Placing a GPU cluster in Kazakhstan as colocation means fixing kilowatts per rack, heat removal, power redundancy and connectivity before anyone talks about empty floor. Typical industry ranges below are not Akashi ratings.
What the design already says about density, cooling and reservation is in AI and GPU in Kazakhstan. This article is the next step: how to plan the placement, from a kilowatt estimate to the choice between a standard cabinet and a high-density module.
How many kilowatts per rack does a GPU cluster need?
There is no universal “kW per GPU rack”, and a published range does not replace the nameplate of the hardware you will install.
For ordinary CPU racks, the colocation guide gives 5–10 kW. Dense configurations and mixed CPU/GPU loads sit at 15–20 kW and above, paired with a cooling scheme: hot aisle or liquid. That is a way to compare halls. It is not an Akashi rack rating.
Density then follows the node. The guide GPU colocation for AI describes a typical eight-GPU server — an NVIDIA HGX node with H100 or H200 accelerators — as 6–10 kW. A rack of those nodes is described at 40 kW and 80 kW. Rack-scale systems of the GB200 NVL72 class are about 120–130 kW per rack. Those are typical equipment figures, not an Akashi rating, and that 120–130 kW figure is not a density for the Astana site. The exact density for the chosen phase is confirmed when the configuration is sized.
Air or liquid?
Every kilowatt drawn at the rack has to leave the room. In the same guide, air cooling — moving cold air through the hall — runs out of capacity around 20–30 kW per rack. Above that typical line the guide names liquid: a rear-door heat exchanger, direct-to-chip cold plates, or immersion. A hall that only offers air cannot host a dense GPU cluster, however many rack units are free. That is an industry frame.
Akashi’s wording describes the design, not a handed-over hall. DC-4 engineering provides for liquid cooling of high-density AI racks. The first phase launches in May 2027, so the liquid loop is not described as already running.
Astana’s climate, in the same description, allows free cooling for most of the season. Free-cooling and hybrid modes are designed to cut energy use by up to 40% against a conventional scheme. The target PUE is below 1.4. Management systems for ISO 50001 and ISO 14001 are being prepared; that is not a certificate of the built hall. How those targets are worded is in the piece on green engineering.
If the calculated rack fits what the chosen path cools with air, liquid is not required. If the figure is above the typical 20–30 kW per rack, ask which liquid scheme is in the design and which module it is tied to. At Akashi that tie is stated for DC-4.
Power redundancy and operational risk
A training cluster sits at full power for hours and days, so a power or cooling failure mid-run throws away accelerator time already paid for. Redundancy is what happens when one element fails without warning.
In the Uptime Institute classification, Tier III is concurrently maintainable: an element can be maintained without stopping IT. Tier IV is fault tolerant: an unplanned failure of any single element does not affect IT. Calculated availability of a Tier IV topology is 99.995%, about 26 minutes a year. That is a property of the topology, not a measured outage of a hall. What enters a colocation contract is covered in Tier III or Tier IV and on the Tier IV page.
Those sentences must not collapse into “Akashi is already Tier IV certified”.
TCDD is Tier Certification of Design Documents. The design documents of the first phase (DC-1, AKASHI Data Center 1, Phase 1) hold Uptime TCDD Tier IV. On the construction timeline this is a completed step in January 2026: the design is certified, not the built hall.
TCCF is Tier Certification of Constructed Facility. It is a separate on-site audit: inspectors test the constructed systems, not the drawing set. The audit is planned for April 2027 and is not a certificate held today. TCDD is not TCCF. TCDD without TCCF means the first-phase documents are confirmed, and the building is not.
The campus power design names the following:
- two independent feeds, active/active, with 110 kV redundancy;
- dual electrical redundancy per building;
- UPS at 5/4N, with up to 10 minutes of battery reserve;
- backup generators with 12 hours of autonomy at full load;
- two fuel suppliers, with a 4-hour delivery SLA.
Main equipment of the 110/10 kV substation is installed, with a dedicated power line in place. The campus description says the facility uses the reserve provided for it, rather than electricity meant for the city’s current consumption. Technical conditions allow capacity to grow as later phases come online. This is a campus under construction.
Three risks follow. Engineering: fault tolerance is confirmed as design documents (TCDD), and the built hall has not passed TCCF. The first phase launches in May 2027, so a long run is planned against that date, not against the TCDD date. Commercial: 5.54 MW is reserved against the 5.28 MW design IT load of DC-1, so capacity is locked by a reservation. Thermal: a rack in the typical band of 40 kW and above is not acceptable on its own. It is checked against the design density of the specific module, and the exact figure for that phase is confirmed when the configuration is sized.
Connectivity, latency and paths for AI and HPC
Inside the cluster, nodes exchange gradients constantly. The GPU colocation guide describes that as low latency and high bandwidth between nodes: InfiniBand, or Ethernet with RDMA, at 200–800 Gbit/s. That is a typical fabric requirement, not a statement that Akashi has this interconnect. The customer designs the fabric; the facility provides placement, power, cooling and outside entries.
Outside, the path to the user decides. The guide How to choose a data center calls 10–20 ms a comfortable threshold for most business applications, and it says to calculate the actual path. That is a typical planning band, not an Akashi measurement and not a site SLA.
The campus architecture provides four independent optical entries and connections to several Kazakhstani and international carriers. The site is carrier-neutral: the customer chooses partners and routes for fault tolerance, bandwidth and latency. No fixed carrier list is published. Four entries reduce the risk that one cut cable isolates the hall. They do not replace a latency figure to a specific exchange, or the fabric between GPU nodes.
A standard cabinet and a high-density module
These are different arrangements. Cabinet rental is a placement format: 1200 mm wide, coded access, dual PDUs, from 1 to 1,000+ cabinets. Floor loading is 2 t/m², and the raised floor is 1.6 m. A non-standard cabinet goes in a cage. Cooling built to the customer’s own requirements is a dedicated room. Design density is stated separately for each module.
The module figures come from the campus description. They are design figures, not a measured density of a handed-over hall. The actual rack count depends on the customer and on power per rack. The layout is in the piece on the Astana campus.
| Module | Design IT capacity | Per rack | How it is worded |
|---|---|---|---|
| DC-1 | 5.28 MW | up to 15 kW | first phase, design supply |
| DC-2 | 12 MW | up to 30 kW | campus module, design supply |
| DC-3 | 12 MW | up to 30 kW | campus module, design supply |
| DC-4 | up to 30 MW | up to 150 kW | designed for high-density AI and HPC, with liquid cooling |
The exact per-rack density for a given phase is confirmed when the configuration is sized. It is not inferred from the design figure of a neighbouring module.
Design capacity of the whole complex after all phases is up to 100 MW: a figure for the project as a whole, not the IT load of a hall in operation, and not a sum of the module IT figures. The full-campus rack plan is up to 4,224; the actual count depends on the customer and on kilowatts per rack. Expansion to 100 MW and 4,224 racks sits at 2027+ on the construction timeline, meaning 2027 and later. DC-4 is intended as the most powerful building on the campus, not as a hall that already accepts 150 kW.
A configuration estimate and a capacity reservation
Sales needs kilowatts per rack and rack count, air or liquid, a 1200 mm cabinet or a non-standard size, mass against 2 t/m², how many independent paths, and the date relative to May 2027. The exact per-rack density for the chosen phase is confirmed in that estimate.
The request goes on the contacts page. Sales replies within one business day, runs a technical pre-qualification and a capacity check, and sends a proposal. Directly: sales.department@akashi.cloud, +7 700 810 3010.
A reservation is discussed now. With 5.54 MW reserved against the 5.28 MW design IT load of DC-1, waiting for a free rack in May 2027 means competing for that power. The same request should not mix TCDD, which applies to the design documents, with TCCF, which the built hall does not hold.
Frequently asked questions
How many kilowatts per rack does a GPU cluster typically need? These are typical ranges, not Akashi ratings. CPU racks are often 5–10 kW, and dense or mixed CPU/GPU loads are described at 15–20 kW and above. One eight-GPU node is described at 6–10 kW, a rack of such nodes at 40 kW and 80 kW, and rack-scale systems of the GB200 NVL72 class at about 120–130 kW per rack.
When is air cooling no longer enough? In the typical industry description, air cooling runs out around 20–30 kW per rack; above that, the same guide names liquid cooling. At Akashi, liquid cooling is in the DC-4 engineering design and is stated for high-density configurations. It is a design property: the first phase launches in May 2027, and the hall has not been handed over.
How does a standard cabinet differ from a high-density module? A cabinet is a rental format: 1200 mm wide, coded access, dual PDUs. Per-rack density in the campus description is stated per module, and those are design figures: DC-1 up to 15 kW, DC-2 and DC-3 up to 30 kW, and DC-4 is designed for up to 150 kW for high-density AI and HPC, with liquid cooling. The exact density for a given phase is confirmed when the configuration is sized.
Does TCDD mean the built hall is already Tier IV certified? No. TCDD (Tier Certification of Design Documents) certifies the design documents of phase one, DC-1. Uptime Institute lists AKASHI Data Center 1, Phase 1 with Tier IV Certification of Design Documents. TCCF (Tier Certification of Constructed Facility) is a separate on-site audit of the built facility, planned for April 2027. TCDD is not TCCF.
How do I request a configuration estimate and reserve capacity? Send a request on the contacts page with kilowatts per rack, rack count, cooling scheme and the date. Sales replies within one business day and checks available capacity. The first phase launches in May 2027. Reservation against the 5.28 MW design IT load of DC-1 stands at 5.54 MW. The exact per-rack density for the chosen phase is confirmed in that estimate.
Need a configuration estimate for a GPU cluster? Send a request to sales — we will go through density, cooling and a capacity reservation.