Ekibastuz: Close to the Power Plant, but How Far from Commercial Compute?

There is a timeline discrepancy in the Ekibastuz "Data Center Valley" that requires scrutiny. A report on November 7 still stated that the project would establish 125 MW of capacity in the first half of 2027 and expand to 1 GW across 2029–2033; however, a government update the previous day listed three Phase 1 facilities totaling 304 MW and set 1 GW as a 2029 target.[Z][G] Although the two reports were released just one day apart, their figures differ. Available evidence is insufficient to determine whether this represents a plan update, differing accounting boundaries, or a media report relying on an alternative plan. Directly adding them together or asserting that one has been superseded lacks a factual basis.

Megawatt figures in data center projects are easily mistaken for completed construction milestones, but they are primarily power capacity plans. Ekibastuz has a geographic advantage due to its proximity to power plants; yet turning this advantage into compute power that customers are willing to pay for still depends on whether electricity can be reliably delivered to racks, whether equipment can arrive legally and be installed and commissioned, whether the network can connect customers, and whether real workloads and service contracts can absorb the planned capacity.

The Project Has Made Tangible Progress

According to the government briefing on November 6, the first facility is undergoing foundation and supporting construction, excavation of the foundation pit for the second facility (the Firebird facility) is 70% complete, and the third facility remains in the design stage.[G] Here, 70% refers strictly to foundation pit excavation, not to completion of 70% of the entire data center; nevertheless, it shows that the project has entered the construction phase and can no longer be regarded merely as a conceptual plan.

Substantive progress has also been achieved on the equipment front. Firebird announced in October that it had secured a US license to export 45,000 NVIDIA Grace Blackwell GPUs to Kazakhstan, and an October announcement from supplier NVIDIA provided indirect corroboration.[F][N] This is a step beyond a purchase intent, but the license addresses only regulatory clearance. When the equipment will arrive in batches, how it will clear customs, and how installation and acceptance testing will proceed still depend on subsequent delivery. Equipment deployments in Astana and Almaty mentioned by NVIDIA also cannot serve as a basis for asserting that the data center in Ekibastuz is already operational.

Power Proximity Is Only a Starting Point

Ekibastuz’s proximity to the GRES-1 power plant makes it an attractive location for large-scale loads. KEGOC's annual report shows that power generation at GRES-1 in 2025 increased year-on-year by approximately 2.029 billion kWh, an increase of 9%.[E] This demonstrates the presence of a large-scale power source and a foundation for expanded generation nearby; however, annual generation figures do not directly answer whether the data center can secure its contracted power on an hour-by-hour basis.

A data center needs more than available generation at a power plant: it also depends on the carrying capacity of lines and transformers, switching arrangements during maintenance or outages, and whether power quality and restoration times satisfy customer contracts. Reports mention that a 215 MW substation and external power supply schemes are being coordinated.[Z] This is a critical engineering link for bringing the location's advantages to the facility, but a substation does not represent new power generation, nor can its ultimate capacity, wiring schemes, and redundancy arrangements be inferred solely from power ratings.

Assuming that 1 GW represents the total load of the facilities and that they operate at 1 GW every hour throughout the year, annual electricity consumption would be approximately 8.76 TWh, equivalent to 7.03% of Kazakhstan's national electricity consumption in 2025.[E] This is strictly a hypothetical assumption used to illustrate scale, not a forecast of the project's future electricity consumption. Actual consumption will vary with the deployment of equipment racks, operating workloads, and facility efficiency; if 1 GW refers to IT load, cooling and power distribution losses must also be included when computing total facility electricity use.

In 2025, national electricity consumption exceeded generation by approximately 1.4956 TWh.[E] This represents national annual balance data, not the available capacity of a specific line in Ekibastuz, and it cannot be directly added to future campus loads to calculate a new power deficit. These data highlight the need for grid coordination, but they do not prove that the local site will inevitably face power shortages; similarly, proximity to the power plant does not prove that interconnection projects and backup power supplies require no investment.

Capacity, GPUs, and Commercial Compute Are Not Equivalent

The Phase 1 capacities listed by the government are 68, 164, and 72 MW, totaling 304 MW, scheduled to launch in Q2, Q3, and Q4 of 2027, respectively.[G] These figures provide milestones, but the government briefing does not clearly distinguish between IT load and total facility load including cooling and power distribution.

Firebird stated that its project segment has a potential capacity of 109 MW, alongside plans to deploy over 45,000 GPUs.[F] There is currently insufficient information to determine the relationship between this 109 MW and the 164 MW listed by the government for the second facility. One cannot assume that 109 MW is IT load and 164 MW is total facility load, and then divide the two to calculate the project's PUE. PUE requires comparing total facility energy consumption with IT equipment energy consumption for the same facility during the same period; planned power figures from two different sources do not meet this condition.

Rules related to this news

Even if all equipment arrives on site, the quantity of GPUs does not automatically equate to model training or inference capabilities that are reliably available to customers. Chip models and configurations, cluster interconnects, software stacks, storage, and actual workloads all affect the end result. High-speed internal data center networks allow accelerators to operate together, whereas external fiber bandwidth, routing, and stability determine how customers upload data, access services, and retrieve results; the two cannot replace one another.

Different workloads impose different network requirements. Large-scale training may involve massive data transfers and long-running jobs, whereas real-time inference for end users places greater emphasis on response latency. Existing materials provide no international line data, measured latency, or service level commitments for the campus, so it is premature to claim a low-latency advantage for European or East Asian clients. More valuable subsequent evidence will be equipment load tests, network service conditions, and the first actual paid services.

Cooling also reflects the distinction between planned capacity and actual operational capability. The scheme described in [Z] adopts closed-loop liquid cooling paired with outdoor dry coolers, which is expected to reduce reliance on continuous water consumption for evaporative cooling; however, heat must still be transferred from servers to the outside, and the system design must accommodate conditions such as summer heatwaves. Domestic, fire protection, and cooling loop makeup water each have distinct demands. The planned 2,300 cubic meters per day of water for domestic and supporting needs mentioned in [P] cannot be directly taken as measured cooling water consumption. In the absence of a full-year thermal load model and measured PUE and WUE, it is inappropriate to label the project "zero water consumption," nor is there any basis to assert that it will inevitably suffer derating due to high temperatures.

Contractual and Financial Scopes Must Be Separated

The June government announcement listed two documents of distinct legal natures: a strategic agreement between relevant ministries and Firebird, and a Binding Term Sheet signed between KT-Telecom and Firebird to establish technical and organizational conditions.[J] The title of the latter holds legal significance that cannot be ignored, but because the full agreement has not been made public, it remains impossible to evaluate clause by clause which provisions are binding and what conditions precedent exist for closing.

In August, KAZAKH INVEST reported that ECG inspected the local power plant and the campus, signing an MoU with KT-Telecom to consider developing a data center with approximately 100 MW of IT load and further expansion.[I] This represents an expression of cooperation intent, not a final investment decision or financial close. The expansion potential mentioned therein may also overlap with the campus's overall target and cannot be mechanically added to the stated 1 GW. Other enterprises in the government update are either at the preliminary agreement stage or still in negotiations.[G] Characterizing these figures from disparate stages collectively as "locked-in orders" obscures the work required for a project to move from intent to construction and operations.

Financial figures likewise require delineation. [B] once headlined an article "Investment Halved," but an update at the end of the text recorded that the competent authority corrected the figure back to $10 billion. Firebird itself projects its own investment at $5 billion,[F] which covers a different scope from the figure announced for the entire campus; the numerical difference alone cannot be taken as an inherent contradiction, nor can the figures be directly added together.

The government briefing also noted KPMG's projection of potential annual export revenues of $3.1 billion and an estimated cumulative investment of $115 billion over the data center's 20-year lifecycle.[G] These figures depict forward-looking potential, not contracted service revenue or disbursed funds. Without the original modeling documentation, external observers cannot verify assumptions regarding customer utilization, service pricing, or capital expenditure accounting boundaries, nor can project profitability be derived from them.

Firebird's commitment to allocate at least 10% of its GPU capacity for use by Kazakhstani enterprises and the government provides a clear opportunity for domestic applications; however, pricing, allocation mechanisms, and service level agreements have not been disclosed, so it cannot be construed as a free resource.[F]

Whether the location's energy advantage can be converted into commercially viable compute ultimately depends on whether each part of delivery comes together: power delivered as contracted, equipment installed and verified through load testing, network and software capable of supporting customer workloads, and revenues sufficient to cover long-term operations and hardware refresh cycles. Ekibastuz has taken two tangible steps with on-site construction and equipment export licensing. Going forward, what will be more telling is not which set of capacity figures is larger, but whether these operational conditions can be fulfilled one by one to keep real customer workloads running sustainably.

References