MacanCompute
Get a quote

Private AI inference

Private AI compute. Rent it or own it.

Run open-weight models on hardware reserved for you. Rent a dedicated machine hosted in Indonesia, or put a Macan M-1 in your own server room. Same API either way.

Get a quoteSee the server
Macan M-1 server: a black tower with a perforated front, three orange slashes, and a glass side showing four GPUs.
GPU memory per machine
32–192 GB
API on every machine
OpenAI-compatible
Hosted and supported
Indonesia

Rent or own

Two ways in. One API.

Start by renting, move to owning, or run both. Your applications call the same endpoint.

Rent

A dedicated machine in Indonesia

A whole machine, reserved for your organisation and hosted by Macan. No shared GPUs. No queue behind someone else’s job.

  • Solo, Duo, Quad, or Pro configuration
  • Hosted and operated by Macan
  • Price and term by quote
See compute

Own

A Macan M-1 on your premises

A server built to order and delivered to your site. Your models and your data stay inside the building.

  • One to four GPUs, up to 192 GB of GPU memory
  • Open models installed and tested before delivery
  • Price by quote
See the server

  1. Your data stays put

    Prompts, documents, and outputs stay on hardware reserved for you. Nothing passes through a shared model provider.

  2. No per-token billing

    You pay for the machine, not for each request. Usage can grow without the invoice changing shape.

  3. Open models, standard API

    Qwen, Gemma, DeepSeek, GLM, and other open-weight models behind an OpenAI-compatible API. Most existing code needs only a new base URL.

  4. We run the plumbing

    Macan installs the models, sets up the endpoint, and keeps the machine healthy. Your team builds the product.

Industries

Built for work that cannot leave the room.

Where data is sensitive or the site is remote, a private machine is the simple answer. Some jobs teams start with:

Health

  • Summarise clinical notes and discharge letters
  • Search internal guidelines in plain language
  • Draft patient explanations for staff review

Typical setupAn M-1 Duo on site.

Finance

  • Review loan and onboarding documents
  • Answer staff questions from policy manuals
  • Flag unusual wording in contracts

Typical setupAn M-1 Pro on premises for larger models.

Mining

  • Search maintenance logs and equipment manuals
  • Summarise shift and incident reports
  • Assist engineers at sites with a weak link

Typical setupAn M-1 Solo or Duo at site. It can run with no internet link.

Energy

  • Query operating procedures and standards
  • Summarise inspection and outage reports
  • Draft work orders from field notes

Typical setupAn M-1 Duo on site, or a rented Quad for central teams.

Government

  • Sort and route incoming correspondence
  • Search archives and records in Indonesian
  • Draft replies for officer review

Typical setupAn M-1 Quad or Pro in the agency’s own data room.

Manufacturing

  • Answer operator questions from machine manuals
  • Summarise quality and downtime reports
  • Translate technical documents

Typical setupAn M-1 Solo on the factory network to start.

Compute · Rent

Rent a dedicated machine.

Pick a configuration. We host it in Indonesia, install the models you choose, and hand you an endpoint. The hardware is yours alone for the term.

MachineGPUsGPU memoryHosted inPrice
Solo1 × RTX 509032 GBIndonesiaBy quote
Duo2 × RTX 509064 GBIndonesiaBy quote
Quad4 × RTX 5090128 GBIndonesiaBy quote
Pro2 × RTX PRO 6000 Blackwell192 GBIndonesiaBy quote

The GPUs inside

GPUMemoryMemory bandwidth
RTX 509032 GB1,792 GB/s
RTX PRO 6000 Blackwell96 GB1,792 GB/s

GPU specifications as published by NVIDIA.

Server · Own

Macan M-1. The same machine, in your building.

A tower built to order around NVIDIA Blackwell-generation GPUs. It arrives with models installed and an endpoint ready on your network.

Front of the Macan M-1: perforated panel with three orange slashes and an orange edge light.
VariantGPUsGPU memoryPrice
Solo1 × RTX 509032 GBBy quote
Duo2 × RTX 509064 GBBy quote
Quad4 × RTX 5090128 GBBy quote
Pro2 × RTX PRO 6000 Blackwell192 GBBy quote

What you get

Built to order
Solo, Duo, Quad, or Pro.
Tested before it ships
Each unit runs a burn-in under load.
Ready to use
Open models installed, endpoint configured.
Warranty
Confirmed in your quote
Delivery and install
Confirmed in your quote
Get an M-1 quote
The four Macan M-1 variants side by side: Solo with one GPU, Duo with two, Quad with four, and Pro with two larger GPUs.
SoloDuoQuadPro
Side of the Macan M-1: four GPUs stacked behind a glass panel, each with an orange light strip.

Images are product renders. Final hardware may differ in detail.

Models

Will it fit?

Pick a machine and a precision. The chart estimates the GPU memory each open model needs and shows whether it fits.

Machine
Precision

7 of 13 models fit on Solo (1 × RTX 5090, 32 GB) at 4-bit.

ModelParametersNeedsResult
Qwen3 8BQwen8.2B4.9 GBFits
Gemma 3 12BGemma12.2B7.3 GBFits
gpt-oss 20Bgpt-oss20.9B12.5 GBFits
Mistral Small 3.2 24BMistral24.0B14.4 GBFits
Gemma 3 27BGemma27.4B16.4 GBFits
Qwen3 32BQwen32.8B19.7 GBFits
DeepSeek R1 Distill 32BDeepSeek32.8B19.7 GBFits
Llama 3.3 70BLlama70.6B42.4 GBDoes not fit
GLM-4.5 AirGLM110.5B66.3 GBDoes not fit
gpt-oss 120Bgpt-oss116.8B70.1 GBDoes not fit
Qwen3 235BQwen235.1B141.1 GBDoes not fit
GLM-4.5GLM358.3B215.0 GBDoes not fit
DeepSeek V3DeepSeek684.5B410.7 GBDoes not fit

Estimate: parameters × bytes per weight × 1.2.

  • Long contexts and many simultaneous users need more memory than this estimate.
  • 3-bit quantisation saves memory but can reduce answer quality.
  • Parameter counts are taken from each publisher’s model card.

How it works

From first call to first token.

  1. Define the job

    Tell us what the model must do, how many people will use it, and where the data must stay.

  2. Test first

    Try your workload before you commit, so you see real speed and quality on real hardware.

  3. Build and burn-in

    We assemble or allocate your machine, install the models, and run it under load until it is stable.

  4. Go live

    Your endpoint goes live on your site or hosted in Indonesia. Macan stays on hand for upkeep.

Quote

Tell us about the job.

A few details are enough. We reply by email with a configuration and a price.

What happens next

  1. We read your request and check which machine fits the job.
  2. You get a configuration and a price by email.
  3. If it helps, we set up a call to go through the details.

I want to

We use these details only to answer your request.

FAQ

Questions, answered.

What is Macan Compute?

Private AI inference from Macan Teknologi Perkasa. Your organisation gets its own machine for running open-weight AI models: rented and hosted in Indonesia, or bought and installed on your premises.

Should we rent or own?

Rent when you want to start quickly or are still sizing the workload. Own when data must stay inside your building or you expect steady use for years. Both expose the same API, so you can move from one to the other.

Which models can we run?

Open-weight models such as Qwen, Gemma, DeepSeek, GLM, Llama, and Mistral. The fit chart shows which models fit on which machine. Models offered only as a hosted service cannot be installed.

Where does our data go?

It stays on the machine reserved for you. With an M-1 on your premises, requests do not leave your network.

How does pricing work?

By quote. The price depends on the configuration, the term, and the support you need, and hardware prices move. We would rather send an accurate number than publish a stale one.

How is an M-1 delivered and installed?

Each M-1 is built to order. Delivery, installation, and lead time depend on your location and are confirmed in your quote.

Can we try before we commit?

Yes. Describe the workload in the quote form and we will propose a way to test it before you sign for a term or a purchase.

Who supports the machine?

Macan does. We install the models, set up the endpoint, and handle upkeep. Support terms are set out in your quote.