45,000 Nvidia Blackwell GPUs Cleared for Kazakhstan Data Center

The approval is an export license for 45,000 Nvidia Grace Blackwell GPUs obtained by Firebird.ai. It allows the Armenian-American infrastructure company to bring the advanced chips to Kazakhstan. The project matters because it would give the country major computing capacity for artificial intelligence development.

The GPUs will be installed in a data center built with Kazakhtelecom. Firebird AI KZ Ltd. is constructing the first phase now. The facility is intended to support AI systems, data processing, and related technology businesses rather than simply export raw energy.

The first phase is planned to launch in the third quarter of 2027. A second phase could eventually raise Firebird’s infrastructure capacity in Kazakhstan to 500 megawatts. Firebird also plans a lab for AI startups and human-capital development, linking the hardware project to local skills and innovation.

A GPU, or graphics processing unit, is a specialized computer processor. Unlike a general-purpose processor, it handles many similar calculations at once. That parallel design was first associated with graphics, but it is also highly effective for the mathematics used by artificial intelligence.

AI models repeatedly process enormous arrays of numbers while learning patterns or generating answers. GPU clusters divide that work across many chips. Nvidia’s Blackwell family is designed for demanding AI workloads, so 45,000 of these GPUs would provide a substantial foundation for data-center computing. The article identifies them as the hardware Firebird plans to deploy.

This hardware matters because advanced AI depends on both software and physical infrastructure. Kazakhstan is seeking to develop local and Kazakh-language models, while Firebird plans a lab for startups and talent. The article does not provide the chips’ exact computing performance, but their scale signals a major planned investment.

The project is large in two different ways. Its first announced hardware allocation is 45,000 Nvidia Grace Blackwell GPUs. That means a very large collection of specialized processors operating together in one AI-oriented infrastructure project. The article does not provide a total number of servers, storage capacity, or computing speed.

Its possible physical scale is also substantial. Firebird says Kazakhstan’s infrastructure capacity could reach 500 megawatts after a planned second phase. A megawatt measures power capacity, not the amount of electricity consumed over a year. The figure therefore describes the potential scale of the data-center infrastructure, rather than directly stating its yearly energy bill.

The first phase is under construction and is targeted for the third quarter of 2027. The 45,000-GPU figure is approved hardware, while 500 megawatts describes a possible later capacity. Exact performance and consumption cannot be calculated from the article alone.

Kazakhstan has energy resources and is considering a different way to use them. Instead of mainly exporting energy, it could direct more electricity into data centers and AI development. The proposed shift could let the country sell computing capacity, AI services, and technology products rather than only the underlying energy.

Jensen Huang described this as transforming energy into intelligence. Firebird’s planned data center provides a concrete mechanism: electricity would power tens of thousands of Nvidia GPUs, while a planned lab would support AI startups and human capital. AI could then be applied across mining, agriculture, and other sectors of the economy.

The potential gain is a broader, more technology-focused economy. President Tokayev’s government is emphasizing AI, including Kazakh-language models. However, the article presents this as a future opportunity, not a completed transformation. The first data-center phase is still under construction, with launch planned for 2027.

The project brings together four main actors with different responsibilities. Nvidia is the chipmaker providing the Grace Blackwell GPUs. Its CEO, Jensen Huang, promoted Kazakhstan’s potential as a global hub for “Physical AI” and discussed deeper collaboration with Kazakh companies.

Firebird.ai is the Armenian-American infrastructure company developing the project. It obtained the export license for 45,000 GPUs, is building the data center, and plans a lab for startups and talent. Kazakhtelecom is Firebird’s local partner in constructing the facility. Firebird’s first phase is under construction, and a second phase could reach 500 megawatts.

The Kazakh government supplies political direction and national coordination. President Kassym-Jomart Tokayev met Huang and Firebird CEO Razmig Hovaghimian. AI Minister Zhaslan Madiyev highlighted Kazakh-language models. Together, these actors connect chips, infrastructure, local networks, policy, and workforce development.

Pax Silica is a US initiative designed to support AI and supply-chain reliability for Washington. It can also give participating countries access to advanced technology and chips. Kazakhstan joined Pax Silica in June, and the Firebird-Nvidia project could strengthen its position in that initiative.

WAICO is described in the article as a competing Chinese initiative. The potential conflict arose because a reported draft State Department note said countries joining WAICO would be ineligible for Pax Silica. Kazakhstan had joined both programs, creating uncertainty over whether it could retain access to the US-backed arrangement.

The situation later became less rigid. During a visit to Central Asia, US Secretary of State Marco Rubio rejected the idea that Kazakhstan had to choose between the two programs. He said Central Asian states were free to cooperate with any nation while building AI infrastructure. The article leaves the final membership outcome unstated.

Advanced AI systems process huge amounts of data through repeated mathematical operations. Specialized GPUs perform many calculations at the same time, making them suitable for training and running large models. The work must be repeated across many chips, so serious AI development requires data centers rather than a single computer.

Electricity powers the GPUs and the wider data-center equipment. It also supports cooling and reliable operation, although the article does not quantify those demands. Firebird’s planned 45,000-GPU deployment shows how hardware and power must be assembled together. Its possible 500-megawatt second phase illustrates the infrastructure scale involved.

This creates an advantage for countries with abundant energy, technical workers, and investment. Huang told Tokayev that Kazakhstan has richness in energy and in science, mathematics, and technology. If the country can connect those resources, it could build AI capacity. Without them, access to advanced models and chips becomes harder.