Firebird says its cluster meets 95 per cent of NVIDIA reference performance

Firebird's AI factory in Armenia has achieved NVIDIA Exemplar Cloud Achievement for AI training workloads on B200 GPUs, the company said on 12 October. It says the status confirms the infrastructure meets at least 95 per cent of NVIDIA's reference performance for the workloads evaluated.

The statement is the company's own. The release carries no quote from NVIDIA and no benchmark values.

The release says the recognition followed technical evaluations conducted with NVIDIA's engineering team. It does not name the workloads, state the size of the cluster tested, give the date of the tests or say whether anyone outside the two companies verified the results. It describes the site as having gone from an empty, snow-covered plot to a high-performance facility in six months, and puts the achievement at less than 10 months from that starting point.

Firebird's website, retrieved on 17 October, lists its first Armenian data centre as 6,144 NVIDIA B200 GPUs across 15 MW. In a 5 June release the company said the first phase would operate more than 6,000 Blackwell GPUs with a total power capacity of 18 megawatts, at an investment of $500 million. It opened the site near Hrazdan on 8 August and named Perplexity as one of its first customers, which it said would use the infrastructure for its answer engine and digital coworker platform.

The 8 August release did not give a utilisation figure or say how much of the first phase is in commercial use. The 12 October release does not either.

The first phase is the smaller part of what the company has announced. In February it said a second phase would add more than 41,000 NVIDIA GB300 GPUs and bring the project to about $4 billion. On 8 August it said NVIDIA intends to invest in Firebird, following an earlier investment by CoreWeave, and gave no amount for either. The 12 October release limits the new status to training workloads on B200 GPUs, the first-phase hardware, and says nothing about the later phases.

The company has not said whether it will publish the evaluation results. The items to track are:

• whether NVIDIA publicly confirms the status and the workloads it covers;
• whether Firebird publishes benchmark data for the cluster;
• how much of the first phase is under contract beyond Perplexity.