Is This the Cheapest Way to Use AI in 2026? (65+ Models Tested) | Iskander Malikov, Trioma

What if you didn't have to choose between ChatGPT, Claude, and Gemini, and instead, the platform chose for you? That's the core idea behind Trioma, a vertically integrated AI ecosystem built by Iskander and his team in the UAE. Craig Smith sits down with Iskander Malikov, Co-founder & CEO of Trioma, to unpack how Trioma's orchestration engine routes prompts across 65 models, why token price is a misleading metric, and how their GPU network lets users participate in AI infrastructure build-out to offset their own compute costs. The conversation goes deep on the economics of AI subscriptions, the coming collapse of traditional SaaS, and why Trioma believes AI should be sovereign, accessible, and community-driven, not locked behind a single company's walled garden. If you're thinking about where AI access, cost, and infrastructure are heading, this one is essential listening. Subscribe to Eye on A. I. for weekly conversations with the people building and deploying the future of AI.

• 0:00Speaker 00

  Something that would have taken months if not years of development, Trioma did it in like 10 minutes. It is exciting and scary.

• 0:06Speaker 01

  There are a lot of these companies out right now that are offering multimodal access with a single subscription. Trioma does something a little bit more.

• 0:17Speaker 00

  Most people actually want certain thing to get done. If that's what users want, why not let them tell us what they want by simply entering a prompt or recording their voice and letting us know what they want. And then the orchestration engine will figure out the best way to get there. We live in a world where so many new applications will emerge. Because now everyone have tools to build something truly unique.

• 0:42Speaker 00

  Uh, Trioma is a vertically integrated AI ecosystem. Uh, where there are three separate layers, the top layer, which we call the interface. It includes the web platform, the desktop application, as well as iOS and Android apps that are coming soon. Then there's the mid-layer, uh, where we have 65 different AI models, as well as the orchestration layer that connects them and makes them work together. Uh, and then there's the bottom layer, which is GPU, essentially chips that power that, uh, that compute. Uh, and all together, we, we call it Trioma, uh, as we have, uh, yeah, three layers essentially.

• 1:19Speaker 01

  Hi, the world is getting increasingly complicated. Every time your team deploys a new cloud workload or AI agent, another identity gets permanent access to your critical systems. Did you know nine out of 10 organizations suffered an identity breach last year? Because adversaries are logging in with valid credentials, not breaking in. Non-human identities outnumber human identities 109 to 1, creating a massive unmanaged attack surface in cloud environments. Legacy tools were built to manage human employees, which leaves modern machine and AI access largely unmanaged.

• 2:04Speaker 01

  That's where Idira by Palo Alto Networks comes in. Human, machine, AI, one identity platform for all. 96% of human users operate with excessive permissions they don't actually need. Idira replaces permanent permissions with dynamic access. So you can lock down every identity without slowing down your business. Idira provisions access on demand and revokes it immediately when the job is done, which removes high-risk permanent permissions. With Idira, you get unified control by replacing siloed IAM, identity and access management,

• 2:47Speaker 01

  PAM, modern privilege access management, and governance tools with a single control plane across human, machine, and AI identities. Secure every identity at your organization with Idira by Palo Alto Networks. Learn more at paloalto networks dot com slash I D I R A. There are a lot of these companies out right now that are offering multimodal uh, access with a single subscription.

• 3:25Speaker 01

  Uh, Trioma does something a little bit more rather than, uh, the, the user picking the model. The Trioma routes the model to the appropriate, uh, model depending on the task. Is that right? Can you talk about why this is the new model for, uh, for access?

• 3:51Speaker 01

  Uh, and who is the user?

• 3:53Speaker 00

  I think the, the old model, the, the original behavior was, uh, something similar to a chatbot, where most users were utilizing AI the same way they were sort of utilizing Google search or any other search, they were asking a question and expecting a response. Over time, that response has shifted towards an actual result, whether it's an image, an audio file, uh, a video, like, uh, initially was a brief video. Now we can do like proper long videos, but it's still a question and answer type of, uh, interaction. What we've realized is that most people actually want certain thing to get done. So when they interact with AI, they are mostly interested in the result itself,

• 4:34Speaker 00

  not so much the process, the reasoning, how the AI, uh, is processing your specific prompt. And we thought, okay, if that is what users want, to us, users always come first. So we thought, okay, if that's what users want, why not let them tell us what they want by simply entering a prompt or recording their voice and letting us know what they want. And then the orchestration engine will figure out the best way to get there. And it could be a combination of different AI models. It could be a single model, uh, it could be an orchestration engine again, that utilizes different models to pick the most optimal and efficient model to get the job done. But what we thought about is that they need the end result. And to answer a question like, who is this for?

• 5:19Speaker 00

  Uh, that is something that we've been thinking long and hard about. Like, who are we actually building this for? Like, do we want to go with a specific, uh, audience, let's say, content creators? Or do we want to go with, say, coders? Or do you want to build something for just, you know, average people like stay-at-home moms? For example, to let them, uh, handle their, like, household stuff, automate their, uh, certain things that they do at home. And we thought that we should not be limiting our product to a specific group of people. Instead, we should go wide and we should let everyone use AI. Because one of our core principles is that AI has been trained by using universal knowledge. And therefore, everyone should be able to fully leverage, uh, that knowledge

• 6:03Speaker 00

  to be able to gain access to the intelligence that has been combined into those models. That was, you know, the, the intelligence that was trained to build those models. So that's our, one of our core concepts. So to answer your question, it's for everyone.

• 6:16Speaker 01

  But consumer as opposed to enterprise or are you looking at both?

• 6:22Speaker 00

  Both. We have, we're seeing different use cases. So consumers are mostly interested in things like content creation, coding, uh, they leverage agents. They leverage the agentic workflow where different models work, uh, hand in hand when, when the orchestration kicks in. But the same thing applies to, uh, businesses as well. As many of them actually need agents in their, uh, in their businesses to build websites, to automate certain workflows, to automate certain processes, whether it's customer support, compliance, all sorts of things. So we're seeing demand for both. Most of our business customers, um, are interested in, uh, the agentic workflow where they automate the entire flow. Consumers, on the other hand, they, they do a bit of both, uh, agentic as well as

• 7:05Speaker 00

  sort of question and answer type of, uh, interaction.

• 7:08Speaker 01

  And of the combining, the orchestrator, it's more than routing, right? You're not just deciding, oh, this task, uh, should go to Gemini or this task should go to ChatGPT. Does Trioma break down tasks into subtasks and use an ensemble of models to get the task done?

• 7:33Speaker 00

  Absolutely. And that is, uh, one of the core, uh, functionalities of our engine. Essentially, it understands that, first of all, it analyzes your prompt. So this is, uh, this is actually some of the feedback that we've been receiving from customers. They're saying, oh, like when I use a frontier, uh, AI, uh, model, it's instant, right? So they're saying, oh, like if I'm using AI, it should be instant. Well, our response is like, it could be instant, but if you really wanted to understand deeply what you're trying to do, especially if your prompt is not something that is simple, but something more sophisticated, it takes time to analyze it and to understand which model or like a, which models should be utilized to get the job done. And that includes both orchestration that uses and leverages models that, uh, are picked,

• 8:17Speaker 00

  uh, based on efficiency, but also models, uh, that can get the job done. And what I mean by that is certain models can get, can be trained to perform certain tasks, but they could be censored not to do certain things, especially this. I mean, this, this is oftentimes the case with, uh, some more US-based private closed models. And that is where, uh, orchestration shines because it literally picks the, a different model to work on that particular task where censorship is stopping it from completing it. And then it goes back to the original model, which is more efficient, for example. So everything happens under the hood. The customer doesn't need to worry or the business doesn't need to worry about what's happening. What the customer needs to get is the end result.

• 9:00Speaker 00

  And all of that orchestration is essentially required to deliver the best result and consume the least amount of tokens for the customer or for the business.

• 9:09Speaker 01

  And on the token consumption, uh, this has become a real issue, particularly for enterprise. Uh, because while, uh, you know, these models and particularly the agentic workflows can get a lot of stuff done, they can also burn a lot of tokens. As these foundation models grow, there's a certain amount of cost arbitrage involved in your platform, right? Your, your, uh,

• 9:41Speaker 01

  Able to, obviously, otherwise it wouldn't be a business, charge more than than your spending in tokens. How you do that? I mean, do you do you get a discount from the foundation model companies as as effectively a reseller? Or can you explain how that works? And what does that mean for foundation companies if people start using third-party

• 10:13Speaker 01

  interfaces?

• 10:14Speaker 00

  Yeah, a couple of things. So first of all, we have direct relationships with many of our partners, including Frontier AI Labs, Research Labs, some of the big names in the industry, which I'm sure everyone on your podcast are familiar with. We have enterprise deals with them, meaning that we get a discount, kind of like a bulk order, so to speak, where we are able to bring in volumes and get a discount in return. At the end of the day, these are commercial organizations, right? So they're they're interested in making profit. And and and so are we. And to them, we are a reliable and stable customer. Because we are a business that consumes millions and billions of tokens.

• 10:56Speaker 00

  And I understand that it makes sense to have a long-term contract with us versus having millions of contracts with individual consumers who can decide to just, you know, cancel their subscription. I mean, the churn rate with these SaaS companies is oftentimes insane. So it's it's a different business, right? So it has its ups and downs, but it's a different business. When you're dealing with retail, when you're dealing with consumer, there are a lot of factors you need to keep in mind. And this is coming from me as I was building and scaling several consumer applications from trading, from charting. So I understand how how sometimes difficult a consumer apps can be. We, on the other hand, are a pretty

• 11:39Speaker 00

  clear business to them, to our enterprise partners. And they're interested in having a long-term relationships with us because, in a way, we are purchasing these tokens in bulks through an API integration, and we are the distribution layer for them. And why this is beneficial to both us and them is because we're building the community, we're building the actual user base, and we're distributing these models, but we're also enhancing them in a sense that we combine different models and make sure that customers get the best of everything. So they get the the combinations, like a bit of a cocktail of different AI models that work in tandem, or sometimes three, four, five different models.

• 12:22Speaker 00

  And the end result is always better than what they would otherwise get if they were to use with a work with a single, single company. That's one thing. The other thing, something that we've realized, and we actually thought about this enterprise consumption, and you're absolutely right, companies today are consuming a lot of tokens. And if you remember earlier this year, like, there was this concept of token maxing, and a lot of corporations were actually encouraging their their employees to sort of consume a lot of tokens. I think that's changing now, which is great, because I think a lot of that spend was probably not optimal, so to speak. And we thought about this, and we're like, okay, like, if there is so much demand for AI token,

• 13:04Speaker 00

  how do we make sure that our clients, both individuals and businesses, are set so that they don't have to overspend on tokens? And they can build their agentic workflows and automate their workflows and not think too much to an, oh, like, at some point, if the price of tokens keeps increasing, they will figure out that it's cheaper for me to hire someone versus letting the agent do the work. And we're we're kind of seeing that in some of these, some emerging economies already. So it's actually cheaper to hire a person versus, you know, paying an agent through through through tokens. So what we've built is something that we call TGN, is the Trioma GPU Network,

• 13:44Speaker 00

  which allows our customers, both businesses and consumers, to participate in that GPU buildout. Essentially, they're able to buy a seat. It's a software license that they can participate in. And that software license allows them to generate tests, which is the energy stream. And that energy stream can be swapped for AI credits. So they decide what they want to do over time. They want to participate in that network buildout and potentially profit from that, or if they want to spend it on their AI agents, for example. And we're seeing different results from different cohorts of users. Some of them are like essentially token maxing, and they're mostly spending it on their agentic workflows, and they're swapping it for AI. And some of them are just simply interested in seeing that growth

• 14:28Speaker 00

  in in the AI industry, because they're they're understanding this is the hottest sector on the planet, and they want a piece of it, basically.

• 14:33Speaker 01

  Wow. Explain that, break that down a little bit. Because that went by pretty fast. So you're building out this GPU network. Is that through virtual private clouds? Or or do you have your own data centers? Or how are you doing that?

• 14:53Speaker 00

  Yeah, currently, we're renting GPUs. We work with cloud companies, so we rent them. But we are slowly moving towards ownership. We understand that the margins are better. We're able to do a lot more when we own the actual chips. And one of our core principles is actually decentralization and AI sovereignty. And we can't really be fully sovereign if we don't own the actual chips, the actual GPUs. So we're we're getting there. But to answer your question, the way it works is, essentially, we purchase the GPUs. The GPUs are then being rented to companies, and they're also being utilized internally by Trioma. As we deploy open weight models and we're able to improve our margins,

• 15:36Speaker 00

  and we, as a business, are able to essentially become more efficient. And that margin, that efficiency is then being passed to our customers. So this is how customers benefit, and this is how customers are able to get additional models, and how they're able to participate in that buildout.

• 15:53Speaker 01

  And what's the the meter or the mechanism for passing that along to customers?

• 16:00Speaker 00

  The the energy, the Trioma energy. So this is why we call it energy. It's literally a a metric for demand of compute. And as we're seeing today, like, demand for for AI compute is basically skyrocketing. And a lot of people want to participate because of that. But essentially, yeah, so they, our users, both businesses and individuals, are able to decide what they want to do with that compute. They can swap it for AI credits, or they can keep it. So it's it's up to them to decide how they want to interact with with AI. But we give access to AI ecosystem buildout to individuals. And the beauty of that is that they're also able to see the actual product, how that compute is being utilized.

• 16:44Speaker 00

  So it's not very often where you see a company that can let you participate in the underlying chip that powers everything. And the interface is just the the front-end layer that they interact with. But the actual compute happens at the at the bottom layer, which is, which is GPU.

• 17:01Speaker 01

  Yeah. So if I'm a customer and I want to participate in the build-out, what do I do?

• 17:07Speaker 00

  You just buy TGN, just participate in our TGN infrastructure. So we have a section called TGN, Trioma GPU Network, and anyone can can participate. I think the minimum purchase amount is something like thirty dollars or fifty dollars, I'm not sure. And we support all payment methods, whether it's bank transfer, credit card, stable coins. Yeah, so once people realize the potential and how they are able to participate in this, they get excited there very quickly.

• 17:36Speaker 01

  And so I, first of all, I have my subscription to Trioma. And then I buy in the the Trioma GPU network a certain amount of energy. Is that right?

• 17:52Speaker 00

  Yeah. You buy you you you buy TGN, which is your your slot, basically, you participate in that GPU infrastructure buildout. And that generates energy. So your your GPU generates energy. And then you decide what to do with the energy. You can swap it for AI credits, or you can withdraw it if you no longer want to participate.

• 18:12Speaker 01

  And is the

• 18:15Speaker 01

  you you is is this a way to get a discount? Or is does this actually generate revenue at some level?

• 18:24Speaker 00

  Both. We encourage our customers to swap their energy for AI credits, because obviously, we want them to use AI. And this is our way of making AI more accessible to them. But if they no longer wish to use AI for some reason, they can always withdraw and and not use it on Trioma.

• 18:43Speaker 01

  I see. So it's a discounting method. They're they're not withdrawing revenue from TGN?

• 18:50Speaker 00

  If if they no longer wish to participate, they can withdraw through stable coins. But so far, the behavior that we're observing is most of them are either just buying more because they see this hype around AI, or they're simply using it for AI credits. And we're seeing a lot of heavy usage on our platform, quite a lot of power users who are consuming, as you said, many of these tokens. And to them, TGN is the is a perfect solution, because they pay for these tokens anyways through Frontier AI models or through aggregators. And they realize that on Trioma, they can purchase their slot once, and they no longer have to think about it constantly, as long as the test generation satisfies their their token needs, so to speak.

• 19:34Speaker 01

  When OpenClaw was out, there was a bit of controversy, because it was essentially

• 19:45Speaker 01

  Doing an arbitrage on the anthropic models, I guess. Uh, and anthropic shut that down. What was that about? And is that something you have to worry about?

• 19:56Speaker 00

  I mean, we have to understand that all of these companies, they are commercial organizations, first of all, and they understand that this whole subscription model uh is something that is a little bit unnatural to AI companies. And the reason is very simple. Normally, when we think of software or SaaS products, like software as a service products, we think of a company or an organization that has spent some time, built something and then scales this and releases this to like the max amount of people. The cost of extra user is tiny, like you're basically paying the servers, like you're covering your server and infrastructure costs, but you don't really have extra spend because of like hundreds or thousands or millions of new users.

• 20:41Speaker 00

  With AI, it's different, right? Because every single user generates something, whether it's text or images or content. And every single generation spends actual compute, right? It doesn't just come out of nowhere. There's an actual chip that does the actual processing. So as you scale and as you have more people and they're fully utilizing and maxing their subscriptions, very quickly you realize that you're probably not running a very profitable operation here because your subscriptions may end up giving more tokens than the actual cost of that subscriptions. And as companies realize that and as more and more organizations and uh coders realize that that gap,

• 21:23Speaker 00

  they start arbitraging the uh tokens that they're getting through those subscriptions and the actual cost of that subscription. And this is why we're seeing all of these limits that are being introduced by Frontier AI Labs. And I think very quickly we'll see more limits. Uh, and eventually, in my opinion, we're just going to shift towards AI usage. Uh, simple as that. At Trioma, we we've realized that very, very quickly. And we thought, okay, if someone is using AI on a regular basis and they want extra value, they're able to buy our subscriptions. And and our subscriptions are very, very straightforward. They just give you extra AI credits, basically. As we value these customers and we give them some extra incentives, if they know that they're going to spend at least twenty dollars or a hundred dollars a month.

• 22:08Speaker 00

  But if they are power users and they don't really know how much they end up spending, if it's going to be five hundred dollars or five thousand users, these are API customers. And I think um many of these frontier AI labs are realizing that as well, and they're starting to uh limit those subscriptions. And they have to be careful because they have millions of customers, right? So they can't just disappoint them all at once. So they're introducing these things. And also, like, there are a lot of, there's a lot of competition coming in from from different parts of the world where people are uh realizing that maybe I don't need the best model. Maybe I don't need to follow the latest uh sort of PR stunt or marketing trend and use the the latest model just because it has been promoted to me.

• 22:48Speaker 00

  Perhaps for my particular task, a an average model is just as fine, but it costs like ten times less. So people are getting smart about this. And and I think this is why it's so important to ensure that people get access to all of these models. And they're not tied to a single model because at the end of the day, we should be able to give people the freedom of choosing uh what model they're they're comfortable with.

• 23:10Speaker 01

  I mean, there are companies that offer uh multiple models and offer routing to the appropriate model and and even offer ensembling where they put together, uh, you know, they split the task between different models. It is, is cost becoming the competitive edge as more companies do this?

• 23:40Speaker 00

  I would say cost becomes an edge, but not in terms of uh token price or price per millions of tokens as we usually count it, but uh the cost of the actual result. So you're no longer looking at, oh, like this company charges X amount of dollars for like million tokens. You're looking at, oh, how many iterations it took me to get to a certain result, to get to a result that is uh that I'm satisfied with. And if you realize that uh you have a, let's say, you're you're working on a single platform and uh it it costs you like twenty or thirty dollars to get a uh, say, a document done or an image edited. And uh and it took you like fifteen iterations, whereas on another platform, it took you like you've one shot it basically.

• 24:23Speaker 00

  You you did a one shot with a single prompt and then maybe like had an additional prompt just to fine tune a little bit. And it costs you like ten times less. You realize that the price of token is meaningless if the end result is like ten times cheaper. So I think people are getting very knowledgeable about pricing. Uh, the the the arbitrage opportunities are are starting to disappear. Uh, and and and this is, I think it's it's just normal, like as as AI proliferates and as we see many of these new models emerging, it's just normal that at at some point eventually, um, I think most people will shift towards the result um and rather than picking a specific model based on price or some other parameters.

• 25:06Speaker 01

  What just launched, something just launched uh yesterday, I think.

• 25:12Speaker 00

  We did our actual launch yesterday.

• 25:14Speaker 01

  I see.

• 25:15Speaker 00

  Before that. Yeah. So we were um um building our products since January. Um, and we have a a pretty uh, say, hardworking group of individuals, coders who've been building the product. Uh, and we've been collecting a lot of user feedback from from day one. And uh, uh, yesterday was our official launch, uh, where we're fully like live and we users were able to sort of use Trioma since August, but I think yesterday we we had like a proper launch. And now we're again, like, we have quite a few users, quite a lot of demand coming in, and we're collecting that feedback to see what people are mostly interested in, whether it's content creation or coding uh or API usage for businesses, et cetera.

• 25:59Speaker 00

  So we're letting people decide where Trioma should go. We we really wanted to become a community-driven platform so that people can tell us uh which direction we should be heading.

• 26:09Speaker 01

  And you mentioned access to open weight models. Does that include Chinese models, the GLM and Qwen and all of these other powerful uh Chinese models?

• 26:23Speaker 00

  Yeah, we have uh, we have models based out of China. We have GLM. We have a few other Chinese models. Our job is to pick the most capable and efficient models to get the job done, right? We're not in the business of having three hundred models just for the sake of having them, right? If no one's going to use them, what's the point of having them? Uh, first of all, it's not going to be used in orchestration and it doesn't really give you any benefits other than some bragging rights maybe. But again, that's not what users care about. Users care about results. So we only add models that are likely to be utilized through our orchestration engine. But yes, to answer your question, we don't discriminate. Us being a UAE-based company, uh, we're happy to work with both uh Chinese models, open weight models, as well as US uh models and European models as well.

• 27:08Speaker 01

  And obviously, Trioma is a global product. Uh, are you accessible in all markets or are there some markets that uh don't want you playing?

• 27:22Speaker 00

  We uh do not provide services in certain restricted markets, um mainly because our partners that we work with uh do not provide services in those countries. And we work with quite a few partners and uh and we're talking about the sort of the usual suspects, so to speak. Some of these regions are restricted. We're talking about the North Korea and Iran and maybe a few other countries. And uh, according to our terms and conditions, we're not uh able to provide services in those countries. But other than that, we're trying to be as global uh as possible because in our opinion, uh, AI uh should be used by everyone. We should not be limiting people. We should not be limiting uh their AI, their usage.

• 28:05Speaker 00

  We should, however, give them the tools and warrant them about the the responsibility that they should be taking when they're using those tools. Because these tools are becoming very, very capable today. And those tools in wrong hands uh can be misused or used by bad actors. So we're being very careful about who actually uses this. Uh, and that's a very fine line, right? You you actually want people to benefit, but at the same time, you don't want bad actors to uh misuse your tools. We view it as individual responsibility of the person who uses these tools. But there are certain limitations that we have to follow uh due to certain agreements and partnerships that we have.

• 28:43Speaker 01

  You know, I use perplexity, uh, but it gets very expensive uh to to to use all of their routing and and orchestration, I think.

• 29:01Speaker 01

  That's their max level, which I don't pay for. Do you match all of their capabilities just at a cheaper price? I mean, how do you compete with a behemoth like perplexity?

• 29:19Speaker 00

  When we set the pricing, we set a very reasonable margin to ensure that our customers get the best value. So if you look at the sheer token price, it may be on par, if not lower, than some of these larger companies, because we're in the business of proliferating AI, making sure that AI is usable by consumers, by individuals. And the only way you can you can make that product huge if you keep your prices low. So we're in the business of obviously keeping them as low as possible. At the same time, we also understand that consumers are looking at the actual product and how much it costs them to get a certain job done.

• 30:00Speaker 00

  So if we can optimize and let them use, because people generally don't always know, and there's nothing wrong with that, what model to use for a particular task, because they normally pick a single model and then they start doing some work with it, expecting that this model is, maybe they understand it may not be optimal and ideal for all tasks, but maybe they don't want to switch constantly back and forth. So they're just picking a certain model and they're doing the work. We, on the other hand, understand that different models are charged differently. And if we're able to let our users spend less money, it doesn't really matter that much, whether like our tokens are like 10% or 20% cheaper or more expensive.

• 30:43Speaker 00

  What matters is how many tokens you've actually used at the end of the day and what models have been used. So it's the end result and the total price that matters versus individual token price. And we're being very transparent about this. Our users are literally able to check their dashboard and see how much they've spent from every single model. I think and compare like our token prices to our competitors. They usually don't do that because they're happy with the result they're getting and how much they end up paying for it. But if they want to, they can always compare as we're being very, very transparent about our pricing.

• 31:14Speaker 01

  So one of the advantages is you don't need, as I do, five different subscriptions and manually switch between tabs on your computer or apps. But the other advantage is that whereas there are other companies that are offering these orchestrations of multiple models, it's just cheaper. Or in the end, it's cheaper if you're more efficient in finding the best model for the task. Can you talk about the orchestrator? That seems to be kind of the the heart of the offer.

• 32:00Speaker 01

  How is it built? Presumably it's an LLM or an ensemble of LLMs.

• 32:09Speaker 01

  Yeah, can you talk about that?

• 32:10Speaker 00

  I think our CTO would be the best person to ask this question. All I know is it is constantly being updated as we add new models. And there are certain weights and parameters that are constantly changing. But we have a team of brilliant engineers and they keep sort of fine tuning this engine as we speak. I wish I could tell you more, but basically it's the, it's the,

• 32:37Speaker 00

  so on one hand, you have the the actual engine that utilizes different models and routes based on particular models and tries to pick the most optimal model for a particular task. On the other hand, you also have a sort of an internal translator, so to speak, that understands your prompt and tries to get the job done based on your prompt. And if you remember, in the early days, we had this term called prompt engineering as people were supposed to learn how to talk to AI to get the best possible result. So we thought from day one that this is, this thing should not exist. People should not learn how to talk to AI.

• 33:22Speaker 00

  People should just talk like they normally do. And the whole point of AI is to make their lives easier and simpler and better. So they should talk to AI the way they talk normally. And this is what we've been working on and it's part of our engine, actually. It recognizes and understands normal speech, normal chats. It analyzes those prompts and gets the job done. It's part of that engine, actually. Because if you look at some of our platforms that focus on content creation, they have a whole bunch of presets and a whole bunch of complicated prompts that are insanely difficult to reproduce, like a normal person would never be able to like describe all of that. And here at Trioma, we understand that this is not how normal people interact with AI.

• 34:07Speaker 00

  They just literally tell what they feel. They tell what they want to see at the end. And it's the AI's job to understand that and to get the best possible result. So we've been building that since day one.

• 34:21Speaker 01

  And on the coding side, if if somebody wants to use Trioma to code, you have access to the underlying coding models?

• 34:34Speaker 00

  Yeah, I think on the coding side, we are mostly utilizing our desktop application. Coders love it. It gives them everything they want. It gives them their their space. They're able to use the terminal if they want to. They're able to leverage different models. They're able to use orchestration. So desktop application is something that our coders are very much interested in. Most of the feedback that we collect from our coders is around desktop. They love it. They just want like extra features. We haven't really seen a lot of web platform usage from coders, although like they can, they can use that too.

• 35:14Speaker 00

  But just the capabilities of desktop applications are far superior to web platforms based on the specific privacy settings of your web browser. It is simply not able to access certain folders and applications on your computer, whereas desktop really gives you the like the autonomy and the full power of your local machine where desktop application is installed.

• 35:40Speaker 01

  And so far, I mean, I know it's very early days. Do you see usage clustering in a certain domain? I mean, is it just research or agentic workflows or coding? Do you see any early indications of where users are going?

• 36:02Speaker 00

  We saw a bit of a shift, actually. Initially, when we just launched, we saw most people using AI like they're using Google search, like they're asking questions, et cetera. Then we saw two cohorts of users, mostly content creators. And this is why we've built a studio for them because of the demand that we've had for for those types of tools and products that they wanted to utilize. So we're seeing a lot of demand from content creators and coders. Currently, what is changing is that people are now trying to automate. And this is mostly coming from more like advanced and sophisticated users. As they go through this learning curve, as they sort of mature in their understanding of AI

• 36:47Speaker 00

  and capabilities of AI, they no longer wish to simply ask questions and get the result. They now are trying to automate this. And we're seeing a lot of this agentic workflow picking up. And because of that agentic workflow, there are many of them are now switching towards the desktop application to be able to automate these processes and workflows. So this is the behavior we've been observing lately. Yeah, so it basically started as a chatbot, as a search engine. Then it sort of shifted towards I ask a question, I get a response. And now it's all about agents, basically.

• 37:19Speaker 01

  Yeah, I've been talking to people about the future of SaaS generally because, and I'm not necessarily talking about the big, very complicated SaaS offerings, but with this agentic coding, you can build an application. I mean, I was talking to a company the other day and I wanted an integration with

• 37:50Speaker 01

  HeyGen, you know, the voice generation company. And they said, well, we don't have it yet, but you can build that with your. Uh coding agent. And uh you know, so a lot of work or I've seen people talk about Photoshop, you know, that you can build uh essentially of Photoshop application using coding agents that doesn't have the myriad of options that confuse everybody that is built really for your use case.

• 38:30Speaker 01

  How do you, how do you see that developing now that these models are becoming so capable? And does that touch you guys at all?

• 38:41Speaker 00

  It does, absolutely. First of all, that's bad news for SaaS companies. I think uh as we're, as we're seeing, many of these SaaS companies are struggling uh for, for the exact reason you've just mentioned. If you remember in the early days, uh internet was all about like emails and if you, it was mostly communication. It was a channel to transfer data. Uh sometimes uh at the, in the early days, we thought it's just a more efficient way of transferring emails, for example. Oh, like it's better than fax or like, like you're able to search, you can look, look something up. Then we started filling internet with data and that's where the social media, uh social networks uh came along, et cetera. And then SaaS was really shining in uh like ten,

• 39:25Speaker 00

  fifteen years ago, uh mainly due to uh the capabilities of modern browsers. As HTML five became available as a technology, you're now able all of a sudden to transform a lot of heavy desktop applications into the cloud, into the browser. And that was part of it, right? When we were building and scaling trading, that's exactly what we did. We tried to use the analogy of, you know, uh Microsoft Word and Excel versus Google Docs. Uh even today, like Google Docs is not as powerful in terms of capabilities as your good old desktop application, but it's good enough for, for most people and they're like perfectly fine. But the, the fact that you don't have to install anything and you can easily share your files uh is uh very convenient for a lot of people. And this is why many people are not even using

• 40:09Speaker 00

  these desktop applications any longer. So that was SaaS, right? That was the beauty of SaaS. It was giving the tools to people uh and they were able to do everything from their browser. They didn't have to install anything. It was cross-platform. Doesn't matter if you're on a Mac or a Windows. It was just great. But now you no longer have to do the work. You literally tell AI what you want to do. So the whole concept, the whole notion of SaaS was around giving the tools to people, giving like sophisticated, powerful tools, but they have to do the job, you know, with their own hands. But now you just tell AI to do it for you. So if you look at tools such as Thigma or Photoshop, why would you do it by yourself

• 40:50Speaker 00

  or hire someone when you can just tell your AI to draw it and, and model it and build it for you? Um and, and that impacts many of these companies and including your use case. We live in a pretty weird age, I'd say, where it's easier to vibe code an app than to uh look it up or find it and download it. Um so we're seeing a lot of that um lately. Uh and Trioma is actually benefiting from that, to be completely honest, uh because uh many of our users are experimenting uh with different models, with the orchestration, and they understand that their capabilities are increasing as we add more models, because now they can uh, you know,

• 41:32Speaker 00

  pick and choose or think you just leverage the auto mode, the orchestration mode uh through desktop. And to give you some interesting examples, my co-founder the other day uh vibe coded his own messenger app and it took him like ten minutes. And he was like, oh, like, get on this meeting. I was like, oh, like, are you sending me a Zoom link? I was like, no, like, I'm sending you this invite. I was like, oh, interesting. I thought it's going to be like this clunky little thing that will be buggy as hell, but no, like it had like proper audio, video, webcam, support chats. I was able to create like groups and everything. It really literally looked like a proper application, something that would have taken months, if not years of development. And uh Trioma did it in like

• 42:11Speaker 00

  ten minutes. Um so it is exciting and scary, but also, again, exciting in a sense that we live in a world where so many new applications will emerge, because now everyone have tools to uh, to, to, to build something truly unique. Um development is no longer a scarce resource. What is important now is the actual product that you're building for people, like what are you trying to build? Like, how are you planning to uh to service uh consumers, businesses? So the, the community and, and people is becoming more important than ever, because now you can no longer just focus on building something and expecting that people will come if your product

• 42:53Speaker 00

  is good. No, now you have to focus a lot more on the people. And I think we live in a beautiful age now, because uh we're gonna see a lot of amazing applications.

• 43:03Speaker 01

  And so Trioma, you were saying there are three elements in the, that, that make it tri. Can you go through there? There's uh orchestration, Trioma.

• 43:15Speaker 00

  Yeah. The reason why we called it three is because we have the three elements, which is the TGN, Trioma GPU network, then there's TESS, Trioma Energy Stream, and then there's TIE, Trioma AI Credits. So TGN is the GPU infrastructure to build out, uh so people can participate in that. TESS is the energy stream, what essentially TGN generates, and TIE is AI credits. What we spend on prompts, what we spend on AI credits. Together, they form this loop uh which we call Trioma.

• 43:46Speaker 01

  So where, where do you see

• 43:50Speaker 01

  Trioma going? Where do you see the industry going uh as things are changing so fast? And are you able to keep up with the model releases? So do you have the latest Anthropic models or the latest OpenAI models? How long does it take for you to integrate those into the platform?

• 44:14Speaker 00

  Our deaths are usually pretty quick. As soon as these models are released through an API, um we add them in a matter of days, sometimes hours. Um so I'd say we are pretty quick on that front. Uh what takes some time is optimizing the engine. So it's not like, just like adding the model itself does not take a lot of resources. Making sure that model is properly utilized within that en- engine is what takes time and what takes a lot of development resources and effort. Uh again, as we're focusing on getting the best results for our consumers or businesses, we make sure that new model that we've integrated is properly integrated into that engine is only utilized when it's needed. So the capabilities of that model uh are being leveraged for a

• 44:59Speaker 00

  particular task, which that model is very good at. And what we're observing today is that there is no single model that is good at everything. Uh even some of these models that make bold claims and statements that are like they're multimodal. They have different capabilities in terms of generating content, in terms of uh text, copy creation, et cetera. Uh they still have certain limitations, because at the end of the day, they were trained by people. They were trained by certain intelligence. So there are certain limitations. Uh however, when you combine these models, you multiply that intelligence and that's what we're trying to build. We're literally giving access to people uh and we're, we're, we're letting them use the combined intelligence of all of these models.

• 45:43Speaker 01

  And have you, presumably you have uh done these tests, uh which I've done. Uh I'll pick a model through somebody that provides multiple models. I mean, I'll say perplexity, but there are others. So I'll pick uh, you know, Fable in perplexity and I'll run a prompt. Uh and then I'll go to Anthropic and run the prompt uh through Fable directly. And I get different, and I'm not talking about the probability of, of answers. You get a slightly different answer every time you ask

• 46:25Speaker 00

  Yeah.

• 46:25Speaker 01

  any model. But uh there seemed to be a, a difference in the capability of the model when I was doing it directly and through uh perplexity. Is there, is there any difference uh in running a prompt through a model within the Trioma stack and running it through the foundation model directly?

• 46:52Speaker 00

  Yeah, absolutely. Uh a couple of things. First of all, um when we interact with Frontier AI Labs, research labs, uh we don't always know what they actually utilize uh under the hood, right? So it's, it's a bit of a black box to a lot of uh users, um as they have their own orchestration. And we don't really know what model was used for what, because like we can pick a certain model, but it doesn't mean that that specific model was utilized and it was utilized all the time. We have no idea what was actually used. It's a bit of a black box. And, and secondly, on Trioma, because of the orchestration, uh you may uh select the auto mode on Trioma,

• 47:36Speaker 00

  or maybe you can even select a specific model and also select that same model on the uh, the, the Frontier AI Lab website. And your result could be different, because they have an orchestration as well. So basically, everyone has it. It's just a matter, but, but they would be limited to their own models, but we would be able to offer all kinds of models that we have integrated. So the result is always going to be slightly different based on the prompt, if it's a simple prompt. If you're asking uh Trioma to build a website and the Frontier AI Lab to build a website, the result could be completely different. Um because, again, because of the orchestration. If you turn on the auto mode on Trioma, then the result is going to be way different, because then

• 48:20Speaker 00

  the orchestration truly kicks in and now you're leveraging these different models uh for like audio, video, images, copy, et cetera.

• 48:28Speaker 01

  You guys, and this is a selfish question. Uh you know, I do these interviews uh live and I end up with four video streams and an audio stream, you know, two cameras, four cameras and then audio. And I use a couple of different programs to try and manage that, but it's uh, it's a lot of compute and, you know, to try and uh sync the videos and then edit, switch between camera views and stuff like that. Would Trioma be able to handle something like that? Or is that too specialized?

• 49:13Speaker 00

  No, absolutely. We actually have a real film director on the team. So I would encourage you to check out our studio on the web platform as well as desktop application. I'd be surprised if it would not be able to handle your specific task. It literally has everything you've just described. We're actually able to generate up to 60 minute videos on Trioma, even though the model itself has a limitation of 30 seconds. So yeah, I'd encourage you to give it a try.

• 49:42Speaker 01

  Yeah, okay. So you guys have just launched this. And where do you go next? I mean, this is a massive market. As I said, there are other companies doing similar things, but the market is so huge. Do you pick a geography or pick a particular use case? Or as you're saying at the beginning, decide whether this is a consumer or enterprise application? I mean, where do you go from this, from here?

• 50:20Speaker 00

  Yeah, I think the beauty, the beauty of AI is that it's a truly global thing. So we're not limiting ourselves to a specific geography. We're seeing a lot of usage pretty much everywhere, to be honest. I think US is probably leading the way in terms of consumer apps and content creation. But Asia also is creating a lot of content, especially East Asia. So I'd say this is a truly global platform. But to answer your question, right now, we're focusing on distribution on users. We want Trioma to be accessible by everyone. While doing so, we understand that users are everything. We're building our community. We're making sure our community is happy.

• 51:00Speaker 00

  So our number one goal is to let people know what Trioma is capable of doing and essentially spread the message. But at the same time, fully participate in this global GPU network build out as that is an essential part of our core principles. For us to be fully sovereign and to be able to deploy our open weight models. This is something where we have a strong belief in. We think that AI should be available to everyone, that this is global, universal intelligence that should be available to everyone. And we also think that AI should be sovereign, meaning we should be able to control what our AI models are producing.

• 51:43Speaker 00

  And in order for us to do that, we need to have our own GPU infrastructure. So that is something that we're actively working on to enable our users to produce amazing results and not get limited by certain factors, whether it's a certain model restriction or geographical restriction. So we're trying to proliferate the AI. But also, one thing that is very important is, again, community and education. The capabilities of these models today are superb. Like, these models can do so much. And a lot of people have heard of AI, but they don't really know how AI can benefit them personally. So this is why we're focusing a lot on community, so that people trust other people, especially if it's all within the same community.

• 52:27Speaker 00

  And we want that community to help each other understand the capabilities of AI through education and internal communities and groups and chats. So we're very much focusing on the people, because I think that will essentially define how AI is going to evolve as we grow as a company and both as the community of AI users.

• 52:50Speaker 01

  And when you say community or focus on the people, you're talking about a relationship between Trioma and its users? Or are you talking about building a user community that communicates among itself?

• 53:05Speaker 00

  Both. To us, it's the same thing. The users who use Trioma and users who talk to each other while they use Trioma to us, that it's the same community. So we want them to talk to each other, to share their ideas, share their thoughts, be creative and help each other. But at the same time, we want to educate them, to tell them what's possible today and how AI can benefit them. Because there's a lot of fear today that because of AI, a lot of people are going to lose their jobs. And some of that fear, to be honest, is justified. I mean, AI is actually creating a lot of threats to entry level jobs, right? If we're talking entry level, customer support, compliance, many of like legal,

• 53:48Speaker 00

  like a lot of, many of these jobs are under a lot of stress right now. Especially imagine if someone is just freshly graduated and now is thinking about starting his or her career. And you have AI now, which basically is able to do a lot more because of that universal intelligence and for a much cheaper price. So we think some of that, some of those concerns, some of those fears are justified. At the same time, AI really empowers people, right? It gives competent people access to knowledge and intelligence that they didn't have access to before. And they're now able to become a lot more competent and productive in what they do.

• 54:31Speaker 00

  So in a way, it gives extra leverage to competent people, extra power, so to speak. And they're now able to become a lot more productive. People who are creative and who are building things and creating content, AI gives endless opportunities. And we're observing a lot of this in content creation. I mean, there is so much new content being created. We're actually seeing proper series being done through AI. So I think we're going to see a lot of very interesting content and create it.

• 55:02Speaker 01

  Yeah, it's an exciting time to be alive, absolutely.