Armenia’s AI push focuses on compute sovereignty

Firebird, a U.S.-based AI cloud and infrastructure company with operations in San Francisco and Yerevan, plans a first phase of an AI compute hub near Hrazdan around 2026. Armenian officials describe the phase as including more than 6,000 NVIDIA Blackwell GPUs, about $500 million in investment, 18 megawatts of power capacity and up to 110.6 exaflops of FP4 Tensor compute.

The project centers on imported, high-end NVIDIA infrastructure installed in rack-scale systems rather than local chip fabrication. Server partners include Dell Technologies. The hardware platforms combine dozens of GPUs with Arm-based CPUs in liquid-cooled racks designed for dense model training and inference workloads. Armenian authorities present a larger second phase that could bring total investment to roughly $4 billion and more than 41,000 additional GPUs.

U.S. export approvals and a bilateral memorandum of understanding are part of the project framework. The United States and Armenia signed an AI and Semiconductor Innovation Partnership MOU in August 2025 that covers secure semiconductor supply chains, electronics development and AI commercialization within U.S. export-control rules. Armenian officials report that regulators approved transfers of advanced NVIDIA chips for the hub.

Armenian government officials and industry backers describe the initiative as aimed at local access to large-scale compute. The country has a concentrated tech workforce in software engineering, mathematics and electronic design automation; local operations of global EDA firms employ more than 1,000 people. The government said local startups, universities and public agencies would be able to run model training and research without relying solely on foreign cloud providers.

The government has signed a cooperation agreement with a European AI company to develop AI assistants and government services using open, adaptable models for Armenian language and infrastructure. Officials stated that part of Firebird’s capacity would be reserved for domestic companies and public use while other capacity could be offered to international customers, including U.S.-based firms in the region.

Planned infrastructure carries heavy energy, cooling and networking needs. Reported designs include a 100 MW data center that, at full load, would consume electricity comparable to a city of roughly 120,000 people. Project documents describe closed-loop liquid cooling that requires only periodic water replacement. Armenia’s current power mix includes nuclear, hydro and thermal generation, and the government says it is upgrading fiber-optic links with capacities up to 1 terabit per second to support data flows.

Project documents and officials note several execution challenges. They list the need to secure long-term customers, maintain reliable power and cooling, ensure network resilience, preserve export-control stability and develop local operations expertise. How local companies and institutions use the capacity, at what price and for which workloads is expected to affect the economic outcome.

Officials and project materials describe the initiative as converting imported advanced chips into domestic compute capacity for training and hosting AI systems. The project’s timeline, scale and commercial terms depend on coordination among Firebird, infrastructure suppliers, regulators, telecom upgrades and local partners.