AI Cloud Operator Firebird Launches 70,000-GPU, 300MW 'AI Factory' in Armenia

The global race to build artificial intelligence is increasingly colliding with the physical limits of the electrical grid and the geopolitical friction of export controls. For most emerging markets, these barriers have effectively locked them out of the foundational infrastructure required to train and run frontier AI models. But a new facility in the Caucasus is attempting to rewrite that geography.

On August 8, U.S.- and Armenia-based cloud operator Firebird officially launched the first phase of a massive AI computing facility in Hrazdan, Armenia. Backed by a $500 million public-private partnership, the site represents the largest AI factory in the Commonwealth of Independent States (CIS) region. The launch brings 6,144 Nvidia GPUs online immediately, establishing an operational beachhead for a project with vastly larger ambitions.[1][2][6]

The Hrazdan facility is built on the Nvidia DSX reference architecture, a system that codesigns compute, networking, power, and cooling into a single integrated platform. By utilizing Dell PowerEdge servers and advanced chilled-water cooling systems, the DSX design allows operators to dynamically manage power at the GPU and rack levels. This recovers stranded electrical capacity, enabling the facility to pack up to 40 percent more GPUs into the same megawatt footprint compared to conventional data center designs.[1][4]

That density is critical for Firebird's roadmap. The company plans to scale the Armenian campus to house more than 70,000 Nvidia Blackwell and Vera Rubin GPUs, drawing 300 megawatts of power by the end of 2027. Achieving that target would position the Hrazdan site among the largest AI infrastructure platforms in Europe and the broader region, shifting the center of gravity for high-performance computing away from traditional Western hubs.[2][3][4]

The project has drawn significant industry weight. Nvidia has announced its intention to invest directly in Firebird, following an earlier funding round led by AI cloud provider CoreWeave. The opening ceremony underscored the geopolitical stakes, drawing Armenian Prime Minister Nikol Pashinyan, Nvidia CEO Jensen Huang, and U.S. diplomatic representatives. Pashinyan framed the launch as a direct dividend of a U.S.-Armenia technology cooperation agreement signed a year earlier.[1][2][6]

Firebird's strategic advantage lies in its regulatory navigation. Advanced Nvidia accelerators are subject to strict U.S. export controls designed to prevent the technology from reaching restricted nations. Firebird secured explicit export licensing from the U.S. Commerce Department's Bureau of Industry and Security to deliver tens of thousands of GPUs to Armenia, a regulatory clearance that many competing regional hubs have struggled to reproduce.[4][5]

The capacity is already finding buyers. Perplexity, the AI-powered answer engine, has signed on as one of the facility's first major customers, utilizing the high-density cluster for its inference workloads and digital assistant platforms. Firebird is also reserving a portion of the compute for local Armenian startups, universities, and government research initiatives, ensuring domestic access to the hardware.[2][4]

Armenia is only the first node in a broader expansion strategy. Firebird has secured 125 megawatts of capacity at Data Center Valley in Kazakhstan, which also received U.S. export authorization. The company is targeting a global pipeline of two gigawatts of AI infrastructure across frontier markets by the end of 2028, testing whether the AI boom can truly be decentralized.[3][4][6]

Yet, industry observers note that scaling from an initial launch to a multi-megawatt campus presents immense physical challenges. While the current 15-megawatt deployment is a proven reality, reaching the 300-megawatt target will require massive grid upgrades in a country that no longer possesses a large electricity surplus. The ultimate success of the hub will depend on whether the ambitious roadmap can survive the physical realities of power generation and cooling overhead.[5]