Firebird switched on the largest AI factory in the Commonwealth of Independent States on Friday, opening a Hrazdan, Armenia facility that will grow to more than 70,000 Nvidia Rubin and Blackwell GPUs and 300 megawatts of capacity by the end of 2027. The site was delivered in just over six months and forms the anchor of an approximately 2-gigawatt regional roadmap spanning Armenia, Kazakhstan and additional frontier markets. Nvidia said it intends to invest in Firebird, following an earlier 2026 investment by CoreWeave.
The opening drew Armenian prime minister Nikol Pashinyan, Kazakh deputy prime minister Zhaslan Madiyev and U.S. chargé d'affaires David Allen — a political turnout that signals how governments in the region are treating GPU capacity as sovereign infrastructure. Armenia has roughly 3 million people; a 300-megawatt AI campus at that scale is a national industrial project, not a colocation build.
The pitch from Firebird and Nvidia is that AI services may be globally available, but training and serving models in local languages and for local industries requires local compute. The Hrazdan factory is meant to give Armenian developers, universities, startups and public institutions capacity that today they would rent from hyperscalers abroad.
“AI factories are the foundational infrastructure for the AI era, providing the computing capacity needed to train, fine-tune and deploy AI models for every domain at scale.”— Rev Lebaredian, Nvidia VP
Key facts
- 01Firebird plans 70,000 Nvidia Rubin and Blackwell GPUs and 300 MW of capacity in Armenia by end of 2027.
- 02The full roadmap spans roughly 2 gigawatts across Armenia, Kazakhstan and other frontier markets.
- 03Nvidia intends to invest in Firebird, following an earlier CoreWeave investment this year.
- 04The Hrazdan facility was delivered in just over six months, with Perplexity among early customers.
- 05Built on Nvidia DSX, the design targets 40% more GPUs on the same footprint versus conventional builds.
The infrastructure runs on Nvidia's DSX platform, which co-designs accelerated computing, networking, power and cooling as a single system rather than assembling them piecewise. Dell PowerEdge servers host the GPUs. Schneider Electric supplied medium- and low-voltage switchgear, three-phase UPS systems and rack enclosures for the power backbone. Vertiv provided a chilled-water cooling architecture coordinated by its iCOM CWM Chilled Water Manager, a closed-loop design tuned for shifting AI workloads.
The efficiency claim is the commercial argument. Nvidia says DSX allows Firebird to run up to 40% more GPUs on the same footprint, which translates directly into more tokens per dollar and more revenue per megawatt — the two metrics that determine whether an AI cloud makes money or burns it.
“With DSX, it can run up to 40% more GPUs on the same footprint, producing more tokens per dollar and extracting more value from every megawatt of capacity.”— Rev Lebaredian, Nvidia VP
Early demand is already booked. Perplexity is working with Firebird to access capacity for its AI agent platform and answer engine, a notable customer given Perplexity's own ongoing negotiations with larger US and Middle Eastern infrastructure providers. Firebird has not disclosed pricing or contract terms with Perplexity.
Firebird's investor stack now includes CoreWeave — itself a $60-billion-plus public GPU cloud — and, imminently, Nvidia. That combination gives Firebird preferred access to allocation from Nvidia's Rubin and Blackwell production, which remains the binding constraint on GPU cloud growth industry-wide. Rivals without that access have watched their expansion plans slip by quarters.
The regional strategy is the more interesting bet. Kazakhstan has cheap power and cold weather; Armenia has a technical labor pool and Western political alignment. Between them, Firebird is assembling a footprint that competes for European and Middle Eastern AI workloads that would otherwise land in Ireland, the Nordics or the Gulf. A 2-gigawatt buildout at completion would put Firebird in the same rough scale bracket as CoreWeave's US footprint circa 2024.
The risks are real. Grid capacity in Armenia is finite, and 300 megawatts is a meaningful fraction of national demand — Schneider's involvement suggests the utility interconnection was itself the critical path. Six-month delivery of a facility this size is aggressive, and early operations at that pace tend to surface commissioning issues that add cost and slip revenue recognition. Nvidia's intended investment has been announced but not sized; until closing terms are disclosed, Firebird's capital runway to the 2-gigawatt target remains partly notional.
For the broader AI infrastructure market, the Firebird launch is a data point that the GPU cloud category is no longer a US-and-hyperscaler story. Well-capitalized regional operators with Nvidia allocation, closed-loop cooling designs and Dell hardware can stand up a competitive facility in six months in a country most AI customers could not have found on a map three years ago. That expands the addressable geography of AI compute — and it puts real pressure on incumbents to defend price and latency in markets they had assumed were theirs by default.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
