A Cold Room, a Hot Field: YSU’s Supercomputer and Armenia’s AI Ambitions

2026-02-27 - 11:54

Listen to the AI generated audio article. Your browser does not support the audio element. Machine learning advances through iteration: run, compare, adjust, repeat. But iteration requires scale. When computing power is limited, research continues, but with constraints. Experiments shrink. Baselines simplify. Some questions stop being worth asking. A quiet ceiling forms over ambition. At Yerevan State University (YSU), that ceiling has just been lifted. “For the first time ever, our AI researchers’ wings have spread,” says Hrant Khachatryan, group leader of YSU’s Machine Learning Research team and founding director of YerevaNN. Housed in a dedicated cold room maintained at 16 to 18°C, YSU’s new supercomputer center gives Armenia a university-based facility powered by NVIDIA H100 GPUs, capable of serious AI training and high-speed experimentation. In a field defined by heat, speed and global competition, progress now begins in a carefully cooled room in Yerevan. A Cold Room, a Real Machine, and a Real Budget In January 2026, Yerevan State University inaugurated a new data processing center built around 64 NVIDIA H100 GPUs. The system isn’t “AI” by itself, it’s a coordinated computing machine designed to run massive numbers of calculations in parallel, turning experiments that would be painfully slow on ordinary servers into realistic research workflows. It’s a physical reminder that this isn’t software, but infrastructure. Official reporting ties the project to a sizable public investment: 3.4 billion AMD allocated in 2024 and an additional 1.17 billion AMD in 2025 linked to ensure full operation and more effective use of the center. That second tranche matters. A supercomputer is never just a purchase of GPUs. It requires sustained investment in power, maintenance and operations. What a “Supercomputer” Means in Practice Before this investment, YSU had 16 NVIDIA GPUs. They worked on small models that required a few GPUs at a time. Now, with 64 additional H100s, well interconnected and usable as a single system, the ceiling has lifted. Some things that were possible before can now be done better. And some things that weren’t possible at all are now within reach. Khachatryan explains the “better” part through precision. In molecule generation, a model might be asked to change a feature and hit a target value, for example “3”. Earlier, it could generate a result, but the output was often loose, landing at 3.2 or 2.8. More compute means more iterations, tighter variance, and results that move from “close” to “reliable.” This becomes even more demanding when the model must change several features at once while keeping the molecule coherent. Then there’s the category that didn’t work at all before. When we use ChatGPT for translation we can give it a few example sentences in the style we want, and it follows the pattern. In research, teams need models that can perform a basic task and then adapt to a small set of examples, following a provided pattern. In some of YSU’s projects, earlier neural networks couldn’t handle even the basic level reliably. Scaling them up turned hour-long experiments into multi-day runs. With the new system, they can start from larger models that already perform the core task and focus on what matters next: learning from examples and matching the expected pattern. Khachatryan insists the shift is as much psychological as technical. Expectations have changed. Now that everyone uses AI models, you can’t surprise people with “another model.” You need to offer something more. But before you can add new features, you have to make the existing model work properly in your environment, which often wasn’t possible before. Why YSU? Why Now? Khachatryan’s version of the story begins with collaboration. For about five years, YSU researchers have worked with Armenian researchers from the United States. That relationship was more than academic exchange. It showed what their workflows could look like with serious compute, and what they could not do without it. He places this in a wider regional context. “The Armenian tech sector is closely tied to the American technological sector,” Khachatryan explains. “No other country in our region has such strong research connections.” The implications are important: Armenia hosts branches of American companies, and Armenians in the U.S. run technology firms that shape expertise and increasingly influence scientific work. The connection, he argues, exposes local researchers to cutting-edge standards and creates pressure to build capacity that matches the level of collaboration. He also credits NVIDIA’s Armenian vice president, Rev Lebaredian, with accelerating the process, from understanding the right system to moving procurement forward. In Khachatryan’s view, Lebaredian’s opinion carries particular weight, which helped secure confidence in the scale of government investment. The timing isn’t about hype. It’s about what Armenia can contribute to international collaborations. When compute-heavy work has to happen elsewhere, local teams end up with narrower roles. Local compute doesn’t replace cooperation. It changes what Armenia brings to the table. The Research Agenda Already Waiting for the Compute YSU has tied the system to concrete research directions already underway. In a university-published interview, Vice-Rector for Scientific Affairs Rafayel Barkhudaryan describes two active lines of work. One group applies large language models to molecular design and discovery, as well as neural networks for radio signal processing. The second group focuses on drone control, developing AI models that allow drones to operate from text-based instructions. The same group has also made progress on satellite image analysis. These examples show why compute is increasingly cross-disciplinary. Imagine a chemistry team evaluating thousands of candidate molecules and an ML team building models to predict which candidates look promising. The bottleneck isn’t the idea; it’s running enough experiments to learn what works and prove it reliably. Compute lifts the bottleneck, making collaboration repeatable rather than occasional. “If a scientist in chemistry knows the university has a supercomputer that can help them discover new solutions, they will feel motivated to work on new things,” Khachatryan says. In his view, the same applies to researchers and students in physics, biology, and other fields where compute can unlock new methods. The presence of the infrastructure changes what feels worth attempting. “They will understand that these are no longer just dreams—they now have the resources to make these come true.” Keeping Researchers Closer to Science For Khachatryan, the goal isn’t to pull people out of industry, but to keep those who want a research career from leaving science simply because they lack resources and a viable path. Khachatryan distinguishes between what’s technically possible in Armenia and what’s broadly accessible. In the private sector, researchers who need GPUs can usually access them through cloud compute. He points to Picsart and the NVIDIA Armenia team as examples of groups that have long had this access, and notes that Async has gained it more recently. “But outside these companies, these resources are not available anywhere else in Armenia,” he explains. That gap is why a university-based supercomputer matters: it serves people who want to stay in science and still do serious compute-heavy work. To make that path viable, two things are needed: good computers and good salaries. The new center addresses the first; the second remains harder. Khachatryan, who also founded YerevaNN, says that when they started, government funding was essentially zero. Now, more than half comes from the government. If funding continues to grow, from the state and other sources, more motivated students could stay in research. The demand is already there; the opportunities aren’t. He also argues that the supercomputer can indirectly strengthen international fundraising. Domestically, most available grants have already been accessed, though that could change as Armenia’s science strategy evolves. Internationally, the new capacity can make the team more competitive in European and American consortia. Armenia becomes a more attractive partner when local teams can contribute both researchers and the compute needed to run large experiments. According to Khachatryan, offers have already arrived. A team in the UK has expressed interest in collaborating on DNA model development, which would require assembling a dedicated group on the Armenian side. Khachatryan expects demand for the supercomputer to grow. A queueing mechanism already exists. As demand rises, he says, a prioritization model will follow. “With the demand growing for the computer among the students, new resources will follow,” he says. Through collaborations such as Firebird and Eleveight, some resources will also be allocated specifically to scientists. What This Could Change If the center works as intended, it raises the ceiling on what research groups can attempt, makes cross-disciplinary collaboration more practical, and strengthens Armenia’s role in international partnerships as a place where serious experiments can run. The point isn’t the cold room or the hardware list. The point is repetition at scale: running the same idea enough times, under enough conditions, that results tighten from “close” to “reliable,” and from “promising” to “proven.” Comment Creative Tech Armenian Women in Tech Lead on Their Own Terms Sona Gevorgyan Feb 17, 2026 In Armenia’s tech sector, women founders are redefining leadership while navigating motherhood, bias and limited representation. Through resilience, education and visibility, they are building successful companies, and in the process reshaping what the “ideal founder” looks like. Read more When AI Speaks Armenian։ Podcastle’s New Voice for a Nation Elen Tovmasyan Jan 23, 2026 Armenian finally sounds natural in AI. Elen Tovmasyan explores how Podcastle’s new Armenian text-to-speech technology could transform media, education, accessibility and diaspora engagement, while raising important questions about ethics, voice ownership and what it means for a “small” language to enter global AI systems. Read more Why 2025 Mattered for Armenia’s Tech Sector Ani Toroyan Jan 14, 2026 2025 may be remembered as a turning point for Armenia’s tech ecosystem, a year of long-term foundations taking shape. Ani Toroyan writes that while these developments open new opportunities they also raise expectations and if the past year was about setting direction, then 2026 will be about delivery. Read more AI, Drones, Blockchain and Agriculture Catherine Odom Dec 17, 2025 Young Armenian entrepreneurs are bringing AI, drones, and blockchain into agriculture, developing tools tailored to small-scale farmers. From smart beehives to precision spraying and traceable wool, agritech startups are reshaping rural livelihoods while navigating a small, cautious market. Read more The Quiet Strategy Driving Global Wins for Armenian Startups Ani Toroyan Nov 18, 2025 Armenian startups rarely win abroad through formal “global expansion” plans. Instead, founders reveal the quiet force that actually opens doors, from Netflix to P&G: the informal, trust-based networks of the Armenian diaspora, where shared identity becomes a powerful competitive edge. Read more When Technology Meets Health Mane Mkhitaryan Nov 6, 2025 From AI-powered diagnostics to digital therapy, Armenia’s emerging health-tech scene is redefining how we think about wellness. Mane Mkhitaryan spotlights Armenian-led startups bridging technology and healthcare to make expert care, better sleep, and mental resilience more accessible. Read more