Home / Private AI / GPU cluster hosting

Need a place to run your GPU cluster?

Haink supplies NVIDIA B300 servers, GB300 NVL72 racks and H200 for your project. If you do not have a site with the power and cooling they need, we can place the cluster with our Tier 1 data-center partner in the United States or Asia-Pacific, in the country you choose, and finance defined projects.

Scope my cluster See available capacity →
22 MWlisted partner capacity in Asia-Pacific, plus US and El Salvador sized per project
Liquid-readyUS, Hong Kong, Kuala Lumpur, Jakarta and Tokyo now; Johor from January 2027
1–2 weeksend-user documentation and supplier review
Financingfor projects with a known end user, use and site

Three questions every cluster project has to answer

Price per server is the easy part. Projects are decided on whether the capacity will really be there, whether the data stays yours, and what the cluster will earn or save.

1. Will the capacity be there?Hardware through allocation to your project, and power, cooling and space in a hall that is energized now or on a known date.
→ Capacity
 2. Does the data stay yours?Hardware you own or finance, in the country you choose, with administrative access and keys that stay with you.
→ Data control
 3. What will the project earn?Owned cost per GPU-hour against what the market charges, for your own product or for capacity you resell.
→ Economics

1. Capacity you can plan around

Hardware through allocationB300 and GB300 NVL72 are supplied through OEM allocation to documented projects. Stock in cluster quantities is close to impossible to find in 2026; a project order is the dependable route, and stock, where it exists at all, costs roughly 1.4–1.7 times as much.
A hall that is readyNew high-density capacity takes a year or more to build. Partner halls listed below are energized now or have a stated ready date, so the site does not become the bottleneck.
One timeline, not threeDocumentation and supplier review take 1–2 weeks; hardware supply, freight, installation and burn-in are planned against the hall's date, so equipment does not arrive before power.

Capacity shown is available as of September 2026 and is reserved for a project once contracted. How supply is committed: How NVIDIA GPU allocation works →

Where your cluster can run

Capacity available with our Tier 1 data-center partner for new projects, September 2026. Hosting is optional: we deliver just as well to your own site or a facility you choose.

LocationCapacityCoolingReady
United StatesSized per projectAir and liquidOn request
El SalvadorSized per projectOn requestOn request
Hong Kong2 MWLiquid-cooling readyNow
Kuala Lumpur, Malaysia2 MW air + 2 MW liquidAir and liquidNow
Johor, Malaysia2 MW air + 2 MW liquidAir and liquidJanuary 2027
Jakarta, Indonesia2 MWLiquid-cooling readyNow
Manila, Philippines2 MWConfirmed per projectNow
Tokyo, Japan3 MWLiquid-cooling readyNow
Osaka, Japan2 MWConfirmed per projectNow
Sydney, Australia3 MWConfirmed per projectNow

For scale: 2 MW holds roughly 120 eight-GPU B300 servers, about 960 GPUs, or about 13 GB300 NVL72 racks at 132 kW each, with headroom for networking and storage. Every destination has its own export and import regime: advanced data-center GPUs for installation in Hong Kong need a US export licence, while a US site needs none. Colocation and energy cost by location: AI server cost →

2. Your data stays yours

You own itThe hardware is yours, or financed with title agreed per project. At the end of a contract it moves with you, subject to export rules for the new destination.
Your jurisdictionYou choose the country your data lives in. For a US company a US site keeps data at home; for regulated industries in Asia the hall can sit in the country the regulator expects.
Your accessAdministrative access, encryption keys and the software stack stay with you. The facility provides power, cooling, physical security and remote hands. Blackwell GPUs add hardware confidential computing.
Tier 1 facilityThe cluster runs with a Tier 1 data-center provider, with the physical security, access control and operating standards that come with it.

How this compares with public cloud: Private AI vs cloud AI →

3. What the project earns

Two ways a dedicated cluster pays: it replaces rented GPUs for your own product, or it sells capacity to others. Both come down to utilization.

Owned cost per GPU-hour3 years, 70% used3 years, 90% used5 years, 70% usedMarket rental, Sept 2026
B300$5.69$4.43$4.59$6.80 blended index
H200$3.35$2.60$2.75$3.30 index · $4.62 settled trades
For your own productAn owned B300 costs less per GPU-hour than renting at every utilization shown, by roughly 15–35% over three years and more over five. H200 matches the lowest index rate at 70% and beats what capacity actually trades at.
For capacity you resellOne eight-GPU B300 server at the $6.80 index brings in about $334k a year at 70% utilization and $429k at 90%, against an all-in cost of about $244k a year over five years. The margin depends on contracted utilization, so we plan it against a named offtaker, not a hope.

Owned cost: server at the middle of its range, plus 15% for fabric and storage, plus colocation, energy and support, less conservative resale value. Rental rates: Silicon Data index 24 September 2026, Ornn Data H200 benchmark 27 September 2026. Indicative, not a quote. Run your own numbers in the cluster cost calculator →

What is included

LayerWhat we deliver
ComputeHGX B300 servers or GB300 NVL72 racks; H200 servers or H200 NVL nodes where they fit the workload
Fabric and storage800G InfiniBand or Spectrum-X, all-flash or parallel storage sized to the GPU count
Site (optional)Colocation with our Tier 1 partner: space, contracted power, cooling and remote hands
DeliveryEnd-user documentation, export and import paperwork, freight and customs
Bring-upInstallation, burn-in and acceptance before handover
PaymentPurchase, or financing that can cover hardware and colocation together

From first call to first token

1. ScopeWorkload, model sizes, location and target date. Days to two weeks.
2. DocumentationEnd-user statement and supplier review. 1–2 weeks.
3. SupplyHardware through allocation, planned against the hall's ready date.
4. DeliveryFreight, customs and delivery to the partner hall. 1–2 weeks.
5. Bring-upInstall, burn-in and acceptance. 1–3 weeks at 8–64 GPUs.

Several stages run in parallel. The detailed breakdown by cluster size: GPU cluster deployment timeline →

Frequently asked questions

Is the capacity reserved for us?

Yes, once contracted. The capacity listed is available at the date shown; power, space and cooling are reserved for your project in the colocation contract.

Do we have to host the cluster with you?

No. Hosting is an option. We deliver the cluster to your own site or any facility you choose; the Tier 1 partner capacity is there if you need a place with the power and cooling a B300 cluster requires.

Who owns the hardware?

You do, or the financing structure agreed for the project holds title until it is paid. Either way you are the end user of record, and the same end-user review applies.

Can you finance a GPU cluster?

Yes, for defined projects: a known end user, a defined use and a defined site. The structure is agreed per project and can cover hardware and colocation together.

Which GPUs do you deploy?

NVIDIA B300 servers and GB300 NVL72 racks for new clusters, and H200 where a lower entry cost fits the workload. B200 and H100 are supplied on request. See B300 vs H200 vs B200 vs H100.

Can the cluster go anywhere?

Every destination has its own export and import rules, and all orders go through end-user screening. A US site needs no export licence; advanced GPUs for Hong Kong or Mainland China need a US export licence.

AI server cost 2026 → B300 vs H200 comparison → Cluster bill of materials → Private AI infrastructure → GPU allocation →

Planning a cluster?

Tell us the workload, where it should run and when. We come back with the platform, a costed bill of materials, a partner hall with capacity, a realistic timeline and the financing options.

Scope my cluster

Not ready yet? Ask a question