AI Inference & LLM
Deploy large language models, chatbots, AI agents, embeddings, and RAG applications with GPU acceleration.
NVIDIA Blackwell GPU Hosting
Kilat GPU.
Make Your AI #JadiKILAT.
High-performance GPU computing powered by the NVIDIA RTX PRO 6000 Blackwell Server Edition. 96 GB of GDDR7 ECC GPU memory for AI inference, fine-tuning, generative AI, computer vision, rendering, and CUDA-accelerated workloads.
✓ 96 GB GDDR7 ECC
✓ NVIDIA Blackwell
✓ Data Center in Indonesia
✓ Dedicated & Multi-GPU
96 GBGDDR7 ECC GPU Memory
24,064CUDA Parallel Processing Cores
4 PFLOPSFP4 Tensor Core Performance
1,597 GB/sMemory Bandwidth
AI & Accelerated Computing
One GPU. Many Possibilities.
Run modern AI workloads with large GPU memory without building and operating your own GPU servers.
Fine-Tuning
Fine-tune models for business and domain-specific use cases with 96 GB of VRAM on a single GPU.
Generative AI
Generate images, video, audio, and creative content using modern generative AI models.
Computer Vision
Run object detection, image recognition, OCR, video analytics, and high-speed vision inference.
Rendering & 3D
Accelerate rendering, 3D visualization, simulation, digital twins, and NVIDIA Omniverse workflows.
Video Processing
GPU-accelerated encoding, decoding, transcoding, streaming, and AI-powered media processing.
NVIDIA Blackwell
96 GB VRAM for serious workloads.
RTX PRO 6000 Blackwell Server Edition combines AI compute, CUDA, RTX, and large GPU memory in a single data-center accelerator.
CPU, RAM, NVMe storage, network, and server configuration can be tailored to your workload.
GPUNVIDIA RTX PRO 6000 Blackwell Server Edition
ArchitectureNVIDIA Blackwell
GPU Memory96 GB GDDR7 ECC
Memory Interface512-bit
CUDA Cores24,064
Tensor Cores752 (5th Gen)
FP4 Tensor Core4 PFLOPS*
FP8 Tensor Core2 PFLOPS
FP16 / BF16 Tensor Core1 PFLOP
TF32 Tensor Core234 TFLOPS
FP32120 TFLOPS
Memory Bandwidth1,597 GB/s
InterfacePCI Express Gen 5 ×16
Media Engines4× NVENC, 4× NVDEC
* Effective FP4 performance with sparsity, per NVIDIA's published specifications. All GPU figures are peak values from the NVIDIA RTX PRO 6000 Blackwell Server Edition datasheet.
Sizing Guide
What can you run?
Use the model weights as a starting point for sizing. Every row below runs on a single 96 GB Kilat GPU Dedicated.
| Model | Precision | Approx. weights |
|---|---|---|
| Whisper large-v3 | FP16 | ~3 GB |
| Llama 3.1 8B | FP16 | ~16 GB |
| Gemma 2 27B | FP8 | ~27 GB |
| Qwen2.5 32B | FP8 | ~32 GB |
| FLUX.1 dev + T5 encoder | BF16 | ~33 GB |
| Llama 3.3 70B | FP4 (NVFP4) | ~35 GB |
| Llama 3.3 70B | FP8 | ~70 GB |
| Qwen2.5 72B | FP8 | ~72 GB |
Weight sizes are calculated from published parameter counts and are indicative, not benchmarked. Plan for an additional 15–30% of VRAM for the KV cache, activations, and runtime overhead — more for long context windows or large batch sizes. Not sure how your model will fit? Send us your model and target throughput and we'll size it with you.
Dedicated GPU
Kilat GPU Dedicated.
One tenant, one GPU. Every Kilat GPU server allocates the full RTX PRO 6000 Blackwell to your workload — nothing shared, nothing partitioned.
Kilat GPU Dedicated
96 GB
Full RTX PRO 6000 Blackwell
The entire NVIDIA RTX PRO 6000 Blackwell Server Edition is allocated to your workload. Maximize compute, VRAM, and bandwidth without sharing the GPU.
Dedicated physical GPU
Full 96 GB GDDR7 ECC memory
Full GPU compute resources
Custom CPU, RAM & NVMe
Linux + SSH & CUDA environment
Kilat team technical support
Best for: larger models, fine-tuning, high-throughput inference, rendering, computer vision, and production workloads that need a full GPU.
Pricing
Custom quote
Based on server configuration, GPU count, and term.
Request Kilat GPU Dedicated
Why Kilat GPU?
Powerful GPU. Local support.
Indonesia Data Center
Run workloads closer to your data, applications, and users in Indonesia.
96 GB GPU Memory
Large VRAM for AI models and workloads that are difficult to run on smaller-memory GPUs.
No Hardware to Manage
No need to invest in GPUs, servers, power, cooling, and data-center operations.
Flexible Configuration
CPU, RAM, NVMe, OS, and server configuration can be tailored to your needs.
Operated by CloudKilat
ISO/IEC 27001
ISO 9001:2015
ISO 14001
ISO/IEC 20000-1
Get Started Easily
From workload to a ready-to-use GPU.
Tell us your workload
Tell us your AI model, application, framework, VRAM, storage, and intended usage.
Choose the configuration
The Kilat team helps configure the right GPU, CPU, RAM, NVMe, network, and environment.
Deploy & run
Access your server and deploy your models, containers, or CUDA workloads.
FAQ
Questions about Kilat GPU.
What is Kilat GPU?
Kilat GPU is CloudKilat's GPU computing service for AI, machine learning, rendering, video processing, and GPU-intensive workloads powered by the NVIDIA RTX PRO 6000 Blackwell Server Edition.
How much GPU memory is available?
Each RTX PRO 6000 Blackwell Server Edition provides 96 GB of GDDR7 GPU memory with ECC, and every Kilat GPU Dedicated server allocates all of it to a single customer.
Can I use it for LLMs?
Yes. Kilat GPU is suitable for LLM inference, AI agents, RAG, embeddings, and fine-tuning compatible with the NVIDIA CUDA ecosystem. See the sizing guide above for how models fit in 96 GB.
Can I use Docker?
Yes. Linux servers can be configured for GPU containers and tools such as Docker, NVIDIA Container Toolkit, PyTorch, vLLM, Ollama, and other CUDA frameworks.
Is multi-GPU Dedicated available?
Yes. Multi-GPU configurations can be provided based on requirements and capacity availability. Talk to the Kilat team to discuss the right architecture.
Got an AI workload? Talk to the Kilat team.
Tell us about your model, application, and capacity requirements. We'll help prepare the right GPU configuration.