In 5 Years, 90% of What You Use AI For Will Run on Your Smartphone | Paolo Ardoino, Tether

If in 2026, people can already run local models on their smartphones to solve only 50% of the use cases of AI. In 5 years, 95% of the population are able to run 90% of the use cases of AI on a smartphone.

Look, Tether was a company born in 2014. Simple idea, digital dollar. There are plenty of digital dollars, but only USD our digital dollar was able to achieve the holy grail of distribution and financial inclusion and impact in the world. Over the last 12 years, our company built the most used digital dollar in the world with 573 million users growing by 30 plus million users per quarter. I call it the biggest financial inclusion success story in the history of humanity. Sometimes my blood boils because how is possible that a little company like Tether was able to achieve more for financial inclusion than all NGOs and charities and whatnot for the last 50 years. And we did it in a very simple way. We used new technologies like blockchain to make the dollar accessible. And you asked me before, well, you know, why not just use the regular dollar? Well, the reality is that for 4 to 5 billion people, so more than half of the population of the world, they cannot have access to that regular dollar. They don't have access to basic financial services. The people that are unbanked, the number of people that are unbanked in the world is just enormous. And these are not bad people. They're unbanked because they're too poor for being of interest to the banking system. They live in countries where their national currency is devaluating so fast against the US dollar. Think about Argentina. The Argentinian peso has lost 94.5% of its value against the US dollar in the last 5 years. The Turkish lira has lost 81% of its value against the US dollar in the last 5 years. The Venezuelan bolivar lost 99.8% of its value against the US dollar. I could go on.

We exactly can. I mean, there are people in Central South America, they have a stable connection. Think about Argentina or Turkey, the digital penetration is incredible. Even Nigeria. Nigeria has 300 million people. Or Philippines. It's huge countries, huge populations. They all have a smartphone. So, they all have digital wallets. So, the reality actually is that the digital penetration for these new technologies and these digital dollars is much lower in Europe and United States. Why? Because they don't need it. But actually, if you talk about remote villages in Africa, well, we can talk a lot about it. We do some incredible things to bring connectivity and energy and access to the dollar there. We could spend hours and it would be the most exciting thing for me in the world. But the reality is that many of these populations have, you know, one of the common denominators of the emerging markets is that they have more youth. The youth has digital means or understands technology faster, of course, than the older generations and they are more prone to change and also have access to smartphones. Like, the average cost of a smartphone for Africa is $80. Of course, you cannot play the best games in the world, but they have browsing, they have like Opera that is a public company is one of the most known browsers in the world has 100 million users in Africa alone, right? So, we partner with them by the way. As one of the main partners, but bottom line is what we learned with USDT over the last 12 years is that we can truly disintermediate. So, if finance didn't perform well in certain parts of the world, and so many hundreds of millions of people were left behind, it was for multiple reasons because there were not enough incentives to solve that problem. And we proved with Tether that we could solve that problem or help to solve that problem. We are still early, but still we are doing a good job. And be very profitable doing so. And we are not taking money from anyone. Actually, well, we are taking money we're earning on the interest rate of the United States. That's it. We are not charging any transaction fee. Sending dollars for the poorest people in the world cost zero in terms of transaction fees. We make the money again on the interest rates of the dollars we keep in the bank account. So, USDT for us and also for us internally was the demonstration that in the world of finance there are too many intermediaries. They all need to make money, and they all need to make money on transaction fees or milking the last mile. And with USDT we proved the opposite. We proved that we could be very successful changing the equation and removing all intermediaries and creating a free product basically. Like just a dollar, simple as that. And we could find a way to be profitable that no one else in the world figured out before us. With the lowest risk possible because, you know, again, we hold the treasury bills in the bank account. So, simple as that.

But, the most important learning point of that is that not only finance is heavily intermediated, is overly intermediated. The entire technology is overly intermediated. You know, internet, I feel internet was born to be peer-to-peer, point-to-point. When you connect your computer to the internet or when your computer or you connect your smartphone to the internet, you get an IP address. And that IP address is basically your home address. And it is kind of unique to you. Could be unique to you. That was the original promise of internet. Let's connect people. And suddenly, internet changed and instead of let's connect people directly, became let's route their connections through data centers. Why? Because also in that case, the intermediaries created a gravitational force so that being intermediaries of people's connections, i.e., people's data, they could earn money hosting, holding, and intermediating people's data. And that is a very huge problem in a growing society. For many reasons, of course, you have data privacy and all the things we know, but there are some issues that not many think about. If you have a person living in Rome, like imagine you are a person living in Rome and it's very likely that your family lives in Rome. And you send a message, I send a message and let's say I'm living in Rome, my family is in Rome, my mom lives in Rome. I send a message on WhatsApp to my mom. That message goes from Rome to Frankfurt or Ireland and back to Rome. Is that good? Is that normal? That adds enormous amount of latency. But even so, that is just a message. Every photo, every video, everything that I send goes to Ireland from Rome, Ireland, Rome. Imagine how much all the governments spent in internet infrastructure in the last 20 years to route packets of data that is completely unnecessary when internet was built to be point-to-point and peer-to-peer. So, that thing is an unnecessary intermediation that is only justified by the broken incentives of internet. So, what we realized at Tether is that we could, but there was already a solution also when it comes to data sharing, how we could build a protocol over internet that could scale to tens of millions of people because of course the classical answer is, oh, but how we can make internet usable if there are no data centers? Well, you know, in early 2000 a new protocol was born, was called BitTorrent and was the iteration of all the file sharing protocols. And was so great that could scale to hundreds of millions of users and exabyte and exabyte of data in the perfect way. And so technically, we had already our answer. We could build, we could reuse structures similar to the BitTorrent protocol and to disintermediate servers and data centers in many applications. That's also part of the things that we are doing at Theta. Happy to elaborate more. But this is basically the spirit that we have in us. Like we learn the concept of disintermediation from finance and what we did to the financial world. And by the way, now everyone talks about stable coins. Everyone, even banks now are talking about stable coins because they understand the power. And so now that everyone understands the power of stable coins, we looked around and we thought, what else we can disrupt. And that's why we built Hole Punch as a protocol and then Qwak as an AI the same concept applied to AI.

Sure. So, I'm a big sci-fi fan and just, you know, for the listeners if they for any reason they would they want to read tonight an interesting short story, it's 14 pages. I think it's the most beautiful short story ever written. Was written by Isaac Asimov 1956. It's called The Last Question. To me, it's beautiful because it's a story of science, AI, physics, religion, universe, philosophy, all coming together in 14 pages. And it's a story of humanity every hundreds of years and dozens of years and millions of years. Always perfecting this automatic computer. Make it better, more powerful. And every single time humanity was asking one single question to this computer, to this AI. It was the last question. The most complex is the most complex question that you can ask and trying to get an answer for. That question was how entropy may be reversed. So, how we can stop the universe from dying. And to me, that is very intriguing, is very exciting, but you know, bringing back that huge complex question to Earth means how we can stop society from dying. How we can stop Earth from dying. How we can make society stable. And society can be made stable going back to and connecting to my previous part is through technology and through connecting society directly without intermediaries. And having, you know, society was built for the last 5,000 years. Society was peer-to-peer. Has always been peer-to-peer. People were meeting in the streets, were talking. People were using cash or salt or coins to interact peer-to-peer. Only in the last 50 years, information and money became fully intermediated. So, we built an entire society over the last 5,000 years that was built with a different construct. And we change it completely in the last 50 years. So, we don't know how society will grow in which direction, how would the decisions of intermediation that we made in the last 50 years, how they will affect society in the long run. So, when I go back to the sci-fi story, I think how we can truly learn from that and if we want to achieve an AI that is even able to answer the most complex question of the universe, that should be an AI that is part of the fabric of the universe itself. It's almost like a new element of the periodic table. An AI that is so intelligent that knows everything that can scale to and be distributed to the four corners of the universe, cannot sit in a data center on Earth. Cannot belong to one person. Even just an AI, how we can plan to have an AI on Mars if the data center is on Earth? The latency between Mars and Earth is too long. And then we add another planet and another planet. Of course, we are years away from that or decades away from that, but still I think we should design technology to scale with the scale of humanity and with our ambition in terms of expanding ourselves in other universes and so on. If we don't do that, we are just here for quick gains. We are not here to build a product for the safety and stability of humanity. We are just here to build a product that makes couple companies rich. And I became very bearish with data centers. Also for multiple reasons. You have, you know, there are some beautiful stories. There was this guy in Australia, this entrepreneur that was able to try to find the cure of the cancer of his dog using ChatGPT and AlphaFold. It's a beautiful story, it's a great thing, but the reality is that 95 to 99% of the people in the world, they will never do that. They use AI for basic things, for search or to take a photo of a grocery list or the grocery receipt and have at the end of the month some accounting. They want some translations, they want some education, they want things that a model, an AI model that runs on a smartphone or on a cheap laptop can already do today. So, if in 2026, people can already run local models on their smartphones to solve only 50% of the use cases of AI, all the use cases the normal use cases of AI, the use cases of AI that everyone all the normal people would run. In 3 years, I'm sure we can bring that percentage to 70%. So, in 3 years, 70% of the use cases of AI can run on laptops and smartphones. And in 5 years, it will go to 90%. So, if in 5 years, 95% of the population are able to run 90% of the use cases of AI on a smartphone, then why the hell we are building tens of gigawatts of data centers when you know, and the subscriptions to these AI models in centralized data centers cannot be justified because then the subscription could cost millions, not hundreds of dollars. And maybe it's still great. I mean, a smartphone will never be able to find the cure of the cancer of a dog. And that is extremely important. I'm not saying that this is not important. Those I think medicine will have huge breakthroughs thanks to AI. But it will become a niche. What will run on centralized data centers on these behemoths will be a niche, will cost billions and billions of dollars, and will be probably subsidized by governments because that is of fundamental importance. But all the normal people, the hundreds of millions of people, the billions of people will use AI as part of their day-to-day lives through their smartphones maintaining privacy or their own data. But look at Apple. Every year they're releasing the M3, the M4, M5 GPU of their iPhone 16, 17, and now 18. And every year the these GPUs can run these llama models two times faster than the previous iteration. So again, that's why in five years what we run on the smartphones will be so good that we will forget about us paying a subscription to OpenAI.

I, you know, coming from this is a very interesting question, right? Coming from the Bitcoin world, you might know that Bitcoin was first Bitcoin mining was first running on CPUs, then moved to GPUs, and then moved to ASICs. AI started from CPUs, now is on GPUs, and last year, the first ASICs for AI were born. So, that right now, if you have a good GPU, you can run Llama 2.2 at 150 tokens per second. The ASICs for AI were able to run Llama 2.2 at 17,000 tokens per second, right? So, I think there is a world where even if you are a bank, let's say a big bank, you could buy in 5 years, 10, 15 powerful ASICs, spend $50,000, $100,000, but you can run an AI cluster that is extremely powerful through the ASICs. So, the ASICs will bring efficiency or energy consumption down to 98%. So, you can run much faster models with much less energy directly on site. And with the added benefit of the fact that you can keep your secret sauce for you, because you probably, you know, the more the time will pass, the more companies will realize that someone else is training their model on someone else's data, right? So, I think over time, of course, now only centralized data center had the capacity to create very cool stuff. And over time, I think that, you know, many banks they run their own data centers, right? So, and the more the models will become good, the more even on the enterprise, I think that they will have their own dedicated hardware. And so, they will be also because think about it, like you don't want every Italian bank, sorry to make all references to Italy, but I'm Italian by birth, will all Italian banks run on a US-based infrastructure, right? Probably not. And so, will all the Italian public administration run on a US-based infrastructure? Probably not. Right? So, and the same thing is like all French, so there is already a big push. If you read one of the coolest new product lines for data centers or cloud providers is actually the sovereign cloud. Means that you go into a country, you install, you build a small data center directly for the country, deploy there your software. And so, even for AI providers, they anyway, they will need to install certain capacity to run models there. So, I don't think we are going to have to see this huge concentration for too mature a long time.

In order to run on smartphones, we need, of course, to ensure we, in order to have a platform that is able to scale and evolve on smartphones, you want to do two things. The first one is inference. The process of inference is you know, you ask something to the model, you get a reply. And so, that computation is called inference. The second part that is very important is how the model can learn from you. Imagine this. Imagine if you could have a model on your laptop, on your smartphone, that could read privately through all your emails and design and learn how to respond in the best way. So, you want to do that, you might want to do that, but you also don't want to do that if it entails to send all your emails to someone else to train the model. So, we wanted to make sure that if I want to have Paolo's assistant that responds in my own way, you know, making, you know, starting the emails in a certain way as I do, and so on. I needed to be able to do that running directly on my laptop and learning directly from me, and maintaining 100% of my privacy. Same with documents, same with how I started the company documents, my strategy company documents, everything that I do. And I want something that is completely customizable and adaptable to me. That process is called fine-tuning. And usually the best way, best technique to do fine-tuning is LoRA, that is low-rank adaptation. So, basically, you take an existing model, and you adjust only a portion of the weights base, and you fine-tune them in order to be respectful and slightly adjust those weights so that they could respect it customized based on your own behavior. And so, we understood that, of course, inference was already a proven task to do on local devices, but...

With Qvoc, we took the llama.cpp, that is probably the most known inference engine for open-source AI, and we built on top of it. And we optimized so much that now it can scale to basically every single consumer GPU. And on top of that, the other breakthrough that was done by Microsoft was called BitNet. The problem is that BitNet is a one-bit model. So usually models have different levels of quantization, they use either 32-bits or 16-bits or 8-bits for representing the weights. Microsoft came up with this BitNet, so one-bit weight model. The problem there is that it was not suitable to run on consumer devices. So technically it was what would have been the perfect solution to run on consumer devices, but was not suitable because it was too much relying on Nvidia high-end GPUs. And so we modified first BitNet to make it adaptable to run on any consumer device and any consumer GPU like the Snapdragon GPUs that you find on the Samsung phones or the Adreno GPUs or the Apple GPUs. Second, we also created in Qvoc a common fine-tuning framework so that not only for the BitNet models, but any model that we support and we support hundreds of models, we give to the developer the same framework to fine-tune any model on all the consumer GPUs. So you have your Llama 3.2 model, you have the MedGemma, you have a Qwen, you have all these models that are supported by our Qvoc platform, now you can run fine-tuning directly on your own device. So I can have my Paolo's assistant, you can have your Craig's assistant, and that will continue to maintain 100% of the privacy.

I must say that I'm very impressed by our team. I rate them that they are doing incredibly well. I mean, and the thing is that we work, you know, the beauty of open source. I come from a huge appreciation for people like Linus Torvalds or Richard Stallman. They are the fathers of the open source and the beauty of working with highest levels of transparency. So when we get out with a claim, it's for everyone up there in open source. Everyone can check the code. Everyone can challenge us. That's I think is the most beautiful thing. I believe, you know, the crypto industry, the Bitcoin industry came up with this motto, "Not your keys, not your coins." So if you don't hold the private keys to your Bitcoin, those are not really your Bitcoins. Well, I would say, "Not your AI, not your intelligence." So if you don't understand how your AI works, it's not you becoming more intelligent. It's someone else becoming more intelligent with your data. And so we are very far along. Now, Qvoc supports OCR, so basically recognition of images and text in the images. Supports text-to-speech, speech-to-text. Supports standard AI chats, like what you are used to with ChatGPT. Supports medical models. Supports financial models. Supports so many different models. Plus has the ability to also interact with other devices at the same time. Supports delegated inference, so from your smartphone, you can use your laptop to run the heaviest of the tasks. It is becoming a very complete product. And we support image generation, video generation. So there's so much that the team on a weekly basis is rolling out. So I'm very excited. On top of that, we are going to roll out very soon our AI assistant. It will be again fully open source for everyone to actually own their own intelligence.

First of all, we want the Qvoc platform is made so that it is built as a software development kit so that every developer. So I want to empower every developer to integrate these AI tools directly into their existing applications. So if you have, let's say that you're building a new version of Uber or you're building a mapping application or you're building a financial application, a training application, or whatever you want. Or like a cooking application or whatever. You should be able to with Qvoc, you take you integrate Qvoc within your application. It gives you already all the primitives to support text-based AI like chat mode. You can or like translation, transcription, voice recognition, voice modification, everything you want. So we created basically the LEGO for AI. I like to describe it in that way. I'm a big fan of LEGO products. But it's really the LEGO blocks so that whatever application you're building, you can take the LEGO block that is needed for you and integrate in your existing application. So that truly, back to the sci-fi story that I like so much, AI can be part of the fabric of the universe. Then, in order to showcase how exactly we work and how we wanted to build our own consumer application. It's called Qvoc AI Assistant. We'll be releasing it in the next couple of months and it will be open source as well and will show how all these different tools and LEGO blocks can come together. Almost like, you know, I was making the wrong analogy with the Death Star because I just finished to build that on the LEGO side. So I don't want to build a Death Star of course with AI, but you know, you have all the schemas from LEGO and they tell you where to put each block, right? So we want to do that to showcase exactly what in our opinion is the best outcome you can reach, but then we give you all the different blocks. So if you want you can take and build something else and it's all completely up to you.

Exactly that. It will be an app that will work on Linux, Windows, Mac, iOS, Android, you name it. And on servers it can work everywhere. And again, it will be fully open source.

We have been collaborating with and we are in direct collaborations with many companies that are building on this technology already. Keep in mind that we open-sourced it less than 2 months ago and yet there are quite some companies that are building on it and every day more are looking at it. But there are two particularly that are very interesting. You know, the other big part of let's say the tech revolution will come with robots. And that is where also things can become scary. I believe that for multiple reasons, of course, but I think a robot's imagine having a robot that can only take a decision if they are connected to a cloud and to a centralized data center. Of course, there is a privacy issue, there is a control issue, many issues, but on top of that there is a latency issue. If you have like your smart car and it will only brake and stop only because the image that it's seeing is analyzed by a data center, well, that can go wrong in many different ways. And the same thing with a robot. Let's say you have a robot that is helping, I don't know, a child, a kid. You want that robot to understand exactly and react in a microsecond or millisecond. Cannot wait for packets to go to Ireland and back if you are in Italy, right? So even robotics will need to have a GPU or an ASIC in their brain to be able to analyze information immediately locally. So I think it's becoming very obvious at least to us that is going to be the case. And so yeah, we have these two robotics companies that are testing how Qvoc is performing in the brain of the robots they're building. So we are going to hopefully showcase something very soon. It's very exciting.