Firebird’s US$ 5 billion Kazakhstan AI factory gets green light for 45,000 NVIDIA GPUs - W.Media

This is not a data center announcement — it is NVIDIA's compute diplomacy taking root in Central Asia. Forty-five thousand GPUs (likely H200/B200 class) form a 200-400 PFLOPS training cluster. The siting logic is brutally simple: Kazakhstan's $0.03-0.05/kWh natural gas electricity costs one-third of Texas. Upstream, this pulls 800G optics, HBM3E, and liquid cooling at scale. Downstream, it deepens CUDA lock-in. The real risk is regulatory, not technical. Kazakhstan currently sits outside the EAR entity list, but Commerce Department discussions on transshipment monitoring in Central Asia are intensifying. If the control radius expands, firmware updates, driver iterations, and spare parts for 45,000 GPUs become a single point of failure — a textbook compliance arbitrage with maximum exposure. Competitively, AMD MI325X and Intel Gaudi 3 have zero traction in the region. The actual contest is compute sovereignty: Saudi Humain, G42, and India's AI4Bharat are all racing to claim the non-Western AI infrastructure narrative. Firebird's $5B is fundamentally about anchoring a compute node for Central Asia. 12-24 month outlook: expect 3-5 sovereign AI projects of similar scale across the region. NVIDIA's business model is shifting from selling chips to selling compute sovereignty — silicon plus software stack plus operations, bundled. Electricity cost will replace land and talent as the primary variable in AI factory siting.