Firebird's Armenia AI factory lands 70,000 Nvidia chips by 2027
A data center that doubles as a diplomatic instrument, and a quiet upgrade to the tool most people already use to run models on a laptop.
Firebird's data center in Armenia is set to reach 300 megawatts and more than 70,000 Nvidia chips by the end of 2027 — hardware that Bloomberg reports the Trump administration agreed to let into the country as part of a wider diplomatic push.
The San Francisco startup began operating the facility in August. Nvidia calls it the largest AI factory in the CIS region, built on the company's DSX reference design and populated with Rubin and Blackwell GPUs. Firebird co-founder Alexander Yesayan told Nvidia the ambition is roughly 2 gigawatts of capacity worldwide within about two years.
The chips were not won on merit alone. The Wall Street Journal reported this month that US negotiators used the promise of Nvidia chip access for the Armenian data center to help broker a preliminary Armenia-Azerbaijan peace deal. Firebird's second phase — scaled to $4 billion and 50,000 GPUs — was announced in Yerevan in February by Vice President JD Vance, alongside Prime Minister Nikol Pashinyan and Nvidia's Rev Lebaredian, and a US export license followed authorizing imports of advanced Nvidia hardware.
That is the export-control regime being used as statecraft in the open, and it cuts both ways. Armenia gets a national-scale computing asset without a domestic chip industry to build one; Washington gets a compliant customer whose hardware it can still switch off at the license level. The bottleneck for projects this size is no longer silicon — it is whether a 300-megawatt load and the cooling behind it can be built on schedule in a country whose grid was never sized for it. Export controls as diplomacy is a pattern we have tracked before — the export-control net around advanced AI chips keeps widening the list of people who move the hardware.
Hugging Face added GGUF support to transformers, so the quantized checkpoints people already run locally now load through the standard Python API — and on Apple Silicon the throughput comes close to llama.cpp.
Transformers reuses ggml's Metal kernels through Hugging Face's kernels library: quantization, normalization, flash attention, the gated delta network inside Qwen's linear-attention layers, plus a top-k routing kernel the team wrote for MoE expert selection. Packed 4-bit weights are read in their stored format instead of being expanded into full-precision matrices before every decode step. On a MacBook Pro M2 Max with 32 GB of unified memory, Hugging Face's benchmarks land within a few percent of llama.cpp across a small dense model, a larger dense model and a mixture-of-experts model.
The near-term limits are real. The packed path is Apple-Silicon-only, covers the Qwen3.5 dense and MoE architectures plus compatible Qwen3.8 checkpoints, and targets one interactive conversation; llama.cpp remains the recommended engine when raw local speed is the point, and a 4B model still takes 2.74 GB at the standard 4-bit setting versus 8.42 GB unquantized. The more consequential change is portability — a kernel operates on tensors, not on a file format, so ggml's work can be pointed at architectures llama.cpp may never implement, vision and audio included. Two changes to the generation loop also help every transformers model, GGUF or not: dropping an unnecessary attention mask early, and deferring the stop check so the CPU keeps scheduling work instead of waiting on the GPU every token.
What to watch: whether the packed path reaches CUDA and multi-request batching, and whether Armenia's cooling and grid keep pace with a 300-megawatt schedule.
Is handing a country chips really a bargaining chip in a peace deal — and should it be? Tell us in the comments.
Sources: llama.cpp (GitHub)
