Firebird's Armenia AI factory is live. The 300 MW promise is not

Firebird's new AI factory in Hrazdan is open. That is the part of the story worth taking literally. The launch makes Armenia an operating site for NVIDIA Blackwell compute. It does not turn every number attached to Firebird—70,000 GPUs, 300 megawatts in Armenia, 2 gigawatts worldwide—into capacity customers can use today.

The distinction matters because an AI factory is not a product when the ribbon is cut. It becomes one in stages: chips arrive, clusters are commissioned, grid power becomes firm, cloud access opens and customers start paying for sustained workloads. Firebird has crossed the first meaningful line. Its larger test begins now.

An operating site with an unresolved capacity number

Firebird's opening announcement describes the Hrazdan launch as the platform's operational debut. It does not state how many GPUs were commissioned on opening day, how much power was energized or what portion of the cluster was available to customers.

The company's current infrastructure page markets the first Hrazdan facility, DC-1, as 6,144 NVIDIA B200 GPUs across 15 MW. An Armenian government technical brief published in June described more than 6,000 GPUs, 18 MW and 110.6 exaflops of FP4 performance. No new commissioning report reconciles the 15 MW and 18 MW figures or confirms that every planned GPU is serving workloads.

That does not make the opening hollow. A built, powered Blackwell site in Armenia is a material infrastructure event. It does mean “operational,” “installed,” “available” and “utilized” should not be used as synonyms.

The megawatt mismatch may have a mundane explanation. Data-center announcements sometimes alternate between IT load, total facility consumption and grid capacity reserved for later phases. Those are not interchangeable. A buyer needs to know how much power can reach computing equipment after cooling and electrical overhead, while a grid planner cares about the facility's total draw. Firebird has not published definitions that let readers map 15 MW to 18 MW, nor a power-usage-effectiveness figure for Hrazdan. Until it does, both numbers should remain attributed to their respective sources.

Firebird and NVIDIA also call Hrazdan the region's largest AI factory. Without a public regional capacity census, that is a company claim, not an independently measured ranking. The more useful question is not whether the superlative survives. It is how quickly this first site converts a carefully financed national project into repeatable commercial capacity.

Phase one was financed like national infrastructure

The most concrete number behind Hrazdan may be neither GPUs nor megawatts. It is the debt.

Six Armenian financial institutions signed a $300 million syndicated loan for the project. Evocabank puts first-phase investment at roughly $450 million; Firebird has used a $500 million figure elsewhere. Those values may reflect different scopes, but the disclosures do not reconcile them. A separate $60 million Ameriabank facility should not simply be added to the syndicate because the public documents do not explain how the packages overlap.

Armenia also signed a five-year, $25 million compute-purchase agreement. That is not an equity investment. It is something more useful during a data center's early life: a sovereign customer committing to buy compute. The arrangement lowers demand risk while the commercial customer base develops.

This is the emerging sovereign-AI formula. Export permission opens the hardware lane; local banks finance the asset; government demand provides an anchor; a global infrastructure partner supplies operating credibility. TECHi found a related version of that structure in South Korea's sovereign AI factory financing. Armenia's version stands out because its domestic bank syndicate is large relative to the country's financial system.

The model can build phase one. It does not automatically fund the next 285 MW.

Debt also gives the project a clock. GPU infrastructure begins aging as soon as a faster generation reaches the market, while interest and principal payments do not wait for cloud utilization to mature. Firebird therefore has to fill B200 capacity while preparing Rubin-era halls without letting the first asset become an expensive demonstration cluster. The presence of government demand helps, but the economics will ultimately depend on commercial utilization, power cost and the price at which Firebird can sell each hour of compute.

The 300 MW target is an energy project

Firebird says it plans more than 70,000 Blackwell and Rubin GPUs and 300 MW in Armenia by the end of 2027. That is a roadmap, not current capacity. Scaling from a marketed 15 MW first site to 300 MW would require twenty times as much power before accounting for changes in rack density, cooling overhead or the timing of new GPU generations.

Armenia's own prime minister supplied the clearest constraint. In a June address to the National Assembly, Nikol Pashinyan said the country no longer had the electricity surplus it possessed 12 to 18 months earlier. He tied Firebird's full potential to new generation and storage.

That turns 300 MW into an energy-development promise. It depends on grid connections, generation, storage, cooling water or an alternative cooling design, and firm delivery schedules. A land reservation or a long-term power option is not the same as energized capacity.

The industry is moving in the same direction. NVIDIA's Rubin pitch increasingly emphasizes useful output per unit of power, which TECHi examined in its analysis of Vera Rubin and tokens per megawatt. That metric will matter more in Hrazdan than a raw GPU count. A 300 MW campus that arrives late or runs below contracted utilization can destroy capital even if its accelerator density looks impressive.

Firebird's roadmap still moves between disclosures

The launch release sets out more than 70,000 GPUs and 300 MW in Armenia by the end of 2027, followed by a global 2 GW target for the end of 2028. Firebird's own site presents a different set of project cards, including two planned 75,000-GPU facilities and a 400 MW headline total. Earlier government material referred to more than 100,000 GPUs by the end of 2027.

These figures may not cover identical geographies, power definitions or construction phases. That is precisely the problem. Investors, customers and policymakers cannot compare a project pipeline when “secured,” “planned,” “available” and “operational” are mixed across announcements.

The pattern is not unique to Firebird. Gigawatt AI campuses are often announced before grid queues, financing and accelerator deliveries have converged. TECHi reached a similar conclusion while examining Z.ai's gigawatt-scale data-center claim: the unit that matters is deliverable compute, not the largest number in the presentation.

Firebird can resolve the ambiguity with unusually simple disclosures: commissioned GPUs, energized IT load, contracted capacity and the date each figure was measured. Publishing those four numbers quarterly would make the company easier to trust than competitors that keep replacing one distant target with another.

Export clearance created access, not demand

Firebird's strategic advantage is real. Advanced NVIDIA accelerators remain subject to U.S. export controls, and Armenia obtained permission for the project. A February Armenian government readout said the United States approved permits for an additional 41,000 GB300 GPUs. That regulatory access is difficult for many proposed AI hubs to reproduce.

The permission answers one question: can the chips legally arrive? It does not answer whether the project can finance, power and fill them.

NVIDIA says it intends to invest in Firebird, but neither company has disclosed an amount, closing date or completed transaction. Firebird and NVIDIA say CoreWeave invested earlier in 2026, again without public terms. Those names strengthen the project. They should not be converted into financing totals that the companies have not published.

Firebird is making the same export-control argument in Kazakhstan, where it says it has secured 125 MW and U.S. authorization. Kazakhstan's official Data Center Valley announcement describes a 300 MW zone expandable to 1 GW, but does not mention a Firebird-specific U.S. license. Until a government or U.S. record confirms it, the authorization should remain attributed to Firebird.

Customer utilization is the quietest line

Perplexity is the only named commercial user in the launch materials. Even there, the verbs are forward-looking. Firebird says Perplexity “will access” its infrastructure; an NVIDIA account of the launch says the companies are working together on access. Neither statement shows that a production workload was live at Hrazdan on opening day.

No public announcement gives a Perplexity contract value, reserved GPU count, term or start date. Firebird's site still asks cloud customers to join a waitlist and bare-metal customers to request access. That can be normal during a controlled launch. It also means the facility's commercial availability is narrower than the word “operational” may suggest.

Naming a prominent customer is valuable because it signals that the platform has passed at least some technical and commercial review. It is not the same as disclosing a backlog. A small evaluation allocation, a burst agreement and a multi-year take-or-pay contract can all be described as access, yet they carry radically different implications for financing the next campus. Firebird has not said which one applies.

The $25 million Armenian government agreement is therefore important but insufficient as a demand signal for the full roadmap. A 300 MW buildout needs customers with long-duration commitments, credible workload growth and the balance sheets to keep paying when model economics change.

AI infrastructure demand is strong in aggregate. Individual campuses can still miss. Location, latency, data-sovereignty rules, network transit and developer tooling decide where workloads run. Cheap land and an export permit do not erase those frictions.

Three numbers would make the next update credible

Firebird's next disclosure should be judged on a compact operating scorecard:

• Commissioned compute: GPUs accepted into clusters and available for scheduled workloads, with the accelerator generation and measurement date.
• Energized capacity: firm IT load actually deliverable to halls, separate from land, grid reservations and future generation.
• Contracted utilization: capacity under signed customer commitments, with start dates and duration, without exposing customer-confidential pricing.

Those measures would show whether Hrazdan is progressing from a sovereign-backed launch into a durable cloud business. They would also let readers compare Firebird with other AI factories without relying on incompatible exaflop claims or end-of-decade power targets.

The opening still deserves attention. Armenia combined diplomacy, export clearance, domestic credit and public compute demand to put advanced AI infrastructure on the ground. That is harder than announcing a campus.

Now the standard changes. Firebird no longer needs to prove that it can hold a ribbon cutting. It needs to prove how much of Hrazdan is commissioned, how much is sold, and how the country will power the other 285 MW.