KONST
AI Training CalculatorContact Us
EN
Glows.aiSERVICE | GPU cloud

Sign Up and Launch,
Rent GPUs by the Second

Glows.ai is KONST's GPU cloud platform partner, with compute from data centers KONST builds and operates under contract. Sign up online and pick an instance type to launch an instance with the OS, drivers, and frameworks ready, so there is no data center or environment to build. Instances are billed per second and stop billing at shutdown, with no monthly fee or minimum usage.

Minimum billing unit1secondBilling stops at shutdown, no monthly fee
SHOWCASE | One-click images and app demos

Open-Source Models and Tools That Run at Boot

Popular open-source models and inference frameworks are packaged as official images across four workloads: language models, image generation, speech synthesis, and agents. The OS, drivers, and dependencies are ready in advance, and tested configurations are available for everything from single-GPU experiments to multi-node deployments.

Three-dimensional render made of social media platform icons
Video generation

Produce Stylized Short Videos at Scale With the Open-Source Wan 2.2 Model

Create stylized short videos with the open-source Wan 2.2 video model and use LoRA to keep characters consistent. Output has no vendor watermark and no daily generation limit.

Wan 2.2 · ComfyUI · LoRARead the deployment guide →
An iris flower on a blue background
Video generation

Turn Photos Into Short Videos With Wan 2.2 Image-to-Video

Load the Wan 2.2 image-to-video template in the preinstalled ComfyUI image and upload a photo to generate a clip. A single RTX 4090 renders a 5-second 720p video in under 9 minutes.

Wan 2.2 I2V · RTX 4090Read the deployment guide →
ComfyUI generating images with a dual-LoRA workflow on a GPU instance
Image generation

Move ComfyUI Workflows to a Cloud RTX 4090

The preconfigured image boots in 30 to 60 seconds, and an RTX 4090 generates a 1024×1024 SDXL image in about 3 to 8 seconds. Custom nodes, models, and workflows can all move to the cloud.

ComfyUI · SDXL · FLUXRead the deployment guide →
OpenClaw calling a local model on an instance through Telegram
Agent

Run an OpenClaw Messaging Agent on a Local Model

The official preinstalled OpenClaw image includes Ollama for running local models. Connect it to messaging apps such as Telegram and Slack so the agent can run commands and scheduled tasks.

OpenClaw · OllamaRead the deployment guide →
Close-up of a chip on a circuit board
Large language models

Self-Host DeepSeek V4 on a 48 GB GPU

V4 Flash has open weights and a context window of one million tokens. The quantized version needs about 48 GB of GPU memory, so pick a 48 GB instance type and launch it in one click to self-host.

DeepSeek V4 · 48 GB GPURead the deployment guide →
GPUStack dashboard showing GPU utilization and deployed models
Self-hosted model serving

Self-Host Multi-GPU model Serving With GPUStack

Self-host a chat interface and multi-GPU model serving on an instance. Models are stored on Datadrive, so the data moves with you when you change instances.

Ollama · OpenWebUI · GPUStackRead the deployment guide →
Chat screen of gpt-oss-20b replying in OpenWebUI
Large language models

Deploy the GPT-OSS API on an RTX 4090

Set up the environment on an RTX 4090 instance, call the API, and go live with Auto Deploy. Llama 3.1 also has a quick-start image.

RTX 4090 · Auto DeployRead the deployment guide →
Waveform and playback interface after BreezyVoice speech synthesis
Speech synthesis

Synthesize a Realistic Personal Voice With BreezyVoice

Generate personalized voice files with BreezyVoice. MimicTalk text-to-speech also has a quick-start image.

BreezyVoice · MimicTalkRead the deployment guide →
Mesh Clustering joining two RTX 4090 instances into one network group
Large language models

Deploy DeepSeek-R1 Across Multiple Nodes and GPUs With SGLang

Deploy DeepSeek-R1 with SGLang across multiple nodes and GPUs. A single-node version can also be launched in one click.

SGLang · vLLMRead the deployment guide →
  • PyTorch
  • TensorFlow
  • JAX
  • Keras
  • Llama
  • Mistral
  • Stable Diffusion
  • FLUX
  • PaddlePaddle
  • Horovod
PRODUCTS | Product lines

Training, Inference, Data, Collaboration:
Four Product Lines, One Account

From single-GPU experiments to multi-node training, and from shared endpoints to dedicated inference, Glows.ai puts compute, data, and team workflows on one platform. Choose the product that fits your workload.

Training

On-Demand Cloud · Clusters

  • On-demand cloudSingle to multi-GPU instances with preloaded framework images, ready in under two minutes
  • Virtualized on-demand clusterMultiple nodes pooled as one resource, with the flexibility of on-demand billing
  • Bare-metal clusterDedicated multi-node H100 and A100 with direct access to hardware and high-speed interconnect
Inference

Cloud Inference · Dedicated Inference

  • Cloud inferenceModels deployed as HTTP endpoints that scale with request volume, and you pay only for compute time
  • Dedicated cloud inferenceReserved fixed resources with predictable performance, for latency-sensitive and regulated industries
Data

Datadrive · Data Engine

  • DatadrivePersistent distributed storage mounted at /datadrive, so datasets and weights carry across instances
  • StorageHigh-performance storage for training data, checkpoints, and model output
  • AI-powered data engineFrom data import and preprocessing to feeding training jobs
Collaboration and apps

Team Edition · NeuroBase · Matrix0

  • Team EditionGPUs, environments, datasets, and permissions in one shared workspace
  • NeuroBaseVisual agent workflows and automation
  • Matrix0A serverless runtime that uses SQL for analytics, feature engineering, vector search, and agent workflows
SOLUTION | KONST's approach

Set Up GPUs by Use: Development, Team Sharing, Online Inference

On-demand cloud launches single or multi-GPU instances for development and short training runs. Virtualized clusters pool nodes into a resource shared by the team. Inference comes in shared and dedicated versions for latency-sensitive online services.

Interactive Development and Short Training Runs

You need an experiment environment you can start and stop at will: launch single or multi-GPU instances on demand and run as soon as you pick an instance type.

Long Training Runs Shared by a Team

Several people compete for the same GPUs: pool multiple nodes into an adjustable resource that the team can allocate flexibly.

Online Inference

Latency matters and traffic cannot stop: the shared option bills per second, and the dedicated option reserves fixed resources.

FEATURE | Service scope and specifications

Launch and Start Training Right Away

No data center or environment to build. Pick an instance type and run, with per-second billing that stops at shutdown.

Ready on Launch

The OS, drivers, and frameworks are ready. Sign up, pick an instance type, and start running.

Per-Second Billing

No monthly fee and no minimum usage commitment. Billing stops at shutdown.

No Rebuild When Changing Instance Types

Data is mounted at launch and kept after shutdown, and snapshots can be restored to a new instance.

FEATURE | Service scope and specifications

Data Persists, and Changing Instance Types Needs No Environment Rebuild

Data and environments outlive the instance, so you can pick up where you left off after changing instance types, an interruption, or a restart. Models and private data are stored separately.

Data Persists After Shutdown

Data is mounted at launch and kept after shutdown, so there is no need to upload it again.

Environment Snapshot Save and Restore

Snapshots save the whole environment, so you can continue right away after changing instance types or an interruption.

Public Access · Private Storage

Public models and datasets are available to use directly, private data is stored in your account space, and a high-performance storage service is available for large volumes.

FEATURE | Service scope and specifications

Self-Service Online Launch, Billed by Actual Usage

Sign up online to pick instance types and launch instances yourself, with no contract. Billing is per second and stops at shutdown, with no monthly fee or minimum usage. Top up first and pay as you go, and check usage and spending in real time.

1secondMinimum billing unit, billing stops at shutdown
0monthly feeNo monthly fee, billing stops at shutdown, no minimum usage commitment
BENEFIT | Benefits

Switch to a High-End GPU and
Training Time Falls From 8.5 Hours to 70 Minutes

The unit price is higher, but run time drops sharply, so total per-second spending goes down and the time saved goes straight into the next experiment.

Training taskGPU classRun time
  1. Image classificationEntry-level workstation graphics card8.5 hours
  2. Same taskSame-generation high-end accelerator70 minutes
  3. Object detectionTop-tier data center accelerator25 minutes
FEATURE | Available instance types

From RTX 4090 to H100,
Across Three Availability Zones in Taiwan

The table lists instance types, memory, and availability zones in the on-demand catalog, all in Taiwan TW-02, TW-03, and TW-04. Bulk multi-node H100 and A100 go through bare-metal clusters.

GPU modelGPU memoryvCPUSystem memoryAvailability zoneBest for
NVIDIA H100 80GB HBM380 GB16224 GBTW-02Large model fine-tuning, high-throughput inference
NVIDIA RTX PRO 600096 GB1448 GBTW-04Large-memory inference, image generation
NVIDIA A100 SXM480 GB1680 GBTW-04Training and fine-tuning
NVIDIA L40S PCIe ×848 GB per GPU1601,792 GBTW-04Multi-GPU parallel training
NVIDIA L40S PCIe48 GB1280 GBTW-04Inference, image and video generation
NVIDIA RTX 6000 Ada48 GB1280 GBTW-04Development and mid-size inference
NVIDIA GeForce RTX 409024 GB1050 to 64 GBTW-03 · TW-04Experiments, ComfyUI, 7B to 20B models

Specs from glows.ai/llms-full.txt (retrieved 2026-10-06). Availability changes over time; see the pricing page for list prices of each instance type.

PRICE | Pricing and fees

Centralized Team Top-Ups, With Credits Allocated by Role

One account manages multiple members. Credits are topped up centrally and then allocated, so members do not pay separately. Administrators can limit which instance types and images members can use, so credits are not spent on high-end GPUs.

Creator1personMember limit

PermissionsCreates and owns the team, with full permissions

AdministratorUp to2peopleMember limit

PermissionsCreates accounts, allocates credits, controls resource permissions, and releases instances

UserUp to100peopleMember limit

PermissionsLaunches and uses instances within assigned credits and resource permissions

EVIDENCE | Trust and partnerships

Security Certification and Vendor Partners,
Built With the Hardware Ecosystem

The platform is certified to ISO/IEC 27001:2022 for information security management, uses NVIDIA-certified hardware, and is built with Taiwanese hardware and system makers.

27001ISO/IEC2022 information security management certification
3availability zonesTaiwan TW-02, TW-03, TW-04
24/7supportEnterprise-grade round-the-clock expert support
2minutes or lessTo launch a single-GPU environment with preloaded frameworks
Partners
  • Advantech
  • GIGABYTE
  • Regulus Technology
  • USPACE
  • Cxit
  • EZ
  • Ense
  • Barking Dog
  • Liger Kernel
FAQ | Common questions

FAQ

How are Glows.ai GPU cloud fees calculated?
Pricing is transparent, with H100, H200, and RTX5090 available, per-second billing, and auto-scaling. Contact our sales team for details.
Are there special plans for schools or research institutions?
Yes. Credits, instance types, and teaching support in education plans are set according to the size of the institution. Contact the sales team to arrange one.
Do I have to rebuild my environment if I change instance types?
No rebuild is needed. Data is mounted when the instance launches and kept after shutdown, so you do not have to upload it again each time. A snapshot saves the full state of the environment and restores it to a new instance, so after you change instance types or after an interruption you can pick up your work directly.
Which GPU should I use for small models and for large models?
It depends on the size of the model and the data. For experimental small-model training, a workstation-class graphics card is enough. Large models with many parameters and multi-GPU parallel training need the memory capacity and interconnect bandwidth of high-end servers.

Ready to Launch Your First GPU Instance?

Tell us your team size and use case, and we will reply with the suitable plan and credit allocation