Business
Firebird, Nvidia Plan 50,000-GPU Yerevan Data Center in $4B AI Bet
April 23, 2026
4
Min
US-based Firebird Inc. and Nvidia Corporation have outlined a plan to build a large-scale AI compute facility in Yerevan, deploying approximately 50,000 next-generation GB300 graphics processors by the end of 2026, with total project investment potentially reaching $4 billion.
The announcement, embedded in the broader US strategic investment push that accompanied Vice President JD Vance's February visit to the South Caucasus, represents one of the largest single AI infrastructure commitments ever targeted at a country of Armenia's size. Backdrop from OilPrice.com.
Why Yerevan? Three structural factors. First, Armenia already has a technically literate developer base, with IT exports growing at double-digit rates and a deep relationship between Yerevan universities and US-based diaspora engineering networks. Second, the US Section 123 civil nuclear cooperation agreement, signed in February, creates a clear line of sight to long-run low-carbon baseload power — critical for hyperscale AI training clusters. Third, a post-TRIPP Armenia will have dramatically improved physical connectivity and terrestrial fiber optionality through southern corridors.
The 50,000 GB300 figure is notable. Nvidia's GB300 class GPUs are among the most advanced systems available for large-model training, and a deployment of that scale would place the Yerevan facility in the same compute-capacity bracket as mid-sized US hyperscaler campuses. Bloomberg's reporting frames the project against a broader $13 billion US investment push toward Armenia.
Near-term execution questions are material. Power availability at the proposed site, cooling design (air, liquid, or hybrid), and access to long-dated PPAs are all unresolved. Armenia's Metsamor nuclear plant currently produces ~440 MW, a number that will not be sufficient alone to cover hyperscale AI plus baseload industrial demand without additional generation. US-backed SMR deployment is part of the longer-term solution.
Governance and export-control considerations add complexity. Any facility of this scale will require BIS licensing for GPU export, Armenian regulatory approvals, and potentially EU-aligned privacy standards if European customer workloads are targeted. The framework is achievable but not trivial.
For Armenia, the upside is transformational. Even partial execution — say, 10,000 GPUs operational by end-2026 — would lift Armenian ICT GDP contribution materially, attract an additional wave of engineering talent, and position the country as a second-tier Caucasus data infrastructure hub distinct from regional peers. The tax stack, combined with recent capital market reforms, supports the case.
For Nvidia, the project is consistent with a broader strategy of distributing compute capacity across allied jurisdictions that meet export-control and alliance-security criteria. The Firebird-led operating model suggests a specialist AI infrastructure platform rather than a sidecar deployment.
Next milestones to watch: site selection, definitive power contracts, first rack deployment, and any follow-on commercial offtake announcements from US or European AI lab customers.
