24 Billion AI Uses Later: What Canva Learned About the Future of Design

0:00Danny Wu:It's such an exciting era to be honest. Like there's just so much potential. But I think the main thing around it is this model trying to understand designs, but the key ingredient in that is really de-layering and being able to understand as well as creating things like presentations or posters or social media posts, anything you can imagine. That's not just a flat raster image. We don't see us as just doing design generation or just doing image generation or changing some pixels from here to there, but it's really a creative option system where we're trying to bring more and more workflows, more and more possibilities and just the entire journeys of design needs into our platform.

0:41Danny Wu:Nothing in the AI technology side is testing us of forcing us to have outcome. The customisation, personalisation, tailing and really like style training and fine tuning and making that accessible, I think that's a very, very good and also very powerful and what we're really excited about is actually kind of continuously trying to push like the definition of this word design as much as we can and in the world now is actually really like using AI to stretch that even further. Well hello everyone, welcome to the Neuron AI Explained. I'm Cory Knowles here with Grant Harvey, ready to bring you another exciting interview with an AI innovator.

1:24Grant, how are you today, man? Doing good, Corey. I'm doing great. How are you? Doing good. Doing good. Excited about this chat. I think it's going to be an interesting one. Yeah. Today, we're talking to Danny Wu, who's the head of AI products at Canva, based in Sydney, Australia, which is very cool. Our first Aussie on the show. That's right. In this role. Is it? He leads Canva. Oh, actually, maybe Tim Davis of Modular was. I think we've had an Aussie before. I take that back. All right. I'm sorry. That's revoked. But anyway, about Danny. So in his role, he leads Canva's AI strategy and product development for the world's all-in-one visual communication platform, which serves over 230 million monthly users as of mid-2025.

2:05Maybe Danny has more updated numbers for us, but that's pretty cool. and then under his leadership he's transformed Canva from a design tool into what the company calls a creative operating system which I'm excited to dig into more about that. Same same well if you're watching and you haven't yet please take just a moment to subscribe so you don't miss out on any of our other upcoming or past interviews and live streams we've had some cool guests and we have a bunch more booked on the way out so you want to make sure you're always around and don't miss out. And with that, Danny, welcome to the Neuron.

2:38Danny Wu:Thank you so much, Corey and Grant. I'm so glad to be here. Oh, we're sure glad to have you with us too. Yeah. So Danny, I want to start with something light just to introduce folks to you. For people who don't know you, when did you become head of AI product at Canva and what was your background before that? Yeah, so I actually took on this role about, I would say, I think four years ago. It's actually a bit hard to keep track with how fast and how there's always something happening in AI. I think it's been three or four years. So I first joined Canva a decade ago as a front-end engineer before moving on to product management for data science and machine learning teams.

3:21Danny Wu:And we were doing machine learning back then, although they were less around generative AI, but more around things like recommending assets as well as personalizing your experience on Canva. But yeah, I was on this role a few years ago, and we've been building GenQ AI products and features ever since. That's really cool. That's awesome. So you go ahead, Grant. I want you to ask this one instead. I know where you're at. Okay, so I was just playing around right before this call with the Canva app in ChatGPT, and I was low-key blown away by it and just how functional it is. So I want you to help us step people through like what actually happens when someone, you know, adds the Canva app in ChatGPT and, you know, like Canva helped me create a pitch deck or mine was, you know, what did I say?

4:14I said, can you make me a presentation about why dogs rule? And it created not just one, but four full presentations with completely editable slides. Like it was really, really amazing. How, what is happening under the hood when that happens?

4:29Danny Wu:Yeah, absolutely. So, I mean, yeah, you know, a simple word, um, MCP. So we've been building a lot of different AI features, like, um, like our great design generation, the presentation, you just saw things having to play around with it, with a grant. Um, and we expose all of those tools, um, through MCP, um, and build on top of the, build on top of, um, SDKs like ChatGPT offers. So it's really giving the same functionality and the same AI, just being exposed to AI assistance. And so while we're doing that, your chatGPT is talking to our MCP server, it's calling functions and tools like design generation, and it's getting the prompt, what you said in chatGPT, pass it on to our tools, we generate some designs and feed it right back to you.

5:24Wow. Well, in just the last few months, it seems like you've integrated the app into ChatGPT, but also Clawed and Microsoft Copilot. And one of the things I'm wondering about is how does each assistant do things a little differently with regard to design creation? Are they pretty similar or is there a substantial difference between them?

5:48Danny Wu:There are individual differences that come down to each platform and model, but what I think has been really amazing in this AI world is how common and how you can basically build once and deploy everywhere, deploy on all the services, all the assistants. So the vast majority of the code and the technology we have to power each of these assistants is shared, it's just MCP. And so that's not just great for developers like us wanting to get consistent experiences, but it also makes the task of supporting all the different AI assistants on one release cycle much easier and much more feasible. It's not like you're building for three or four different completely open systems.

6:38For years, cyber resilience was treated as a recovery problem. backups, restores, continuity plans, and that mindset alone just isn't enough anymore. Cohesity's modern approach reflects how leaders are thinking about resilience today. It can't live in one system, with one team, or in one moment. It has to span data, environments, operations, and decision making before, during, and after disruption. The reality right now is start. Threats move faster than traditional response models. Data sprawl makes risk harder to see. Expectations are rising around readiness and not just prevention anymore. Resilience today is less about tools and more about confidence, assurance, and the ability to keep operating when it matters most.

7:27If you want to see how leaders are building resilience everywhere today, go check out Cohesity at www.cohesity.com slash resilience everywhere. Right. On that point, I've always wondered this, and I'd love to ask an engineer, like, why is it so hard to, like, let's say, like, when you build an app for one platform, like for Mac, why is it so hard to port it to PC? And could AI help us speed that process up? I've always wondered this.

7:55Danny Wu:I think that's a great question. And actually, if you don't mind, I'd love to go through all the challenges we've been facing over camera on cross-platform support. So, yeah, so I'm not sure if you know, but back in, back maybe six or seven years ago, we used to have, so we had an iPhone team that was working on iOS apps. We had an Android team that was working on Android apps. And fundamentally, when you look at an operating system and when you look at the web, which is what camera was originally built on as an environment. The APIs, the way you interact with, the way you like say call things, like display something, or the way you handle like a user click or a user tap, like those things are fundamentally different.

8:41Danny Wu:There isn't really a standard framework or like a standard protocol. There are cross-platform ones, but generally speaking, for building native apps for each operating system you kind of have to go through this journey of building things um building things building things out for each platform which isn't it doesn't sound that bad in theory i mean a lot of companies a lot of tools a lot of platforms do it but uh if you're trying to when you're launching a lot of features really really rapidly and iteratively and you want all of our users to have the same experience we start getting increasingly lots of feedback about saying an iPhone user wondering why they can't use this feature we just launched the web or if sometimes in other ways around and so we actually went through a journey that took a maybe a slightly painful like um year and a half to go cross-platform and basically rewrite um real app to to focus on like that common set of code and a very small amount of um native platform um adapters for the different operating systems so that this way like our product teams um engineering teams can more or less like mostly just build a feature just once and have it across and then gather release features on the same day on all of our platforms which is great it's definitely something our community as well as our team really appreciated and so that's um that's a bit about um about journey to go cross-platform and that's not being i think what's really cool about the ai world is firstly like we're seeing standards like MCT become adopted not by just a single frontier company but actually being adopted by nearly all of them and this degree, this amount of freedom when it comes to cross compatibility is actually something that I've been in technology for a little bit.

10:35Danny Wu:I think this is actually like a quite a new shift in the world of technology. that's fascinating yeah it makes sense that you would want some kind of a that there has to be a translator somehow that can can disperse that out and and understand those differences you would think so that's that's really awesome you've managed to get a hold on that over time i'll bet it is a bear to to do it manually with multiple teams and and nothing's happening at the same speed probably you know one thing's done you're waiting on the other i guess it yeah i don't know absolutely i think um like one example i'll give is when you're working on some additional photo and text effects when you you said when you have multiple clients like a client is something like um a web platform or android app for example um not only do you have to for something like canva which is a visual editor not only do you have to ensure that um features are available on all the platforms but if um if you use a feature on one platform say on web where it's available first then developers also have to handle how that happens when you open up the same design the same presentation another device what if it's not supported what if um it visually looks different and you're consistent then you have no true state of design so for us like just like you know like something like uh like docs like spreadsheets etc like um it's really important for us to ensure that like the same design displays and renders the same across all platforms uh which uh which basically which definitely made um i would just say definitely made development journey until it went um for a unified unified architecture yeah very cool that that makes total sense and what's amazing about canva in my opinion is that so much of what you build and and design uh is for a lot of free users like anyone can use Canva for free on the web and it's a great resource like so many people got their start in in design through using Canva like so it's cool that you're doing all this really hard intense work and there's a lot of times you know like it's it's it's you're giving it away for free I'm not all of it but some of it it's very very generous well most of it thank Thank you.

12:54Danny Wu:It's quite intentional. So roughly about nine or ten of our users are actually free users. And that's very intentional. Our mission is and has always been to empower everyone to design. And I think a lot of times mission statements, what I've seen, they're just a mission statement. that's just something you save and may not actively, I guess, um, guide a lot of the principles and all the actions you do. But for us, like the empower everyone is a really important aspect. Like ever since the early days, we have thought about, you know, um, using family education about wanting to make sure that students who are say learning graphic design or just, um, or just learning in general, like, um, have availability to amazing, great, um, design tools that enable them to transform their creativity and imaginations.

13:50Danny Wu:We wanted to ensure that nonprofits also had access to this tool. So making sure that we have had a great free and available product that everyone in the world can use, that's been part of our DNA from day one. That's awesome. So the Canva design model is described as the world's first, if that's correct model trained to understand design not just images or just video i assume what is it uh what does it actually understand that typical models don't if you don't mind yes the so we're really excited about the model i think the the i think the main thing around it is it is a model like you said cory this model trying to understand designs but what actually the key ingredient in that is really delaying and being able to to understand as well as create things create things like presentations or posters or social media posts anything you can imagine um that's not just a flat raster image so a flat raster image think of it as like a jpeg or a png for example when you have designs when you have things like text when you have things like elements then you want to have the ability to be able to directly edit them.

15:13Danny Wu:Like, you probably don't want to... I'm sure all of us have maybe, like, SAI to read prompt an image to change some text or maybe move this thing, like, to the left or to the right. And I think, I don't know about you all, but a lot of times I find it's much, much easier to just grab my mouse or grab my finger and just directly drag and drop something from A to B. and this element of direct control that is enabled through the generated designs being essentially deconstructed and actual designs instead of just flat. That's what's unique and what we think is really exciting about it in terms of actually helping people through the end-to-end design journey.

15:57Danny Wu:But if you would like, I'd be also really happy to give a quick demo or show you the modeling action. Let's do it. Yeah, that sounds great. We're all about demos. Let's do it. All right. Awesome. Perfect. So we have the camera homepage and camera AI right here. And here's our lovely prompt box. And you can see a list of, I guess, examples here. Any ideas? Or I might actually, since, I mean, you see my background being a bit dry, so I might actually choose a housewarming example. So I just have my prompt for generating invitation, to invite friends to my cozy and welcoming housewarming, featuring all the middle colors and a picture of a house.

16:41Danny Wu:So you can see we've got a few results. And these are images. These are not just images, but these are actually designs. So, for example, if I go here, I can tap right in to head to the camera editor. And if I want to make some changes to actually personalize it, like, for example, come cozy to our house. I can change all the text boxes. I can edit the images, the background. So for example, if I want to make this, use any of our floating tools or apply a filter, or for example, let's say I want to make this a little bit warmer. You can see that only, I mean, it's a little bit hard to see, but I can perform direct mini relations for the background.

17:28Danny Wu:I can also, of course, easily add something. Like if I just wanted to add a style element. These are just composable designs. So they can be published. They can be edited really easily. And, of course, just downloaded with a click. And that deconstruction is really the core of it. Let's see if I... How is it doing that? Like, how is it... Does it generate any of those assets? Does it just know, hey, I can create... I can use these assets that I understand that I already have? like how does it how does that work? Yeah our design model does generate the underlying assets like the images and the backgrounds you've seen and one of the things that we're working on at Research is expanding the list to more and more more and more elements things like decorations like little stars maybe borders around your design things like that so but so it is a transformer based model that has been trained on trained on designs like our internally created designs as well as um designs from our camera creators program so that it actually understands um the inputs to their ai model isn't just images or it's not just visual tokens like um most um most models are operating in the image space it's a model that has been that has been trained i guess end-to-end um the inputs of designs and so the outputs of designs oh wow that's different so cool it's a little bit this way yeah so you're training it on then like on the design files then is what it's seeing and understanding yes so we have um we've got millions of um professionally created templates created by designers through camera creators and this that has been especially pivotal to um actually giving examples um to ai of okay here is um when you here's a design and here is all the different elements, here is all the different ingredients that is in the design.

19:30Danny Wu:Because we have the list and we have the structure of the design as the input and during the training process, then we can make an AI model that outputs things similar to the input, just like how most AI works. And that's how you get layered designs. Wow. I felt like that was missing. I thought that was missing from AI image models forever. And so I'm glad that you've done it. It very much is. It is. Well, since you brought Canvas AI tools to market, I understand you've seen about 22 billion uses. What do you find people are doing with it that maybe has surprised you? What's getting the most traction?

20:18Danny Wu:we have we have actually seen 24 billion uses of our air 24 which is um go uh pretty pretty incredible number now i'll note that is uh that that 24 billion number is for ai usage across all of our tools not specifically by the design model but some of the things we're actually really impressed by is all the creative use cases we haven't really like imagined or building that so So, of course, presentation is a major use case. It's something we know that a lot of our users ask for. It's something that we ask for, like, that we love internally. But we've actually been surprised at some of the things we were creating, like storybooks, for example.

20:59Danny Wu:Like, we heard of one community member who actually used our design model to create a storybook for her daughter that was 20 pages long. And that was an example we never have, I guess, expected. I think it's kind of one of the good showcases of the power of, I guess, AI, because you're no longer just building things for, like, say, a specific job to be done. Instead, you're moving more towards being a builder of tools and what your customers do with their tools is something that will probably very often surprise you in a good way. yeah we've heard that we've heard that in terms of i think um dan shipper had a really good article about this about how to build with a with an agentic uh lens and one of the things that he said was like trying to make the most general purpose tools possible because you don't know how your users are gonna you know put string them together yeah users will surprise you yeah and other than that of course they've got um social social media like all sorts of um all sorts of content creation and another thing that's been particularly interesting is actually logos and this is something that actually holds a little bit close to my heart so one thing I skipped from my intro is before I became a self-engineer I was actually a freelance graphic designer so I was very often um doing things like logos for like uh maybe like 20 30 dollars or designing like designing a brand kit or a website and um it's a little bit crazy to see how things have turned full circle i actually have a actually have a question about that because this is something i've been thinking about a lot and i would love to get your take because you're you know deep in the design world and building ai products what do you think is the best way for freelancers to use these new capabilities that are available to us to not get replaced by ai but to make themselves irreplaceable by ai that's a that's a great question um i think like to me like It first starts with actually establishing your creative signature and establishing and really like thinking like I'm not thinking about tools for the moment, actually thinking about what you do as a creative, like what you are, what your experiences, what your style is, what is your signature about communicating something.

23:20Danny Wu:Like ultimately, you're trying to turn a mishmash of your ideas as well as say your client's ideas or like say your employer's ideas and turning this unimaginable thing into something that's real and tangible, whether that's a logo or whether that's a piece of artwork. And I think that this first step is something that I sometimes see that's missed or skipped, but it can be done in a lot of ways. Like for example, exploring your past works and actually feeding this into, for example, an AI model and asking it like, you know, what prompt would you give me for these images? And really understanding how AI systems, how they understand, how they describe your work thinking about some of those um characteristics and that in itself like um is is definitely one of the first decks in going from just um just from like hey i would say generic like um looks like ai produced outputs to actually taming ai tools a little bit to create things um to create things like that's um more in your style and be more of your partner um and i think the second thing is really like um relying on thinking about how you can rely on ai as a way to accelerate your productivity and make your work work easier how you can scale yourself um and not not so much focus like um like like like you're not like using ai as a freelance um freelance creative isn't really about packing prompts that is um um that can be one part of it and um and for me like it is actually like a pretty significant part of it but it's ultimately the whole workflow of um starting with nintendo mind um using sometimes using a lot of the advanced capabilities like sketch for design so that you can even like um with a pencil sketch you can nail the layout you can nail the hierarchy and the composition and let ai fill in the details and you might do some edits you might do some refinements that when you start treating AI as um individual tools like paintbrushes you can use on a toolkit and try to try to build um try to build your own repeatable and scalable and high velocity workflows that's where AI really becomes magical and uh and just like a companion for you as a creative very similar to how to how digital artistry came about in the first place as well with dealing with, you know, digital brushes and things along those lines.

25:58There was a lot of resistance there as well, as I recall. And I think it's really interesting. And it leads into something. I'm going to step back for just one second. There's such a big difference when it comes to a model in static image creation versus creating editable assets. And I assume that was probably, I understand that it was trained on such things, but I feel like that had to have been quite the technological challenge to solve for as well for your team.

26:32Danny Wu:Yeah, it definitely hasn't been an easy challenge. And the difficulty over it is one of the main reasons why we have to spend more than a year researching, training, and building this ourselves, instead of just relying on something off the shelf. and part of it is when you look at diffusion models like generally what's used to generate things like images, it's ultimately trained on RGB pixels and these things never have the layer information in the first place. So if you think of a photograph for example, if you take a picture of say my cup of coffee, are you in front of me? Then the information that's behind my cup Like that's fundamentally not present or what information is behind my body over the wall over there, for example.

27:28Danny Wu:That's not part of the input. Rasterized words aren't a thing on there. Yeah. It's not. So it becomes a very challenging task when AI model to impute what is there and think. It's almost like, I guess, training, I guess, a generation model in three-dimensional space instead of two-dimensional space. That's more of a metaphor than like a literal, I guess, terminology that is kind of expanding the dimensionality to the Z axis. Would it ever be worth it for your team to develop a 3D model? Because we see that there's a lot of tools where it's turning from 2D assets to 3D. We've talked to some folks who do overworking on the world model side where you almost can kind of create an entire 3D environment.

28:17What are your thoughts on that?

28:19Danny Wu:I will first say like the pretty and also world on world generation simulation models, the advancements, they're actually pretty crazy. Like I'm blown away almost every week. And as for us, like we do look at a lot of different areas for research, but the main question we try to answer is like circling back to, I guess, like why we build AI in the first place is to really help our users complete the end-to-end design journey and let them go from an idea into a final output that's on brand, that meets the requirements they want, that they're actually ready and happy to publish. And so the question we always ask is, does this capability already exist?

29:03Danny Wu:Like, what are all the existing models, open source models, commercial models, APIs that's out there? And are those sufficient? Are those heading on the right track? And at least when it comes to 3D and World Gen, that's an amazing stream of research and development. So it's not really our top priority list. And that mostly comes down to we're not focused on building all the AI models ourselves intentionally. Our strategy is about... That's smart because there's other companies that are doing that and they're raising billions and billions of dollars to do it. There are other companies raising billions of dollars with a lot of expertise that we can, I mean, the good thing is, you know, we as a platform, we can integrate and we can use like what's out there.

29:50Danny Wu:And so we're not like, I guess, I guess like we're actually really happy. Like we're not ashamed in the slightest about using like on about much of our AI using what's of say open source models or APIs that's available, whether that's or even things like or even things like GPT models or court models to us. Like that just kind of like reinforces and highlights our strategy of trying to integrate and bring AI into one page instead of, you know, like trying to do everything ourselves. Actually, I have a question. I think this will lead nicely into Corey's next question, which is, how do you think about, you know, you have a lot of different surfaces with which you're building.

30:28You know, you have the Canvas surface, you have Affinity, and then you also have now, you know, the ChatGPT app inside of ChatGPT. How do you think about how you're integrating AI in all three of these different services and more potentially? How are you deciding what you do where? Yeah.

30:44Danny Wu:yeah great question so we actually have quite a few more services like little nardo for example as well which um we acquired um we acquired i think in late 2024 2025 so i time is a bit of a bit of a blur it really comes down to like a general principle if we want to we want to meet our users where they are so when it comes to chat gbt when it comes to gemini um and claude like um We know that this is like a part of people's day-to-day workflows. It's certainly a part of like most of our workflows at Canva. And conversations can go in any direction. Like it might be taking a picture of something like asking Chachi to identify an object.

31:31Danny Wu:And maybe a few more messages later, you want to create a little like lost poster or maybe not a lost poster, something more positive. But when we see these conversations evolve, we see the interaction patterns changing. And we just, like, it's a really simple belief. Like, we want to meet our users where they are. We want to make it really, really easy. We want to just make it easier for people in the world to design. And so one key part of this has been, and something back to what we were talking about earlier, like the standards like MCP that actually enable Canva and developers like us to be able to build features and build functionality and have a very viable way of releasing and supporting and deploying that to so many different platforms.

32:21Danny Wu:That has been quite critical and really a game changing and making this possible. With that said, is there any concern that those other platforms like that where you're working and serving up your tools for free, that that's cannibalizing your core paid product at all? I think, I mean, it's a valid concern, but I forgot what this is attributed to, but I think there's a great quote about basically you shouldn't be afraid to cannibalize your own product and um not sure if not to acquire a grand either of you like um no i understand that's just something like we try i mean sometimes we can be a little bit uncomfortable about if we think a little bit ahead like you know what are the both the positive ramifications of it and the negative ramifications of it in terms of saying cannibalization but ultimately like if you look at the experience um now like you know like it doesn't really matter what your AI system of choice is like hundreds of millions, like probably billions of global AI users are.

33:29Danny Wu:Like you can access Canva. And the second part of it really comes down to the whole platform we offer. Like we don't see us as just doing design generation or just doing image generation or changing some pixels like here and from here to there. that it's really and like um as as we've been positioning it's really a creative operating system where we're trying to bring more and more of like the more and more workflows more and more possibilities and just entire journeys of design needs into our platform so they can start anywhere they can start on camera they cannot start on camera it's um that ultimately like it is the value of the entire platform that we offer things like things like camera grow things like content scheduling things like all the things like your brand kit all the functionality that like um that that is on the camera platform that's more than just a single task yeah and i'll add something to to to back that up is that i was playing around with both canva you know ai in the canva platform and canva um with the app and i was able to you know in the app on on chat gpt create exactly what i wanted um you know and and then i could just go to canva and start messing about that so So I think having both surfaces is good.

34:49Like if I'm starting from like, we will often make the header images for our newsletter in Canva we have for years. And full disclosure, we've used Canva. And I would go in there and I would use that as my starting point. But now I could technically go there and say, hey, using these assets, but let's make this, I already have a preview that I can go in and just make some tweaks to. So I do think there's cross-functional benefits to having both. Yeah, absolutely.

35:18Danny Wu:And one more thing I'll add is it's actually been pretty, it's been an emerging but growing source of new user acquisition for us as well. We'll find a lot of users are actually discovering and coming to Canva and signing up for a new account through these AI assistants. So it is almost the new SEO in a way. Like in the older days, you would Google for something like presentation creator or like create a poster. And then you'd probably click on all the blue links like Canva. Now the conversation, quite literally, so I intended has shifted towards where users are discovering like ways of how to do something or completing the task through AI assistance.

36:02Danny Wu:So another benefit that we're actually seeing is that this is actually something to shape, help, and grow as a new use and traffic acquisition source as well. That's interesting. I would love, I mean, I'm sure you're not going to tell us, but I would love some hard numbers on that. But honestly, just because for a little while, I think people have been sleeping on ChatGPT apps. And the reason I say that is because it came out right at the end of the year last year, and I really hadn't used it at all. And over this past weekend, I actually went in and tried to add every app, like see as many apps as I could to test them out.