, which I would consider one of the best uh prompt consistent models that's out there right now, is effectively the same price as a single video. As long as I know roughly what kind of video I want, why even bother with sketching? Now I just want to save time. Now I just want to save on iterations. Production time and cost are not the bottleneck anymore. Now it's where it's where we test, where it's safe to ship. We want to come in and say, "No, no, no. Straight to video. Let's go ahead and start releasing videos now and skip all the previs steps that we used to do right there, which is weird. It's a weird world." And it's only going to get faster because as we roll out node-based workflows, which Layer just rolled out today, you have the ability bored you. I promised that I was only going to have one slide. You tricked me. You got me in here. I'm going to I got to pump our own company. We just rolled out node-based workflows today. This morning right there, teams

are building full UA creative sets when they get a cup of coffee. And we'll talk about that a little bit more right here. What that means is that last November, we had Nodevember. every AI company out there was suddenly saying, "Dude, we've doing this thing that you used to be able to do a lot before, but now you can do it with AI. It's very exciting, which is you can go ahead and you can drag things through visual scripting, which is very exciting." And now you can take these different assets and pump them together. And you say, "But wait a minute, Alex. Uh, you keep kicking glasses around up here and you keep coming in and you keep talking about nodes. I've been using nodes forever. Sure, I've been generating height maps with Gaia and I've been building visual scripting and Unreal for many a year. Uh, what does that mean for me? Why would I care? Right? We've all been doing node-based workflows forever. Well, the difference is that with AI, suddenly you can start delegating actions to people. And we're starting to see this with some of our customers and some of the other people in the space which is I want to go ahead

and I want to generate not one video. I don't want to generate four videos. I want to generate uh 80 videos and I want those 80 videos to be built around different hypotheses and I want those hypotheses to be testable immediately. Right? These are things where as a product manager myself, it used to take me, you know, maybe a day for a uh a testing document where I would come in here, I'd create a hypothesis and brief and then I'd have to create 20 to 50 variants right there. Then those variant creations would take time. Myself, my designers, my artists, we'd come in there and generate all these things and then we would say, "This is going to take some time. Now we're going to have to go ahead and build out a bunch of these different variants." But we're seeing that flip now. Actually, the largest amount of time is spent on that hypothesis and brief, which is weird because that's not what I'm used to. I used to be doing a whole bunch of this testing in mobile games back when I worked in the mobile game industry as a product manager, building out and owning the P&L for a bunch of games. Suddenly,

that's the most important thing for me. Oh, I got to spend all this time writing this document. What do I think is going to work? I heard Christmas is popular. It's December. Surely we should build some Christmas UA, right? That would be I I think that sounds like a plausible hypothesis. We're in December. Uh theoretically Christmas does well in December. I think we've got some data to bear that out. Let's make that happen. So hypothesis and brief. Let's create 20 to 50 variants. Then let's make sure that we can export into all formats automatically. That means all aspect ratios, app store safe crop areas, and things like that. That means that we have to be able to export 1080 x 1080 for Instagram, you know, 1080 x 1920 portrait, 1920 x 1080 landscape. You want to throw some Google ad in there, 300 x 250, how you going to deal with that? You got a limited amount of space, but we can work with that. Maybe you want to do some fandom takeovers and do a 150 x 600 banner. All those things there were super annoying back in the day to make, right? You go into Photoshop, you build your templates, you

go into Figma, you get your Figma frame set up right there. Nah, we're going to skip all that. We're just going to let AI deal with all that stuff. Then we're going to use Gemini 3 and like we talked about earlier in some of the other talks, we're going to implement QA gates. We're going to have Gemini 3 and reasoning models double check the outputs and say, "Hey, did you really do that correctly?" And sometimes the AI models like, "You got me, man. You're right. You said Halloween." And I'm like, "No, I said Christmas." And it's like, "Oh, right. Christmas." Yeah, I totally That's what I meant. So, we can go ahead and add these QA agents. And so in the 5 minutes it took for me to go grab my coffee when I was doing this example, I generated 32 static image creatives, 8x video creatives, and they were all ready to go on the platform of my choice. So then I just expanded it even further, right? Like so this is all the time I saved right here. This is me running our workflow and this is what we think all of people are going to be doing in AI in the near future for UA, which is you're going to save all the base image generations. You're going to go ahead and compose those images into different aspect ratios. You're going to

generate CRM titles. You're going to go ahead and modify those images to incorporate them uh natively. You're going to go ahead and generate videos based on those things. And all those outputs are going to be automatically dropped into the UA platform and creative set of your choice, which sounds crazy, but we're here. So, let's talk about how we might actually solve the iron triangle. If we've got good and we've got cheap and we're working on fast, how will we in this room know when we've actually solved the iron triangle? We can go home. Robots automate all of our stuff. It's great. It's a it's a brilliant new world, right? But there's still a lot of work that's got to be done, right? First thing, you've got to deal with this agentic idiation issue, right? You got to be able to take UA data from places like Sensor Tower and App Magic, uh, automatically turn those into concepts, hooks, scripts, and tones for new creative sets. that's out there. We can do that. Scrape that data, paste it through something like GPT, Gemini, or some other uh reasoning agent and say, "What are the top 10 performing

creatives that are currently out there?" Assess those actions and figure out what's going on. Split that even further. What's working well this week? What are the Reddit posts that have come out on the sentiment here? What do we think is hot? What's hot? Right? And you come and you say, "What's really hot is jelly beans. Jelly beans are really hot right now this week. That's dumb, but let's do it, guys. Let's make it. Let's scrape that information. And then let's go ahead and modify our wellperforming videos using Clingo1 video edit edit to go ahead and modify a known good video with a new subject. Yeah, we used to blow up the primary game logo and it turned into confetti. Confetti's out. It's jelly beans. So, we're going to go ahead and we're going to replace all the confetti with jelly beans using Clingo 01 video edit. We're doing it. That's got to happen right there. Then, you've got to have this workflow maturity. We're still early here. Nodevember may have been popular, but we're still building out a lot of the things that I would consider table stakes for true automation scripts and true automation here. Four loops, do end nodes, metadata for image generations, you uh easy and

seamless UA platform integration. These are things that we're building out right there. But people are realizing that many to many video generation systems and asset generation systems do need things like how do I assess a certain position of an array of an element and then make sure that that array element is matched for other elements. So I want to have a bunch of these different videos and I want to modify array position zero for this video and then I want to modify array position three and stitch that together as part of that and then I want to prompted it. There's a lot of these utility nodes for things like workflows or nodebased systems that we ought to build. There's also a lot of personalization that has to be brought in. You all know a lot about audiences, right? You know a lot about who your customers are. How do we pull CRM data? How do we pull audience data into this system so that we can actually have real- time ad creative variance and tailoring based on your audiences? And then lastly, how do we unify generations? Right now, audio, video, and images are related, but they are not cleanly integrated. So, you want to have a talk over over all your videos. That

is an annoying thing to do. We've got to smooth that process out and make sure that we can smoothly blend all three core modalities of video, audio, and images into a single solution. So, if we solve the iron triangle, what could our future actually look like here, right? What could we be doing here as part of our UA uh brand here? We're all here saying, "Hey, what how does that affect me? How do I has it helped me at my job? Well, we could actually be looking, like we mentioned, at this trend to creative pipeline. This is what 2026 could look like. You might be saying that sounds pretty bananas, but it's not that far away. We think it might only be a few months away that we have agents mining last week's trends. Generate ads automatically. Then integrate AI agents with MCPs and skill protocols to go ahead and kick off new ad runs. Those new ad runs come in and those new ad runs get generated and automatically replace your lowest

performing creative set on your UA platform. Those creative sets are automatically updated with the content that you have generated, ended, replaced, rerun automatically via APIs in your UA platform of your choice. what a world we would live in where all the robots take that away for us. But it's not that far away. It sounded crazy and many of the things that we were doing last year felt like magic and I think that we will only accelerate that trend this year. Uh I don't think that this hypothesis and this solution is more than a few months away. I would bet that most other AI uh systems are working on this as are we of course we are working on our a uh AI agents that will allow you to do that. And I also think that the number one concern that that follows up which is oh god if we let the robots run everything how long is it going to be before we get things

that either run a foul of our brand guidelines or run a foul of human decency? Uh well that's where we start building in things like QA agents. We're not that far away with reasoning models to assess based on system prompts and system information whether our outputs actually match your expectations, whether you as a brand owner, an IP manager can actually go in and say, "No, no, no. There's things that are red lines. Do not cross this red line. And if you do, reject that ad and regenerate it." That's not that far away either. In fact, I would say it's probably only a few months away as well. So, as crazy as it would sound to go all the way back here, we're going to go all the way back to the beginning. As crazy as it sounds as a product person to say, we could actually possibly solve the iron triangle for UA where we create good, cheap, and fast content at scale for you and your ad platforms. And I say we because I say we intentionally meaning

someone here is going to do it. We're going to be here at PGC London next year and we're going to be talking as if this is fact because I would not have believed early, let's say last year, I would not have believed that we were in a position where we could do things like generate style consistent assets, extract them easily for our games, create spine animation components, animate it automatically with V3, have those V3 images be faster and cheaper than sketches, have them be generated at scale, have those scales then be built in through CRM. M platforms and delivered to my customers, have it take the amount of time and for the amount of cost it takes for me to go down to Starbucks and get a coffee and a pastry and do that in 2025 felt like magic, but I think we're going to be there this year. It's a scary new world, but it is the world that we're in. Uh, and I will now do the final thing. Obviously, layer.ai sponsoring some of the AI track here at PGC Connect. Come talk to us if

you want to learn more. We're here at PGC London. come and say hello. Uh, scan the QR code. You can come talk to us afterwards. We'll be outside. Uh, right over there. I have my wonderful colleagues here, Borju, Vulcan, uh, Raquel, and Ilia all here to chat. We've got a good chunk of the team here. We're here to help guide you through this weird new world that we are in. Um, and we are trusted by a wide variety of enterprise uh, customers as well as indie customers who come in here across the mobile game space. We're here to solve your problems for you uh, and guide us through this brave new world. 15 seconds left. BOOM. [applause and cheering] THAT'S IT.