On Saturday, Firebird, a US-based AI cloud company, opened what NVIDIA describes as the largest AI factory in the CIS region, in Hrazdan, Armenia. The ribbon-cutting included Armenian Prime Minister Nikol Pashinyan, Kazakhstan’s deputy prime minister Zhaslan Madiyev, and David Allen, the US chargé d’affaires in Armenia. That guest list captures the deeper story: a US company, with American diplomatic presence and a Kazakh official in attendance, opening advanced AI infrastructure in a country historically inside Russia’s sphere of influence.
The event was not simply a commercial launch. It was the physical result of a US export licence, a regulatory decision that allowed advanced NVIDIA hardware to enter Armenia in the first place. Firebird secured that licence in November 2025, and the subsequent scaling announcement in February 2026 made the political dimension explicit.
Key facts
- Firebird opened DC-1 in Hrazdan, Armenia, with 6,144 NVIDIA B200 GPUs across 15 megawatts.
- The facility was built on NVIDIA’s reference architecture and went from plan to operating capacity in just over six months.
- Firebird received a US export licence in November 2025, enabling advanced NVIDIA hardware to enter Armenia.
- Phase 2 was jointly announced by Firebird and the US government during Vice President JD Vance’s visit to Yerevan, scaling the project to $4 billion and around 50,000 GPUs.
- Targets include more than 70,000 Rubin and Blackwell GPUs and 300 megawatts of capacity by the end of 2027.
- NVIDIA, CoreWeave and Perplexity are among the key commercial names connected to the project.
An export licence as load-bearing infrastructure
None of this would exist without export approval. The US government controls the flow of advanced AI chips through export licensing, and Armenia is not a country that would normally receive cutting-edge data-centre equipment at this scale. The licence Firebird obtained in November 2025 was therefore not just a paperwork milestone; it was the foundational condition for the entire project.
The politics became explicit in February. Firebird and the US government jointly announced Phase 2 at a Yerevan press conference during Vice President JD Vance’s visit. That phase scales the project to $4 billion and roughly 50,000 GPUs, including licensing for a further 41,000 NVIDIA GB300 units. The first phase was $500 million, and the announcement claimed Armenia would host one of the world’s five largest AI GPU clusters.
Bringing a US vice president into the announcement turned a commercial project into a diplomatic signal. It showed that the United States is willing to use AI infrastructure as a tool of influence in the South Caucasus, a region where Russian influence has historically been dominant.
What is actually running
The opened flagship facility, DC-1, is specified at 6,144 NVIDIA B200 GPUs across 15 megawatts. It is liquid-cooled and built on NVIDIA’s reference architecture, which is designed to reduce the engineering risk for new entrants to the AI data-centre market. The target is more than 70,000 Rubin and Blackwell GPUs and 300 megawatts by the end of 2027.
That implies a twentyfold increase in power capacity in roughly 17 months. A planned DC-2 is listed at 75,000 VR200 GPUs across 125 megawatts, with a third site at similar scale. If these numbers hold, Armenia would move from near-zero AI data-centre capacity to one of the densest compute regions in Eurasia.
The delivery speed is genuinely notable. NVIDIA says the site went from plan to operating capacity in just over six months. Schneider Electric supplied switchgear, three-phase UPS systems and rack enclosures, while Vertiv provided chilled-water cooling. The rapid deployment suggests that the export licence was timed alongside a fully engineered construction plan, not an exploratory pilot.
The power trick
One of the most important technical details is how Firebird expects to make the economics work. The facility runs on NVIDIA’s DSX platform, which codesigns compute, networking, power and cooling as a single system rather than handling them as separate procurements. NVIDIA claims this allows up to 40% more GPUs on the same footprint by recovering stranded power.
Stranded power is a common problem in data centres: the electrical distribution is often provisioned for worst-case loads that never actually occur, leaving capacity unused. By integrating power management with the compute platform, DSX can allocate energy more dynamically and pack more accelerators into the same shell. For a site scaling from 15 to 300 megawatts, that coordination is where the capacity maths gets made. It is also, in NVIDIA’s own framing, a tokens-per-dollar argument rather than a performance one.
This is not just an engineering optimization. For a neocloud competing with hyperscalers, the ability to deliver a given number of tokens at a lower cost is the whole business model. GPUs are increasingly sold as a service, and the margin depends on how much inference work can be squeezed from every megawatt.
Who is paying for it
Firebird said NVIDIA intends to invest in the company, following an earlier CoreWeave investment this year. The pattern is familiar: NVIDIA takes a position in a company that then buys NVIDIA hardware at scale. It is a way for the chipmaker to anchor demand for its most advanced accelerators without building and operating clouds itself.
NVIDIA has now committed more than $40 billion to AI equity positions in 2026, and the structures have grown more creative. It took a $2.1 billion warrant in IREN as part of a five-gigawatt data centre deal. It also invested $1 billion in Naver while striking a $500 billion arrangement with SK Group, an approach critics have labelled circular financing.
In the Armenian case, the investment is smaller but symbolically important. A NVIDIA equity stake gives Firebird credibility with future customers and lenders, while also aligning Firebird’s purchasing decisions with NVIDIA’s roadmap. The same logic runs through many sovereign AI deals announced in the past year.
The neocloud playbook
Firebird fits a category that barely existed three years ago: companies that buy GPUs at scale and rent them to model builders. These firms, often called neoclouds, sit between NVIDIA and the application layer. They do not own foundation models or consumer apps; they own hardware and sell access to it.
CoreWeave, an early Firebird backer, has signed a multi-year deal with Anthropic to run Claude at production scale. That deal made CoreWeave one of the most important infrastructure providers in the AI boom. Other neoclouds have raised huge sums on the strength of long-term contracts with AI labs rather than revenue diversification.
Europe has its own version. Nscale committed €695 million to Portugal with Microsoft, reaching a $14.6 billion valuation in two years, on much the same thesis about sovereign capacity. The idea is that countries and regions will want AI compute located within their borders, and that a company able to build it quickly can capture a strategic premium.
Firebird’s first named customer is Perplexity, which is using the Armenian site for its answer engine and AI agent platform. Perplexity is not a hyperscaler, but it has become a significant consumer of GPU compute, and its willingness to use a new facility in Armenia is a sign that workload routing is becoming less geographical and more commercial.
Where this goes next
Firebird is pursuing an approximately two-gigawatt roadmap spanning Armenia, Kazakhstan and further markets, which explains why a Kazakh deputy prime minister was at an Armenian ribbon-cutting. Kazakhstan may be the next site, and the presence of its official signals that Central Asian governments are watching closely.
The strategic logic is straightforward. Compute is becoming the thing countries align around, and the United States has found a way to place it in a region where its influence has historically been thin. Export licences, equity stakes and joint announcements are all instruments of that policy.
The open question is demand. Building 300 megawatts in the Caucasus assumes customers will route workloads there rather than to cheaper or closer capacity, and one answer engine does not settle that.