The city of San Sebasti\u00e1n, in Spain\u2019s Basque region, is a relaxed surfers\u2019 haven that feels a world removed from any war. Yet atop a pine-forested hill overlooking the city, engineers in a conference room at Multiverse Computing are training their focus on combat of the kind raging at the other end of Europe, in Ukraine. They\u2019re demonstrating one of their latest creations: a small AI model designed to help drones communicate from high above a chaotic battlefield.<\/p>\n
On a laptop, the engineers demonstrate how a drone can pinpoint precisely what is in its sights. Using the ordinary, workhorse computer processors known as CPUs, the device can identify encroaching enemy tanks and soldiers, for example, and zip only that information back<\/a> to military units, using a compressed AI model that\u2019s vastly cheaper and more energy-efficient than the behemoth large language models that power chatbots like ChatGPT. \u201cYou need an AI system that is super-frugal,\u201d says Enrique Lizaso Olmos, Multiverse\u2019s CEO and one of four cofounders, as the program quickly picks out a tank. \u201cThe drones use very, very, very little energy,\u201d he adds, even when monitoring a situation that \u201cis getting more and more complex.\u201d<\/p>\n Multiverse, like its AI models, is currently small\u2014predicted sales this year are a modest $25 million. But it\u2019s on to a big idea. Its work focuses on compressing large language platforms, or LLMs, and creating smaller models, in the belief that most consumers and business customers can do just fine with lower-powered but thoughtfully designed AI that needs less power and fewer chips to run.\u00a0<\/p>\n Some experts question how well compressed AI models can truly perform. But the concept has plenty of believers. Multiverse\u2019s clients include manufacturers, financial-services companies, utilities, and defense contractors, among them big names like <\/a>Bosch, Moody\u2019s, and Bank of Canada. The company recently redesigned the customer service system for Spanish mobile operator Telefonica<\/a>, drastically cutting the cost of the LLM it had been using. Lizaso and his team envision their SLMs\u2014small language models\u2014being used for \u201csmart\u201d appliances, like a refrigerator that can tell owners instantly what food needs replacing.<\/p>\n More recently, Multiverse has begun collaborating with Deloitte<\/a> and Intel<\/a> on running public services in the U.S, including a state Medicaid platform, using its SLMs. \u201cThere are tons and tons of applications where to a user you will not see any big difference,\u201d says Burnie Legette, AI senior solutions architect<\/a> for Intel\u2019s government technologies group. But the savings to taxpayers are potentially huge. \u201cTo run an LLM is very, very expensive,\u201d he says.\u00a0<\/p>\n By focusing on creating super-small, affordable AI, Multiverse is tackling head-on an issue that has become increasingly urgent in Silicon Valley and in corporate C-suites. In the scramble to ramp up AI capabilities, many have begun wondering whether the giant investments AI requires will pay off \u2014or whether the costs that LLMs\u2019 power demands inflict on the environment will outweigh the benefits. (For its potential in addressing the latter issue, Multiverse earned a spot on Fortune\u2019s 2025 Change the World list<\/a>.)<\/p>\n \u201cThere is a big problem with the way we are doing AI,\u201d says Rom\u00e1n Or\u00fas, 42, Multiverse\u2019s chief scientific officer. \u201cIt is fundamentally wrong.\u201d He and Lizaso see an opportunity to get it right, while it\u2019s still early days for the technology.<\/p>\n Quantum computing brought the founders together<\/p>\n As far back as 2023,<\/a> OpenAI CEO Sam Altman predicted that giant AI models would eventually fade, given the dizzying expenditures involved. Nvidia<\/a> CEO Jensen Huang has estimated that a single AI data center could cost $50 billion,<\/a> of which $35 billion alone goes to acquiring the GPU chips, the category that Nvidia dominates. As engineers race to create next-generation AI models capable of reasoning, the ever-increasing tab is becoming more evident, as are the voracious electricity and water needs of AI data centers.<\/p>\n Or\u00fas and Lizaso believe that the AI arms race is foolish. They argue that the great majority of AI users have constrained needs that could be met with small, affordable, less energy-hungry models. In their view, millions are unnecessarily downloading giant LLMs like ChatGPT to perform simple tasks like booking air tickets or solving arithmetic problems.\u00a0<\/p>\n Multiverse\u2019s founders came to AI in a roundabout way. Lizaso, now 62, originally trained as a doctor, then worked in banking. But he found his \u201ctrue passion\u201d as a tech entrepreneur in his mid-50s, after joining a WhatsApp group of Spaniards debating an esoteric question: How financial firms could benefit from quantum computing. The group, whose members came from different generations and professions, eventually published an academic paper<\/a> in 2018, arguing that quantum computers could far more accurately and quickly price derivatives and analyze risks than regular computing.<\/p>\n The paper was, and still is, largely theoretical, since quantum computing hasn\u2019t yet seen wide commercial deployment. Still, the response was immediate. \u201cWe started getting phone calls and realized we were on to something,\u201d recalls Or\u00fas, a quantum physicist and seasoned academic. The University of Toronto\u2019s Creative Destruction Lab<\/a> invited the authors to an accelerator bootcamp in 2019. There, they discovered that VC firms and others had distributed their paper to prospective startups, suggesting they jump on their idea; they nicknamed their work \u201cthe Goldman paper,\u201d since it had caught the attention of Goldman Sachs execs. \u201cWe were famous,\u201d Or\u00fas laughs. The friends quit their jobs, and Multiverse was born.<\/p>\n Six years after its launch, Multiverse now calls its products \u201cquantum inspired\u201d: The team uses quantum-physics algorithms to train regular computers, a combination they say enables faster, smarter operations than traditional programming does. These algorithms enable Multiverse to create SLMs\u2014models that can operate on a few million parameters, rather than the billions found in LLMs.<\/p>\n Multiverse\u2019s core business is compressing open-source LLMs with such extreme shrinkage that most of its versions can run on CPUs, or central processing units, of the kind used in smartphones and regular computers, rather than GPUs, or graphics processing units. Because it works with open-source models, it doesn\u2019t need the LLMs creators\u2019 cooperation to do the shrinking.\u00a0<\/p>\n The company has so far raised $290 million in two funding rounds, for a valuation of over $500 million. It\u2019s hardly a household name, although Lizaso confidently predicts it could grow to the size of Anthropic, which projects $5 billion in revenue<\/a> this year.\u00a0<\/p>\n Last April Multiverse rolled out its \u201cSlim\u201d series of compressed AI models, including versions of three of Meta\u2019s Llama models and one from France\u2019s Mistral AI, using an algorithm Multiverse developed known as CompactifAI<\/a>. The company says its versions increase energy efficiency by 84%, compared to the original, with only a 2% to 3% loss in accuracy, and that they drastically cut compute costs. Its so-called Superfly model compressed an open-source AI model on the <\/a>Hugging Face platform to such a great degree that the whole model could be downloaded onto a phone.<\/p>\n In August, the company launched another product in its \u201cmodel zoo,\u201d called ChickenBrain, a compressed version of Meta\u2019s Llama 3.1 model that includes some reasoning capabilities. Intel\u2019s senior principal Stephen Phillips, a computer engineer, says Intel has chosen to work with Multiverse among others because \u201cits models did not appear to lose accuracy when compressed, as SLMs often do.\u201d<\/p>\n \u2018The energy crisis is coming\u2019<\/p>\n The sense that something is going \u201cwrong,\u201d as Or\u00fas puts it, has been echoed even by some leading AI scientists. One consequence is already clear: The potential environmental cost to the planet. U.S. data centers now use about 4.4% of the country\u2019s electricity supply, and globally, data centers\u2019 electricity consumption will more than double by 2030, according to the International Energy Agency<\/a>. By that date, according to the IEA, America\u2019s data centers will use more electricity than the production of aluminum, steel, chemicals and all other energy-intensive manufacturing combined.<\/p>\n Switching AI applications to small, CPU-based models might stem that trend, according to Multiverse. Lizaso believes tech companies are less concerned about the environment than the costs. But the two issues are converging. \u201cIf green means cheaper, they are fully green,\u201d he says. \u201cThe energy crisis is coming.\u201d<\/p>\n Some experts question Multiverse\u2019s claim that for most people, they are just as good as LLMs running on GPUs.. \u201cThat\u2019s a big statement that no one has proven yet,\u201d says Th\u00e9o Alves Da Costa, AI sustainability head at Ekimetrics, an AI solutions company in Paris. \u201cWhen you use that kind of compression, it is always at the cost of something.\u201d He says he has not found a small language model capable of working in French as well as an LLM, for example, and that his own tests found that models slowed down markedly when switching to CPUs.\u00a0It\u2019s also generally the case that open-source models of the kind that Multiverse compresses don\u2019t perform quite as well as proprietary LLMs.<\/p>\n Multiverse\u2019s argument that compressed models significantly cut energy use might also not hold up over time, because cheaper, more accessible AI models will likely attract billions more users. That conundrum is already playing out. In August, Google AI reported that<\/a> the energy consumed for each prompt on its Gemini AI platform was 33 times smaller than one year before. Nonetheless, power consumption at Google<\/a> data centers more than doubled between<\/a> 2020 and 2024, according to an analysis of Google\u2019s report by news site carboncredits.com<\/a>, because so many more people are now using Google AI.<\/p>\n