Gcore and TengriAI Eye 100 MW AI Data Center in Kazakhstan

Gcore and TengriAI have signed a memorandum of understanding to explore building a GPU data center in Kazakhstan with a target capacity of up to 100 megawatts, phased between 2027 and 2029. Gcore announced the agreement on November 12, 2026. The deal is exploratory and does not represent a final construction commitment, but it sketches a concrete commercial model for bringing large-scale AI compute to a region the companies say has almost none of its own.

The stakes are straightforward. Central Asia currently lacks domestic AI compute, according to the companies, which leaves organizations in the region dependent on capacity in Europe, the Gulf or East Asia — capacity that often sits outside their regulatory jurisdiction. A 100 MW GPU facility inside Kazakhstan would change that arithmetic for every customer within reach of the country's power grid and fiber connections.

What does the agreement actually commit?

Not construction — yet. A memorandum of understanding is a study-and-scope instrument, and both companies framed it that way. What it does commit is a division of labor that shows how each side expects to make money if the project advances.

Under the MoU:

• TengriAI would handle land, power, permitting and construction — the physical and regulatory layer of the proposed facility.
• Gcore would provide the AI cloud stack, operations and customer demand, using the same platform it runs in its own GPU clouds as an NVIDIA Preferred Cloud Partner.

The arrangement gives each company a distinct layer of the proposed project, running from the powered site all the way to the sale and operation of GPU capacity. Customer development is built directly into Gcore's role, which makes demand generation central to any move from exploration toward the phased 100 MW target. A data center without tenants is an expensive building; this MoU names the party responsible for finding the tenants.

Gcore CEO Andre Reitenbach framed the project's regional ambition in one sentence: "The project would put large-scale GPU infrastructure in the country, operated from the country, for customers across the region."

How would customers use the capacity?

The proposed facility would offer two operating models, and the difference between them matters for the kinds of business the project could support.

The first option lets an operator meter, bill and resell capacity. This supports a regional cloud business serving multiple users — a local company or carrier could buy GPU capacity wholesale and sell it onward, in the way telecom operators already sell bandwidth and colocation.

The second option lets organizations run isolated workloads that never leave the facility. This targets customers that need tighter control over where their AI workloads run — the class of government, financial and enterprise users for whom data residency and jurisdiction are not negotiable.

The two models are not mutually exclusive. The companies' emphasis on both metered resale and isolated deployment suggests the proposed data center could support a multi-customer local AI-cloud business alongside sealed, single-tenant deployments.

What sits inside Gcore's AI cloud stack?

Gcore's AI Cloud Stack documentation, cited in the announcement, describes the software as a private AI cloud platform for turning GPU infrastructure into enterprise services. The component list explains how raw megawatts become billable products:

• Orchestration, monitoring, billing and around-the-clock operational support
• On-demand virtual machines or bare-metal GPUs
• Managed Kubernetes
• GPU access as a service

The platform also carries the controls a multi-tenant facility requires. Identity controls, quotas and multi-tenancy across CPU, network and storage resources allow multiple customers to share infrastructure while keeping separate access and separate resource allocations.

Two features of the stack bear directly on the Kazakh project's business case. Complete white-labeling would allow a local operator to brand the platform as its own — a Kazakh carrier or cloud firm could sell GPU services under its own name without building the software layer underneath. Reseller-channel support provides a mechanism for selling capacity through intermediaries rather than only direct to end customers.

Those features explain the emphasis on metering and resale in the MoU. The software is built to support exactly the operating models the proposed facility would offer.

Why Kazakhstan?

The companies' site rationale rests on four inputs: competitively priced energy, available land, lower cooling costs and a government that treats AI infrastructure as a national priority. Energy price is the anchor. GPU data centers are, above all, electricity buyers, and a 100 MW facility runs 24 hours a day; a meaningful per-kilowatt-hour advantage compounds into a large operating-cost gap over a facility's life.

The rationale extends beyond operating costs to regional reach and legal setting. Kazakhstan's location positions a facility there to serve neighboring Central Asian markets. The Astana International Financial Centre's common-law framework gives customers in those markets a familiar legal environment for contracts — a consideration the companies weight explicitly, since jurisdiction is one of the pain points the project aims to solve for regional customers.

TengriAI General Director Nurassyl Medeu set the ambition in regional terms. The goal, he said, is to make Kazakhstan "a leading AI token hub in Central Asia." He cited Gcore's experience operating GPU clouds at scale and working with telecom operators as the basis for that objective.

What happens next?

The MoU links Kazakhstan's energy, land and jurisdictional advantages to a specific commercial model: local physical infrastructure, Gcore-operated cloud software, and capacity that can be resold or kept isolated. The 2027–2029 phasing gives the partners a runway of roughly one to three years from signing to first power, if exploration converts into construction.

The open variable is demand. The agreement assigns customer development to Gcore, and every element of the software stack — metering, billing, white-labeling, reseller channels — exists to convert interest into contracted capacity. The distance between a memorandum signed in November 2026 and a phased 100 MW GPU campus operating in 2029 will be measured in signed customers, not megawatts on paper.

If the project proceeds, it would give Central Asia its first large-scale domestic AI compute — and give organizations in the region, for the first time, the option of running AI workloads inside their own regulatory jurisdiction at facilities built and operated for exactly that purpose.