Firebird has secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs for an AI factory in Kazakhstan, according to NVIDIA and Firebird. The first phase is under construction with Kazakhtelecom at Data Center Valley, with commercial launch targeted for the third quarter of 2027. These are licensed and planned GPUs, not 45,000 accelerators already serving customers. (NVIDIA)
NVIDIA’s October 22, 2026, regional infrastructure announcement gives the project a concrete scale and schedule. Firebird’s October 23 release adds an estimated $5 billion investment, a cluster with potential capacity of up to 109 megawatts, and a commitment to reserve at least 10% of GPU capacity for Kazakhstani companies and government agencies. Those details make the project worth examining beyond its headline GPU count. (Firebird)
The useful distinction is between three things: permission to export hardware, construction of the infrastructure that will run it, and commercially available computing capacity. Firebird has announced progress on the first two. The third has a 2027 target.
What the Export License Covers
Firebird explicitly ties the 45,000-GPU export license to Phase 1 of its Kazakhstan AI Factory. Both companies identify Kazakhtelecom as the construction partner and Data Center Valley as the location. Q3 2027 means a targeted commercial opening between July and September, not a guarantee that every planned accelerator will become available simultaneously. Neither announcement provides a detailed commissioning schedule. (Firebird, NVIDIA)
The hardware descriptions become more specific in Firebird’s release:
- Export authorization: 45,000 NVIDIA Grace Blackwell GPUs.
- Planned cluster: More than 45,000 devices described by Firebird as NVIDIA GB300 GPUs.
- Investment estimate: $5 billion.
- Potential cluster capacity: Up to 109 MW.
The release does not reconcile its exact 45,000-device license figure with the cluster description of “more than 45,000” devices. The defensible headline number is therefore the licensed 45,000, while GB300 is Firebird’s more specific description of its planned equipment. Neither figure should be treated as a delivered inventory. (Firebird)
I’d put the construction status and launch quarter ahead of the investment estimate when assessing this announcement. A spending estimate indicates ambition; an identified site, partner and commercial target provide more useful milestones for tracking whether that ambition becomes accessible compute. The $5 billion figure remains Firebird’s estimate, not evidence that the full amount has already been spent or financed.
Installed GPUs and Planned Deployments Are Different Totals
Kazakhstan already has NVIDIA infrastructure operating outside the proposed Firebird buildout. NVIDIA says Kazakhtelecom has deployed 400 H200 GPUs and 480 B300 GPUs across Astana and Almaty. Its announcement also describes additional deployments planned for 2027. (NVIDIA)
| Deployment | Reported scale | Status in the announcement |
|---|---|---|
| Kazakhtelecom in Astana and Almaty | 400 H200 and 480 B300 GPUs | Already deployed |
| Kazakhtelecom at Data Center Valley | More than 1,000 GPUs, model unspecified | Planned for 2027 |
| Kazakhtelecom in the Ekibastuz region | More than 1,000 Vera Rubin GPUs | Planned for 2027 |
| Firebird AI Factory, Phase 1 | Export license for 45,000 Grace Blackwell GPUs | Under construction; Q3 2027 commercial target |
These entries describe distinct deployments. NVIDIA does not explain whether Kazakhtelecom’s additional Data Center Valley GPU plans overlap with any part of the Firebird project, so adding every announced number into a single national total would risk double counting. (NVIDIA)
GPU counts also cannot provide a clean performance comparison across H200, B300, GB300 and Vera Rubin deployments. Different hardware generations and system configurations can deliver different results depending on the workload. Neither announcement supplies a common benchmark, networking specification or usable cluster configuration for a like-for-like comparison.
The supported conclusion is narrower but still significant: the licensed Firebird deployment would add a much larger number of accelerators than the 880 GPUs NVIDIA identifies as already deployed by Kazakhtelecom. Its eventual contribution to training speed, inference capacity or customer pricing remains unquantified.
Why 109 MW, 500 MW and One Gigawatt All Appear
Firebird’s announcement contains three infrastructure scales that are easy to confuse. They refer to different scopes.
| Figure | What the published sources describe |
|---|---|
| Up to 109 MW | Potential capacity of the Firebird AI cluster associated with the announced investment |
| Approximately 500 MW | Firebird’s total potential Kazakhstan AI infrastructure capacity after a planned second phase |
| One gigawatt, or 1,000 MW | The broader Data Center Valley AI Compute Hub’s phased expansion ambition, described by Kazakhstan’s AI minister |
NVIDIA and Firebird agree on the approximately 500 MW figure: a planned second phase would bring Firebird’s total potential capacity to that level. It is not an additional 500 MW on top of the initial cluster. (NVIDIA, Firebird)
The gigawatt reference appears in a statement from Deputy Prime Minister and AI Minister Zhaslan Madiyev about the wider Data Center Valley hub. He describes “phased expansion planned to one gigawatt.” That wording does not establish a one-gigawatt Firebird commitment or give a completion date for the broader hub. (Firebird)
My reading is that the apparent 500 MW-versus-one-gigawatt discrepancy is principally a difference in project boundaries. One describes Firebird’s potential Kazakhstan footprint after Phase 2; the other describes a broader hub ambition. The sources do not fully map how those plans fit together, so treating either number as a replacement for the other would overstate the evidence.
Megawatts measure power capacity, not AI performance or annual electricity consumption. The announcements also leave unclear whether each figure uses the same boundary for computing equipment and supporting facilities. That limits precise comparisons between the power numbers, even before hardware choices and utilization enter the calculation.
The Local Access Commitment Matters for Sovereign AI
Firebird says it will allocate at least 10% of total GPU capacity to Kazakhstani companies and government agencies. Applied to the licensed 45,000-GPU scale, that percentage would correspond to 4,500 GPUs’ worth of capacity. This is a simple scale illustration, not a disclosed allocation of dedicated physical devices. ()
Frequently Asked Questions
3 questions
1Are Firebird’s 45,000 GPUs Already Operating?
No. NVIDIA and Firebird describe an export license and a first phase under construction, targeting commercial launch in Q3 2027.
