Kazakhstan Adds Vera Rubin Plans: What NVIDIA Says Is Deployed, and What Is Still a Plan

NVIDIA says Kazakhstan's AI buildout is "moving from infrastructure plans to deployment," and it put numbers on that in an October 2026 blog post published during its GTC Berlin coverage. Kazakhtelecom, the country's largest telecom company, has already deployed 400 NVIDIA H200 and 480 NVIDIA B300 GPUs in Astana and Almaty. Next year it plans more than 1,000 GPUs in the Data Center Valley and more than 1,000 Vera Rubin GPUs in the Ekibastuz region.

Those are small numbers next to the headline figure from June, when the government, Firebird and NVIDIA signed a package described as $10 billion with a target cluster of 100,000 GPUs. This post lays out what is deployed, what is planned, and how to read the gap. For the wider context on nations building their own compute, see our explainer on what sovereign AI is.

TL;DR

What does NVIDIA's post actually say?

NVIDIA frames the update as three layers: infrastructure, models and applications. The infrastructure layer is Kazakhtelecom's hardware. The post says the planned Vera Rubin deployment in Ekibastuz will serve "diverse workloads with infrastructure that enterprises, developers and institutions can use at scale."

The application layer is the second announcement. inDrive and Pony.ai are collaborating "to explore robotaxi services in Kazakhstan, with Astana as the first deployment city." They will assess the feasibility of bringing Pony.ai's seventh-generation robotaxis to the inDrive app, supported by NVIDIA AI infrastructure built on the NVIDIA Hyperion sensor and compute platform with the NVIDIA Halos safety system. The wording is deliberately tentative: "explore," "assess the feasibility." There is no launch date and no vehicle count.

The post closes with a point worth keeping: "Kazakhstan's AI opportunity isn't defined by GPU counts alone. Connecting that infrastructure with developers, universities, startups and industries can turn accelerated computing into useful applications."

How does it connect to the $10 billion Data Center Valley deal?

The June deal is documented by the Kazakh Prime Minister's office. On June 15, 2026, Prime Minister Olzhas Bektenov met NVIDIA Vice President Rev Lebaredian and Firebird's co-founders, and agreements totaling $10 billion were signed. Two are named: a strategic cooperation agreement between the Ministry of Artificial Intelligence and Digital Development and Firebird, including Firebird Labs Kazakhstan at Alem.ai; and a binding term sheet between Kazakhtelecom and Firebird on the technical and organizational parameters of the Data Center Valley project.

Key claims in the government's release:

• The Data Center Valley in Ekibastuz has 300 MW of power available and is designed for phased expansion to 1 GW.
• The planned $10 billion investment would build a computing cluster of 100,000 GPUs, including GB300 and Vera Rubin, per Deputy Prime Minister and AI minister Zhaslan Madiyev.
• Madiyev estimated at least $3 billion in annual export revenue. That is an official projection, not a result.
• Firebird CEO Razmig Hovaghimian said the launch planned for 2027 would put Kazakhstan among the world's top ten countries in AI infrastructure. That is a company claim.
• Kazakhtelecom chairman Bagdat Mussin described the plan as turning "Ekibastuz coal into digital export revenue," and said Firebird already has service agreements with major global players. No customers were named.

Reading the two together: the 100,000-GPU figure is the end-state ambition, the 880 deployed GPUs are today, and the 1,000 plus 1,000 GPUs of 2027 are the next step. Nothing in NVIDIA's October post says the 100,000 target has moved or been met.

Deployed versus planned

For scale, consider how other announcements in this series read. Our coverage of Google Cloud's Vera Rubin general availability and the first Vera Rubin cluster at CoreWeave shows how fast Rubin capacity is appearing at hyperscalers. A planned 1,000-GPU Vera Rubin site is a national-scale pilot by comparison, not a hyperscale campus.

Why does Kazakhstan fit the sovereign AI playbook?

Lebaredian told the Kazakh officials in June that AI is a "five-layer cake" and that "everything begins with energy." Kazakhstan, he said, has energy in abundance. That matches the sovereign AI pattern NVIDIA has been promoting at GTC Berlin: a country with power and capital buys accelerated computing, offers it to local developers and enterprises, and tries to sell capacity abroad. We have tracked similar efforts in the UAE and Saudi Arabia and across Europe, including the Nebius hub in Madrid.

What differs here is the export framing. The Kazakh officials explicitly pitch the cluster as an export business, selling computing to foreign companies for foreign currency, rather than only serving domestic users. That raises questions the sources do not answer: who the customers are, what the contracts look like, and how data residency and chip-export rules apply to the buyers. Treat those as open.

Timeline so far

• 2025: Kazakhstan launches a supercomputer based on NVIDIA H200 chips, according to press coverage of the June deal.
• June 15, 2026: The government, Firebird and NVIDIA sign the $10 billion package of agreements, with the Data Center Valley in Ekibastuz at the center. The same week, Kazakhstan adopts its "Digital Qazaqstan" national strategy, and 2026 has been declared the Year of Digitalization and Artificial Intelligence.
• October 21 to 23, 2026: NVIDIA's GTC Berlin run includes the Kazakhstan update with the 400 H200 and 480 B300 deployed-GPU figures.
• 2027: Planned Vera Rubin deployment in Ekibastuz, more than 1,000 GPUs in the Data Center Valley, and Firebird's stated launch target.

How to read a national GPU announcement

Three habits help with any announcement like this one. First, separate the tense: "has deployed" is a present fact, "plans to deploy" is a forecast, and "aims for" is an ambition. NVIDIA's post uses all three, and only the first describes hardware that exists. Second, check who is speaking. NVIDIA benefits from showing demand for its chips, and officials benefit from showing progress, so neither is a neutral counter. Third, look for the unit of account. GPU counts hide differences between an H200, a B300 and a Vera Rubin part, and between a single accelerator and a rack. A thousand Vera Rubin GPUs and a thousand H200s are very different amounts of compute, and the post does not say how the planned Rubin GPUs are packaged.

It is also worth asking what the compute will be used for. The post points to enterprises, developers, institutions and robotaxis, but names no model, no university partner and no pricing. Until those appear, the most defensible summary is that Kazakhstan has a real, modest cluster, a signed framework for a much larger one, and a government eager to say so.

What is not known

1. Timeline of the 100,000-GPU cluster. Firebird says launch is planned for 2027; the October post gives no new date for the full build.
2. Financing. The $10 billion is described as planned investment. The breakdown between Firebird, Kazakhtelecom and the state is not in the sources we read.
3. Customers. Firebird is said to have service agreements with "major global players," none named.
4. Power delivery. 300 MW is available and 1 GW is the design target; the Vera Rubin phase's power needs are not given.
5. Robotaxi regulation and timing. No permits or dates are mentioned.

What this means for people building with AI

For developers, the practical effects are indirect:

• More regional GPU supply in Central Asia could eventually mean lower-latency inference and more local-language model work. That depends on whether capacity is opened to outside developers or sold to large foreign buyers.
• Another data point on Vera Rubin rollout. Planned Rubin capacity is spreading from hyperscalers to national projects, but in small numbers at first. See the NVIDIA DSX Pod claimed-versus-verified review for how skeptical to be of deployment-speed claims.
• Watch the inference market. If more national clouds sell GPU time, rental prices and availability can shift. Our tracker on B300 rental prices shows demand is still running hot.

What to watch

• Independent confirmation of the 880 deployed GPUs, for example from Kazakhtelecom or Firebird filings and customer announcements.
• Whether the 2027 Vera Rubin deployment is on schedule.
• The first named customer for Data Center Valley capacity.
• Any permit or pilot date for the inDrive and Pony.ai robotaxis.

The broader GTC Berlin announcements are summarized in our keynote roundup.

Bottom line

NVIDIA's Kazakhstan update is a progress report with concrete but modest numbers: 880 GPUs deployed, about 2,000 more planned next year, and a 100,000-GPU target that remains an ambition. The story is worth following as an example of a resource-rich country trying to turn power into AI compute exports. It is not yet evidence that the full cluster exists.

Related reading

• What is sovereign AI? Layers, models and trade-offs
• GTC Berlin 2026 keynote: everything announced
• Google Cloud A5X Vera Rubin general availability
• CoreWeave's first Vera Rubin cluster
• UAE and Saudi Arabia AI landscape
• Nebius Madrid AI hub
• NVIDIA DSX Pod: claimed versus verified

Details are accurate as of October 23, 2026, based on NVIDIA's blog post and the Kazakh government's June 15, 2026 release; all deployment figures are company or government statements.