WorldIndustrialWeekly General Briefing: Kazakhstan AI Infrastructure as NVIDIA and Firebird Push Data Center Valley Forward

For manufacturing, energy and digital-infrastructure readers, the Kazakhstan programme involving NVIDIA and Firebird is less a single deal than a sequence of industrial milestones. It moves from a US export licence for 45,000 NVIDIA Grace Blackwell GPUs to construction at Data Center Valley in Ekibastuz, then to planned commercial operations in the third quarter of 2027, while the broader build-out references tens of thousands of GPUs, the Vera Rubin platform and partner Kazakhtelecom.

The timing was set by the 23 October 2026 visit of Jensen Huang to Astana, where he met Kassym-Jomart Tokayev and took part in a session with Zhaslan Madiyev, the minister responsible for artificial intelligence and digital development. That meeting placed hardware supply, power availability, national policy and local ecosystem development in the same conversation.

Who builds what, and where

Firebird is a US-based AI infrastructure company. Its first phase is being built with Kazakhtelecom at Data Center Valley, Ekibastuz. The licensed GPU block is described as 45,000 NVIDIA Grace Blackwell GPUs; Grace Blackwell is the architecture referenced for that volume. The commercial start date attached to this phase is the third quarter of 2027.

Existing deployments are separate from the new Firebird phase. Kazakhtelecom has already placed 400 NVIDIA H200 GPUs and 480 NVIDIA B300 GPUs across Astana and Almaty. The H200 and B300 systems are operational references, while the next set of additions is scheduled through 2027: more than 1,000 additional GPUs at Data Center Valley and, in the Ekibastuz region, deployments exceeding 1,000 Vera Rubin GPUs.

Capacity figures are not one single installed number

A practical way to read the announcements is by layer. The first layer is a licensed, under-construction block. The second is a possible second-phase ceiling that could lift Kazakhstan AI infrastructure capacity to approximately 500 MW. The third is a longer-term Data Center Valley expansion concept reaching 1 GW. The earlier $10 billion agreements, announced by Kazakhstan’s government with Firebird and NVIDIA, also refer to a wider ambition, including a computing cluster with up to 100,000 GPUs. These are not interchangeable installed-capacity measures.

Construction status published in August

Construction updates published by Kazakhstan’s government in August pointed to work on an initial 50 MW facility and acquisition of a 215 MW electrical substation, with that first facility expected in June 2027. That update provides a shorter-term marker than the commercial operations date in the third quarter of 2027 cited for the first Firebird phase.

From compute hardware to rack-scale platforms

Vera Rubin is described as combining next-generation GPUs with the Vera CPU and high-bandwidth memory in rack-scale computing systems. As eeNews Europe previously reported, NVIDIA began supplying customer samples of Vera Rubin hardware earlier this year, ahead of broader deployments. For industrial observers, the platform matters because it links procurement, physical data centre design and operational software stacks: GPU architecture, CPU, memory, racks, power and cooling are decided together.

Local ecosystem and application focus

The programme also includes commitments outside pure capacity. Firebird plans Firebird Labs Kazakhstan to give startups and researchers access to computing resources, funding and research partnerships. NVIDIA separately announced an investment in Higgsfield, a Kazakhstan-founded company building generative AI tools for video and image production; the investment amount was not disclosed. Higgsfield plans to expand computing capacity through Vera Rubin infrastructure supplied by cloud service providers.

Policy discussion during the visit pointed toward possible use of Kazakhstan’s energy resources to serve international AI computing demand, and toward Kazakh-language models and AI use in manufacturing, mining, agriculture and logistics.

Common mistakes when interpreting the project

• Treating the 45,000 NVIDIA Grace Blackwell GPUs licence as installed capacity. It is a permission and volume reference, not a live rack count.
• Equating 1 GW, approximately 500 MW and 50 MW as competing versions of the same number. They reflect different horizons and scopes.
• Reading up to 100,000 GPUs as part of the first commercial date. The earlier agreement set a wider ambition.
• Ignoring the difference between already deployed H200 and B300 systems and future Vera Rubin additions.
• Assuming construction updates and commercial-launch dates mean the same thing; the first Firebird launch is in the third quarter of 2027, while an initial facility was tied to June 2027.

Action suggestions for industrial readers

1. Track the 2027 milestones separately: June 2027 for the initial facility in government updates and the third quarter of 2027 for Firebird commercial operations.
2. Map announced capacity against physical infrastructure: Data Center Valley, Ekibastuz, the 215 MW substation, and the additional generating and distribution capacity implied by larger expansion.
3. Watch platform transition. Vera Rubin, Vera CPU and high-bandwidth memory may change server design, energy density and cooling expectations.
4. Use local ecosystem programs, such as Firebird Labs Kazakhstan and the Higgsfield investment, as leading indicators of application development, not just data-centre capacity.
5. For supply-chain and project teams, distinguish licensed hardware (US export licence, 45,000 NVIDIA Grace Blackwell GPUs), already deployed hardware (400 NVIDIA H200 GPUs, 480 NVIDIA B300 GPUs) and planned additions (more than 1,000 additional GPUs; exceeding 1,000 Vera Rubin GPUs).

FAQ: What the Kazakhstan AI infrastructure plan does and does not confirm

Q: Is the project only about GPUs? A: No. It links hardware, Data Center Valley construction, Kazakhtelecom deployment, national policy discussion, energy considerations, Firebird Labs Kazakhstan and Higgsfield application development.

Q: What does tens of thousands of GPUs mean here? A: It is the broad scale framing used for NVIDIA and Firebird expansion; the specific licensed block is 45,000 NVIDIA Grace Blackwell GPUs.

Q: Which numbers are closest to operation? A: The H200 and B300 systems in Astana and Almaty are already in place; the next planned additions are scheduled across 2027 at Data Center Valley and Ekibastuz.

Q: Why is energy central? A: The Ekibastuz development is linked to competitive electricity costs and access to existing power infrastructure, but expansion will require more generating and distribution capacity.

The first commercial launch in 2027 is the point at which announced capacity begins to become available compute. Until then, the story is one of staged commitments: a licensed Grace Blackwell block, construction at Data Center Valley, platform migration toward Vera Rubin, and a policy discussion that connects Kazakhstan’s energy position with AI applications across industry and language models.