Building Sovereign AI: Yang Li, CEO of Cosine on AI Models, Coding Agents and the UK's Frontier AI Future

0:00Hello and welcome back to another episode of Riding Unicorns. Today we are joined by Yang Li, co-founder and CEO of Cosign. Yang, it's great to have you on. We're really excited to get into it. You guys are building AI company at the forefront of like AI coding and we can't wait to discuss how this all came about and how you're doing at the moment. So without further ado, maybe you could start with just like a little bit into your background of what led you to starting the business? Yeah, so my background academically is in statistics. So I never learned to code to build software, but I learned to code essentially to crunch numbers.

0:39But my own sort of personal life was I was a very big gamer. So I applied a lot of those skills of learning to read code to make early mods and, you know, to essentially build hacks to fight my friends on a lot of FPS games. And so that essentially was what got me started playing with a lot of early versions of language models. It definitely wasn't like large language back then, but I had an early experience with stuff like BERT and GPT-2. But essentially, both my co-founder Ali and I have always kept up in our own way the development here. And, you know, eventually when GPT-3 came out, and this was, I think, seven months before ChatGPT was a product, right?

1:16You know, we got to see this and Ali was actually the first one to spin up a prototype about building something to do with AI enabled coding. We realized the aha moment for us was we realized that it could output JSON. And that was a point where we realized it could bootstrap. You can make it a circular, the inputs are the output and so on and so forth. Well, we feel like this was a light bulb moment that we could actually build something. And of course, intelligence and token windows were very small back then, but it showed us the spark of what it could have been. And so you saw what was happening with AI, but you decided to verticalize and go heavy into the coding model.

1:51So today, what are the use cases? Who are your customers? Sure. So I think we'd like to say we have earned the right to call ourselves an AI model lab. We definitely started our life as post-training and customizing other people's models, whether it's a Chinese open source model, or we were very early sort of partners with OpenAI to even post-train GPT-4. We were part of the post-training alpha, been fortunate enough to deal directly with Sam Altman and a lot of the senior guys there. We started our life customizing models to focus on what our customers wanted, right? For us, because we've always wanted to build AI and enabled coding tools and workflows and make software engineers' lives better, we essentially started doubling down and making models better and navigating complex codebases, undocumented or under-documented codebases, legacy coding languages.

2:45But as of today, into 2026, we are a model lab that trains our own model end-to-end. You know, we pre-train, mid-train, post-train. We generate our own synthetic data. You know, we create our own reinforcement learning environments. And, you know, we are now earned the rights to call ourselves a model lab. I think it's quite different in the market. And to speak very bluntly, you know, neither Ali and I have a deep mind CV. You know, we didn't author the original papers that transformed large language models. And we had to figure it out. What a good thing is now is that because we were forced to bootstrap, we had to really listen to our customers.

3:24We had to find ways where simply capital alone was not a solution. We spoke with Scale AI very early on and the team came back to us and said, well, it's a minimum$3 million commitment. And we didn't have$3 million. We couldn't use human labelers to actually generate a lot of the early data sets. So we had to build data pipelines that were predominantly synthetic database. So we had to build agentic workflows internally to do recursive loops, right? And all of these fashionable names, which a year ago or two years ago didn't have a terminology around it. We were forced to do it. And as a result, we now end-to-end have become capable and we can stand up straight and say very truthfully, we are a frontier model lab training our own model end-to-end.

4:07And we compete in terms of either public benchmarks or private evals or even just workflows inside our customers where we are just state-of-the-art in performance, specifically for our customer needs. And you said who our customers are. Our USP is that we are able to deploy not only the model, but also the product surface areas, whether that's a CLI, a Mac, Windows app, a web app, a headless integration into a tool that you already use. We can deploy all of it end-to-end, fully air-gapped. So on-premises, I'm not talking about a VPC inside a hyperscaler. We can do that at the customer's request.

4:44But at the most extreme, we deploy inside environments or onto hardware that our customer actually owns, like the actual GPUs inside our facilities. That inevitably pushes us towards a direction where there are customers who care about that. No data goes in or out of their environments. So if you use an OpenAI or Anthropic or you use a DeepSeek or a Kimi model that's hosted by other people, so with a cloud provider, you have to send your data in and out in order to get a response back. The organizations that care about that and having that air gap nature are defense, financial services, critical national infrastructure, a lot of public services, insurance, health care.

5:23It's a reason that at Rachel Reeves' event in London Tech Week a couple of weeks ago, we announced that we launched this coalition of, I think it was 15 at a time and now 18 today, publicly traded companies or super large enterprises like BT, BAE, Babcock, Talus, Lloyds, Barclays, NatWest, GSK, like even, you know, sovereign cloud providers like ERA4. Or we launched this coalition where we are training our next model based on their requirements. We are sharing data for certain use cases and training a fully sovereign UK model. So both on We Are UK, we're UK based, you know, it's trained on infrastructure and hardware that's located in the UK, primarily through the Isambar supercomputer that's down in Bristol and owned by the UK government.

6:12and also with data and use cases, the requirements given to us by British companies or British subsidiaries. For example, we work with Leonardo, which is an Italian defense company, but a UK subsidiary. And Yang, tell us what it takes to compete with an open AI or an Anthropic with a fraction of the resource, a fraction of the compute, a fraction of the headcount. How do you do it without just doing post-training? I think having a second mover advantage has its perks. So I think if you rewind three years ago, you needed at least five, six, seven hundred million dollars in order to figure out how do you scrape and compress the Internet?

6:53You know, you didn't know what data sets. A lot of the machine learning techniques were not published because they weren't invented yet. To a certain extent, the pre-training, I guess, recipe that now everyone uses is now much more set. Right. And three years ago, that was a genuine advantage. You needed the capital, you needed large research teams, and you just needed a human capital resource in order to scrape and label that data. Today, in order to get us to do a model where a checkpoint of a model that at pre-training that is very competent, or at least ready to receive post-training, right, to really give you the last 10 % of the performance that you need, that baseline intelligence is a relatively proven recipe.

7:29You also now even have open source data sets. So if you look at NVIDIA open source, the entire pre-training data set that they use in order to get their own open source models to a very good pre-training base or checkpoint. And so I think if I were to do this three years ago, I would definitely struggle. And I think that there's a few examples of ISO at poolside or you look at Artur at Mistral. you needed to raise that money in order to do it then, in order to get you to even at least at the starting line. For us, you know, in 2026, we have the advantage of knowing the recipe, having access to the datasets, and to get it to a baseline knowledge.

8:08Where we really excel, and I think it's the same for every Frontier Lab, right? You see OpenAI and Esprope opening up huge offices in London. And from my understanding, and I don't know for sure, is that the tech team that they're hiring for is almost like very heavily focused on post-training teams. that's where you kind of get the alpha now where you actually the baseline pre-training side is table stakes but the rest of it is how do you make opus feel like opus how do you make glm 5.2 feel like 5.2 that feeling the taste and that last mile of what really wins customers hearts and minds and use cases and repetitive usage that's all in post-training right and so today to get me to the point where pre-training i'm not saying it's easy it's still obviously very difficult and the government giving us access, the UK Sovereign AI Fund giving us access to Isambard has been a huge help.

8:54This wouldn't have been possible without their support. But it's relatively orders of magnitude smaller. We're now no longer talking hundreds of millions. We're now talking tens of millions. Yeah. I mean, we're all hearing Sovereign AI spoken about lots. I want to just ask you what you think Sovereign AI means for our country and for any country going after it. It is interesting because sovereign means something different to everyone, right? On the surface, it feels very straightforward, but actually if you dig deeper, I think it means different to different people. I think for me, there's three defining features of sovereignty that spans across nations and industries.

9:32It is certainty, control, and resilience. I think what sovereignty means to our customers is definitely they want certainty over that no one else can possibly turn off their access to AI, right? I think that in 2026, every large organization feels and believes that AI will play a big role in that organization. How they apply it or how much they are and what speed they deploy is obviously variable, but I don't think there's any doubt to say AI is going to make a huge impact in the coming decade. So if you are going to change structurally how your organization works and functions and operates and grows, you want that resilience and certainty of knowing no one can turn it off.

10:14I think the export ban with Donald Trump, the slowdown and release of the new OpenAI model from the Trump administration, the reporting from Reuters and Bloomberg yesterday that the Chinese are now considering export bans or somewhat equivalent analogous controls over open source models has made a lot of large enterprises worried to say, well, we need to be certain, you know, you would never build the rest of your company on something that could be turned off by someone, particularly with geopolitical risk there. The control side, I think, is also something that's very important when it comes to sovereignty.

10:44It is control in terms of what data goes into training a model, which is why our coalition partners like BT, Babcock, BAE are very keen to help us and make input into what goes in. Because not only do they want to know the lineage of our data, they also want to know that the behavior and the competencies of the final model is being fed into from the very start. I think the CIO and the head of AI at HSBC, who we recently spent a lot of time working with, they essentially said it is analogous to imagine if you could rewind time five years ago and tell Anthropik exactly what your requirements are, then you're at that stage for cosine, right?

11:27You get a model that you have certainty over deployment. You have certainty over who controls it, who can turn it on and off. And now you get control over what data goes in. And then the opposite of what data comes out, right? The ability to deploy fully on-prem in an air-gapped environment on your own hardware means that you can be certain that no data is being sent over an API or any other data connection. And the last piece, I think if people underestimate how important this is, but in large enterprises, you know, a lot of the decision-making is made by the CFO and the procurement team. Everyone has seen Uber CTO publicly saying that they blew their annual budget on AI in a single quarter.

12:03Everyone's beginning to see that token maxing is just not sustainable. It's not a secret, but everyone knows that the big fund from American Labs are heavily subsidizing token cost. I think the number I've seen floating around is up to 1600 to 1. That's the amount ratio that they're subsidizing the token cost for consumers at the moment. And that control over cost is worrying. If for CFOs or head of procurements, your company wants to use as much AI as possible, you already feel like it's expensive, even though they're being subsidized. And on top of that, all of these companies are going public.

12:34So at some point, they're going to have to not make such heavily losses, right? So price increases are on the horizon. So the best way to control cost is actually if you deploy it on your own hardware. Yes, there is, of course, an upfront capital cost, but you amortize that over four to six years. actually at that point the more you use and the higher the utilization rate of those graphics cards the lower your average toss per token right becomes right so because it's a fixed cost and then you actually you're as a cfo you get to control costs but you also get to encourage ai adoption which is a win-win inside a lot of organizations and yeah what is the future for like human engineering how are they going to adapt to using ai writing codes not writing code what's your view on the sort of evolving role of engineers within tech?

13:24I'm not a doomer. I think that actually software engineering demand will increase. Inevitably, as I think you alluded to, what you actually do day to day as an engineer will change quite significantly, either internally at Coastline with our own engineers or our customers who are using our product. What you see is that workload does increase if you measure it by the number of tickets completed. So people are just doing more tickets and doing more tasks, but they often don't feel as stressed in the doing part. There definitely is an increase in a burden in sort of requirement and energy and focus and reviews, right?

14:02And so to say, well, actually, I don't quite know exactly how the AI reached this kind of final output, this PR or MR, depending on what you pull requests, no merge requests, depending on what platform you use. But what you do have to put a lot of burden on is to say, well, how do I go and evaluate? How do I review this to make sure that it is adequate? It is adhering to our policy. So to answer your question directly, I think there's two things. One is definitely much more of a burden in terms of reviewing and making sure it's right. I think it's creating more of an architect and creating systems.

14:31Like how do you write better tests? How do you make sure that your own CI is more automated, right? And so all of these things, I think is much more of a, how do I create systems that enable us to kind of continuously be more productive but not actually at the writing part but much more so on the review side of it and beyond the sort of competition we've heard a lot about export controls and i wonder if that's relevant to you guys obviously we've all listened to what happened around mythos project glass wing fable being pulled and now returned how does the politics of ai interfere and accelerate what your day.

15:12When the Mythos export ban was announced, a lot of the conversations that we had ongoing suddenly accelerated. My phone was just littered with WhatsApp and text messages and emails as soon as that happened, because the hypothetical of, yes, the US could put an export ban, or yes, they could shut off access. You know, opening, I could decide they no longer want to serve the rest of the world. But it's fairly hypothetical, right? Like, to a certain extent, Apple does not worry that their supply chain in China is going to shut off overnight. I guess there's somewhat of an understanding that globalization means that it's good for business and everyone does it.

15:49That hypothetical was a very small risk in people's minds. And whilst I think I probably saw the risk more than most, I still agree that it was a hypothetical. That has turned into a precedence. And not only just once, now twice, with the slow rollout open AI and potentially the Chinese open source models now being at least somewhat slowed or curtailed, right? That has really helped reinforce our USP to our customers and also further beyond the UK jurisdiction. We had people from many countries across the GCC reach out to us in terms of asking, how do we replicate this solution? How do we make sure that we don't have a supply chain risk in our own nation as well?

16:30I think that is the biggest kind of accelerant for us, which is what could have been a long tail, sort of almost a black swan event has actually happened. Yeah, and Yang, before Cosign, you have been involved in companies that weren't AI and weren't around coding. What have you found is transferable from your previous experiences? I think what is pretty consistent across every fast growing startup is hiring. Sorry, James, I wish I had a more unusual contrarian view here. But I do think particularly in early stage, you need to hire fast and well, right? I think that you're really talking about a small group of people that can really move the needle here.

17:08And I think the way that I think about it is early on, you need to hire people that will just have high agency. And it's just a personal framework. I think about if I hired this person, could they do at least 50 % of a good job, if not better? So they may be way more qualified, but at least 50 % better of the job that I think I could do. In which case, then you should hire them if they have the right qualities, if they have to view, obviously, other traits. But fundamentally, you need to have that capacity. And you also have to give candid feedback very quickly. We try, right? Internally, we try to do two rules, again, stolen from other people.

17:45One is that where possible, we do try and practice radical candor, where we give people feedback very immediately. We try and instill people to say most people, at least for most times, are not trying to be mean, right? And if they do, then absolutely everyone should come to me immediately because we have no tolerance for people being mean and people not people who are treating with respect. But there is a fine line between that and being very candid very quickly. And secondly, I think for us, we do practice the Amazon, I guess, policy of disagree but commit. We understand that the vast majority of decisions are very subjective, right, or opinions are very subjective.

18:19We think about, OK, there is a point where we say, well, I disagree with you, but let's commit to it. Let's try it. And, you know, we've got to move forward. Again, that cliche, which is indecision is worse than the wrong decision. And Yang, where are you guys at currently in terms of commercials? And we're seeing the big labs scaling revenues insanely fast. How does that translate to what you're doing in the sovereign AI space and taking a slightly different approach? Yeah, no, I think that for us, we're definitely slower because we don't have a consumer side of the business. We have a product that consumers can access, but we focus on huge defense primes, large multi-country financial institutions.

18:58The sales cycles are just longer. We anticipate to end this year at 10 million ARR. So we're on track. And I think that the good thing about large enterprise sales is also larger ACVs. You do the grind, you get the reward. But yet we have ambitions to say we end 2026 at 10 million ARR. And Yang, I believe there's an interesting story with Masayoshi-san and Sam Altman. Can you tell us a bit about that? Yeah. So last time we met with Sam Altman, this was before OpenAI moved from the mission office to the one next to Chase Center, which is the old Uber office. So you have to buzz in and it's quite secretive and high security.

19:35There's no branding, but it's a very small intercom that you press. And then your name has to be on the list in order to be let in. And you're kind of surrounded by plainclothes security all the time. We hadn't registered, but in front of us was Masayoshi-san and his team. and they were on the intercom trying to say, well, it's Masa and they were spilling out his name and he wasn't on the list and no one was on the list either. And it was just like us awkwardly standing behind them. And then when we got to us, we were like, we just buzzed and we said our names and we got in and we were a bit sheepish in terms of like, okay, like, you know, it was really quite surreal.

20:08And then turned out, obviously, we're a very small part, you know, SoftBank is much more important than we are to open AI. Sam was actually double booked. He was meant to meet with Masa I do it all at the same time. So after a few minutes, Sam Olman came out, took us to the canteen, pulled the chair up, then sat backwards on it and just said, essentially sat down and say, hi, what can I do for you? Well, no, what were you like, right? And so we quickly rattled off all the list of requirements that we needed to post-train their models. And, you know, we wanted to have access to custom reward functions and all the stuff we wanted to request.

20:38And very frankly, he was like, yep, these things are coming in weeks. These things are coming in months and these things are never going to come at all. So like it was very, very efficient. and then he zoomed back into a boardroom to go meet with NASA. And so it was short but sweet. It was very effective, but it was just a surreal moment of standing right behind all these SoftBank people and us just waiting patiently behind them. What's the vision for you guys at this point? Do you foresee buying a lot of compute capacity like the labs? Do you see yourself increasing headcount hugely? What's the future?

21:11What's the next 12, 24 months? We think quite deeply about this internally. I think for us, there are a few different options for an AI lab to scale. You could become an inference company where a lot of anthropic and open AI sit. You could become a Palantir type company or a faculty, which is now part of Accenture. You could be a services type company or solutions based one where you're competing against the GSIs. I think for us, where we would like to carve out new space is to be neither. There's elements that we take from both of them, but we don't want to be either. I have no intention of becoming an inference provider.

21:44I will not need to buy as much compute, right? I will have to scale up hundreds of times of what I'm doing now for sure, but not to the level of OpenAI or Antropic needing, you know, Colossus from Elon Musk because we don't want to become an inference company. We will partner up with other cloud or Neo cloud providers to host our models and they manage a lot of the inference side, but we don't fundamentally want to own or have contracts on that side. And a large part of our companies that we work with own their own GPUs already, in which case, you know, we don't need to provide the inference on our own hardware.

22:15They already have, you know, B300 clusters or H200 clusters, nodes. On the other side, I don't intend to make the majority of my revenue by using huge teams of engineers to build a very esoteric and custom solution for our customers. We provide the model and the product, and we let our customers become much more efficient. And inevitably, if you look at our coalition, we have people like PwC and Deloitte. We have the GSI as partnered up already because they often do a lot of the last mile work already. But I do think that, again, uncontroversially, we will need FDEs, right? We will need people to go in to implement a solution and deploy the model because the fact that it is fully on-prem and air-gapped means that every, whilst the model weights and the docker images are somewhat the same every time, there was custom setups and the infrastructure is slightly different.

23:05So we need a team dedicated to both deploy, figure out the idiosyncrasies inside the environments of each individual customer. And I think some point in the future, we've got some demand already, well, not some, quite a lot of demand from our coalition partners to give them the ability to continuously post-train incrementally themselves. And we are with a few customers already providing them or working towards deploying this. All of the tooling that we have internally to allow them to, without letting us know, again, this goes back to certainty and control, that we have no visibility. We're given the tools and the ability to post-train incrementally the model themselves.

23:40And all of that means that we'll have to have a team that provides forward-deployed engineering support. But it will never be at the scale of Cognizant, for example. Cognizant is wonderful and a business that has a huge place in the economy. But fundamentally, that is not a position that I am trying to get to. So I'm somewhat in the middle. I think that's where we are headed, in my opinion. it. Super interesting, Yang. Thank you. And interesting to compare against different models out there. And you're obviously throwing something that's very important to a lot of big companies as they think about how they use AI in the long term and not just today firing up 1000 anthropic accounts.

24:14It's, you know, they got to think more strategically about their usage over the long term, as it becomes more and more central to their processes. So we're going to move on to our future unicorn prediction. I'm sure you've seen lots of really interesting companies. Are there any that you've come across that you think have the potential to go all the way to being a unicorn? I'm gonna say two. There's an AI video editing app called Veed and they're actually British and I'm just a personal huge fan. Like I create lots of content in my own personal life. I'm a dad of three so inevitably I've got lots of home videos that I kind of stitch together and I post in my family groups and also in the company all of the content that we generate whether it's a short video for a customer demo, or it's our marketing content, or a custom sort of onboarding or training, the company is fully bought into Vita.

25:01It's fantastic. It's super easy, intuitive to use. And I suspect most people on the surface will say, ah, it feels like a wrapper. I actually think there's a lot of innovation behind the scenes to make it so seamless and nice to use. Similar to, I guess, the iPhone analogy. The fact that it feels so natural in extension of your arm is the fact that it did a lot of hard work behind the scenes. I think the next one is a partner that we kind of work closely with, I am very confident in them. It's ERA4, a UK sovereign cloud provider. They're relatively new compared to the hyperscale is relatively small.