NVIDIA’s August 8 announcement calls the facility the Commonwealth of Independent States region’s largest AI factory and says it will serve local developers, universities, public institutions and overseas AI companies. Firebird’s parallel announcement describes the site as its first operational factory and names Perplexity as an early customer for compute supporting its answer engine and AI-powered digital coworker product.
For Windows administrators, infrastructure teams and AI builders evaluating new GPU capacity, the launch creates a potentially consequential alternative geography for hosted bare metal and sovereign-compute projects. But it also arrives with a familiar AI-infrastructure gap between what is installed today, what is reserved or approved, and what is being projected several build phases ahead.
The operational site is far smaller than the headline roadmap
Firebird’s current infrastructure page identifies DC-1 in Armenia as a 15-megawatt installation with 6,144 B200 GPUs. Armenia’s Ministry of High-Tech Industry had previously described the opening phase as more than 6,000 Blackwell GPUs at 18 MW, so the public record has a small but material mismatch on power capacity. Neither Firebird nor NVIDIA explains whether the 15 MW figure is IT load while 18 MW is total site capacity, or whether one of the numbers reflects an earlier design target.
That distinction is more than bookkeeping. A 6,144-GPU B200 deployment is a serious cluster that can support large training runs, batch inference, simulation and dedicated enterprise tenancy. It is also roughly one-twelfth of the 70,000-plus GPU figure NVIDIA now cites for Armenia by the end of 2027.
NVIDIA and Firebird describe that later figure as a three-phase expansion to 300 MW, using Blackwell hardware and NVIDIA’s future Rubin platform. It is a roadmap, not equipment already operating in Hrazdan. Firebird’s own current site map makes the difference clearer: it lists the operating Armenian factory at 6,144 B200 GPUs, then a 75,000-GPU Vera Rubin expansion at 125 MW, plus another planned 75,000-GPU, 125-MW facility.
Those numbers do not line up cleanly with the NVIDIA Blog’s 70,000-GPU, 300-MW Armenia goal. Firebird’s site map implies a much larger long-term three-site footprint—more than 156,000 GPUs when its listed Armenian phases are added—while its opening-day release uses the more conservative 70,000-GPU figure through 2027. Readers should treat the 6,144-GPU figure as the usable baseline and each later total as a separate stage of an expansion plan.
The claim that this is the CIS region’s largest AI factory deserves similar care. NVIDIA provides no inventory of competing facilities, no definition of “CIS region,” and no methodology for comparison. Firebird itself says “largest NVIDIA DSX AI factory in the region,” which is a narrower claim. The Hrazdan installation may well be the largest announced or operational Blackwell-focused cluster in the Caucasus and surrounding markets, but the broader regional superlative is not independently substantiated in either company’s announcement.
What Firebird is actually selling: bare metal, cloud access and Armenian data residency
The practical product is not a national supercomputer reserved exclusively for Armenian research. Firebird is positioning the facility as an AI cloud and bare-metal platform, with Kubernetes, Slurm workload scheduling, managed inference endpoints, GPU-aware provisioning and software-defined L2/L3 networking. Its public materials describe NVIDIA InfiniBand for east-west GPU traffic, Ethernet for north-south traffic, and NVIDIA BlueField-3 DPUs on GPU nodes for tenant isolation.
That is the part enterprise infrastructure teams should focus on. An organization that needs dedicated GPU nodes, a private cluster or HPC scheduling can work with a provider offering physical capacity rather than queueing for shared public-cloud instances. Firebird says it offers both on-demand cloud access and single-tenant servers, although its website presently directs prospective customers to request access or join a waitlist rather than publishing instance types, availability zones, SLAs or pricing.
The service also comes with a clear residency consequence. Firebird’s security policy says its cloud infrastructure is hosted exclusively in Firebird-operated facilities in Armenia. For organizations that specifically need Armenian data residency or want compute in the South Caucasus, that may be the point. For a U.S. or European enterprise, it means data-location, backup, identity federation, incident response, contractual jurisdiction and cross-border transfer reviews belong at the start of procurement—not after workloads have been placed.
Firebird’s site indicates it intends to devote 90% of Armenia-based compute to U.S. companies operating in the region, an unusually explicit statement for a project framed around domestic AI capability. Armenia is also putting money behind local use: the country’s High-Tech Industry Ministry announced a five-year, $25 million agreement to buy high-performance computing resources from Firebird. That gives the factory a domestic anchor customer worth an average of $5 million annually, but it does not answer how much capacity will be set aside for state projects, researchers, startups or foreign commercial customers.
Perplexity’s announced role is encouraging but narrow. NVIDIA and Firebird say the company will access Firebird infrastructure; neither release specifies GPU count, contract duration, deployment date, workload location or whether Perplexity will run production inference in Armenia rather than use the capacity for testing or overflow. Perplexity has not published a separate announcement detailing the arrangement.
NVIDIA DSX is a reference architecture claim, not a measured efficiency result
The Hrazdan site uses NVIDIA’s DSX AI Factory reference design, Dell servers, Schneider Electric power equipment and Vertiv chilled-water cooling. Those are established infrastructure suppliers, and the split of responsibilities is plausible: Dell supplies the compute platform, Schneider Electric supplies medium- and low-voltage switchgear, UPS systems and racks, while Vertiv delivers thermal management and chilled-water controls.
NVIDIA says DSX can fit up to 40% more GPUs in the same footprint and improve tokens-per-dollar economics. The wording matters. “Up to 40%” is a platform-level design claim, not an independently disclosed result from Firebird’s live site. There are no published figures for power usage effectiveness, rack density, actual per-GPU power draw, water consumption, cooling redundancy, generator runtime or the network oversubscription ratio.
For a GPU facility, those measurements determine whether a theoretical cluster becomes dependable production capacity. A model-training environment can absorb some scheduling delay; customer-facing inference cannot. The company’s claims around automated provisioning, physical isolation, observability and managed endpoints are promising, but Firebird has yet to publish the operational detail an enterprise buyer would normally request: availability commitments, maintenance windows, supported operating-system images, firmware lifecycle policy, incident escalation, data-destruction process, or a compliance portfolio.
The deployment speed is also being presented in the best possible light. Firebird says it went from an empty construction site to operational readiness in just over six months. Earlier Armenian government reporting placed the phase-one investment at about $500 million, with the project funded in part by a $300 million syndicated loan from six Armenian financial institutions. That financing and the visible June shipment of initial Blackwell hardware show the project moved beyond a memorandum or render. Still, “operational readiness” does not disclose utilization, customer availability or whether all 6,144 GPUs were online on opening day.
The investment announcement is an intention, not a disclosed transaction
NVIDIA says it intends to invest in Firebird, following an earlier CoreWeave investment this year. It did not disclose the amount, ownership stake, closing date, valuation, board rights, conditions or whether the investment has already closed. Firebird’s own release uses the same future-tense language.
That is a material omission because this buildout will require far more than a successful first site. Firebird is targeting 300 MW in Armenia by the end of 2027, has secured 125 MW of capacity at Data Center Valley in Kazakhstan according to its announcement, and says its broader pipeline could reach about 2 GW by the end of 2028. Those expansion objectives depend on export approvals, hardware allocation, financing, grid capacity, cooling construction, network delivery and customers able to commit to large GPU contracts.
The U.S. export component is central. Armenia’s Ministry of High-Tech Industry has tied the Firebird project to the U.S.-Armenia memorandum on AI and semiconductor cooperation signed in Washington on August 8, 2025, while Firebird says its Kazakhstan expansion has received U.S. Bureau of Industry and Security authorization. For buyers, this makes export compliance part of vendor risk assessment: access to the most advanced accelerators is enabled by approval processes that can change with policy and destination controls.
The first procurement test is availability, not GPU totals
Firebird has crossed an important threshold: it has a named location near Hrazdan, hardware at the site, major OEM and facility partners, a government compute-purchase agreement and at least one named commercial customer. That is more concrete than much of the worldwide AI-factory pipeline.
But the launch should be read as phase one of an infrastructure business, not proof that Armenia now has 70,000 GPUs or 300 MW of AI capacity. The operational figure is 6,144 B200s, with published power figures ranging from 15 MW to 18 MW. Everything above that depends on later construction and procurement phases.
The next concrete milestone is whether Firebird begins publishing the information customers need to use the capacity: available configurations, allocation sizes, network performance, service commitments, prices and evidence that the advertised cloud and bare-metal services are open beyond named partners. Until then, Hrazdan is a real and meaningful new Blackwell site—while its much larger numbers remain an expansion plan.
References
- Primary source: NVIDIA Blog
Published: August 8, 2026 at 10:24 AM UTC
Loading…
blogs.nvidia.com - Related coverage: firebird.ai
Loading…
firebird.ai - Related coverage: dell.com
Loading…
www.dell.com - Related coverage: en.armradio.am
First batch of NVIDIA Blackwell GPUs arrives in Armenia as Firebird AI megaproject enters implementation phase – Public Radio of Armenia
The first major shipment of NVIDIA Blackwell graphics processing units (GPUs) for Armenia’s ambitiouen.armradio.am - Related coverage: firebird.ai
Loading…
www.firebird.ai - Related coverage: old.hightech.gov.am
Loading…
old.hightech.gov.am - Related coverage: 0e190a550a8c4c8c4b93-fcd009c875a5577fd4fe2f5b7e3bf4eb.ssl.cf2.rackcdn.com
Loading…
0e190a550a8c4c8c4b93-fcd009c875a5577fd4fe2f5b7e3bf4eb.ssl.cf2.rackcdn.com - Related coverage: firebird-technologies.com
Loading…
www.firebird-technologies.com - Related coverage: firebird-technologies.com
Loading…
www.firebird-technologies.com - Related coverage: music.amazon.in
Loading…
music.amazon.in - Related coverage: nvidianews.nvidia.com
Loading…
nvidianews.nvidia.com - Related coverage: aiomnitech.com
Loading…
aiomnitech.com - Related coverage: investor.nvidia.com
- Related coverage: semiconductors.einnews.com
Loading…
semiconductors.einnews.com - Related coverage: nvidianews.nvidia.com
Loading…
nvidianews.nvidia.com - Related coverage: medium.com
Loading…
medium.com - Related coverage: old.hightech.gov.am
Loading…
old.hightech.gov.am - Related coverage: hightech.gov.am
Loading…
hightech.gov.am