Appearance
Computing Rental
Use computing power like electricity — light assets, high elasticity, zero barriers
In the early stages of AI projects, many enterprises face a practical dilemma: the business direction is not yet fully validated, model performance is uncertain, but hardware procurement costs hundreds of thousands or even millions, plus data center, electricity, and operations — a huge investment. If you buy a batch of servers first, and then the project direction changes or the business doesn't take off, the hardware becomes sunk cost; if you only use public cloud, the long-term costs are staggering, and data security is hard to guarantee.
Magicsoft Computing Rental Services are designed to solve this dilemma. We provide high-performance GPU computing to enterprises in a "service-oriented" way — no hardware procurement needed, no self-built data centers, no operations teams required. You only need to apply on demand, pay as you go, and stop when you want, just like water and electricity.
■ Product Positioning
Creating a "Use Computing Like Electricity" Service Model
Enabling enterprises to acquire computing resources on demand, achieving low-cost, high-efficiency, and zero-burden computing usage.
🎯 Value Proposition in One Sentence:
Transform capital expenditure (CAPEX) into operational expenditure (OPEX), turning computing from a "heavy asset" into a "light service."
Computing rental is not simple "cloud server leasing," but an exclusive service deeply optimized by Magicsoft based on its powerful computing infrastructure and scheduling platform for AI training, inference, Web3 nodes, and other scenarios. Unlike general cloud vendors, we better understand the workload characteristics of AI and Web3: pre-configured mainstream frameworks, optimized distributed communication, and visual task management — truly enabling enterprises to "use it immediately and use it efficiently."
■ Core Product Portfolio
1) Elastic GPU Computing Services
Product Description:
Users can apply for GPU resources on demand through the Magicsoft computing platform, supporting resource startup speeds from seconds to minutes. The platform provides multiple GPU type options (A100 / H100 / RTX 4090, etc.), meeting different computing needs from inference to training.
| Specification Type | Applicable Scenarios | Startup Speed |
|---|---|---|
| Lightweight (Single RTX 4090) | Model inference, small-scale validation | < 30 seconds |
| Standard (Single A100) | Fine-tuning, medium training | < 1 minute |
| High-Performance (4×A100) | Large-scale training | < 2 minutes |
| Flagship (8×H100) | Large model pre-training | < 3 minutes |
👉 Problems Solved:
- Slow computing resource acquisition → Compressed from weeks of traditional procurement to minutes
- Inflexible scaling → Supports manual or automatic elastic scaling, one-click expansion for business peaks
Imagine this scenario: You're training a model and suddenly realize you need more cards to accelerate convergence. In traditional mode, you'd need to reapply for resources, wait for approval, and reconfigure the environment — at least half a day gone. But on the Magicsoft Computing Rental platform, you just slide the slider on the console from 4 cards to 8 cards, and the system automatically allocates new resources and joins the training task. The entire process takes less than 5 minutes, and training isn't interrupted. This is the true meaning of "elasticity."
2) Computing Scheduling and Management Platform
Product Description:
Magicsoft provides a one-stop computing management platform (MagicCompute Portal), where users can complete all operations including resource application, task submission, monitoring, and cost analysis through a visual console. The platform supports real-time resource monitoring, including key metrics such as GPU utilization, memory usage, temperature, and network traffic.
| Function Module | Description | User Value |
|---|---|---|
| Resource Dashboard | Total GPU cards, used/idle, quotas | Clear at a glance, no resource waste |
| Task Management | Submit, stop, priority, queue | Flexible control of training tasks |
| Real-time Monitoring | GPU/CPU/Memory/Network charts | Timely detection of performance bottlenecks |
| Cost Center | Statistics by project/user | Refined cost allocation |
| Alert Configuration | GPU temperature too high, task anomalies | 7×24 automatic notifications |
👉 Problems Solved:
- Unable to unify computing resource management → One platform manages all, with clear isolation for multi-tenant and multi-project
- Low usage efficiency → Real-time monitoring + scheduling optimization, GPU utilization improved by 30% on average
Many enterprises, after renting cloud GPUs, simply log in to machines via SSH and manually run scripts, completely unaware of how much resources are used, which tasks are running, or whether there are idle cards. Magicsoft's management platform completely changes this "black box" state. You can see the status of every card, set up task queues to automatically queue training tasks, and generate monthly cost reports for various business departments. The management platform is not just a tool, but the foundation of enterprise computing governance.
3) Multi-Billing Models
Product Description:
Magicsoft provides flexible billing methods to adapt to different enterprise usage habits and budget structures. Users can choose to pay by actual usage duration, purchase package plans for better unit prices, or reserve exclusive resource pools for large customers.
| Billing Model | Billing Granularity | Suitable Scenarios | Unit Price Advantage |
|---|---|---|---|
| Hourly Billing | Minute-level (less than 1 hour counts as 1 hour) | Short-term validation, temporary tasks | No commitment, lowest barrier |
| Daily Billing | 24-hour units | Continuous training tasks for several days | 20% discount compared to hourly |
| Monthly Package | 30-day fixed resources | Long-term stable training | 50% discount compared to hourly |
| Exclusive Resource Pool | Reserve N cards, pay monthly | Large customers, multi-team sharing | Customizable discounts |
👉 Problems Solved:
- Uncontrollable computing costs → Multiple billing combinations, precisely matching business rhythm
- Serious resource waste → Release after use, idle cost drops to zero
We encountered an AI medical startup that needed to train models once a week, each time for about 20 hours. If using hourly billing, that's about 80 hours per month with controllable costs; but if they mistakenly reserved monthly resources, they would spend 3 times more. Magicsoft's billing model design allows customers to choose the optimal plan based on their business's real rhythm. Moreover, the platform provides cost suggestions: for example, if the system detects you've used more than 12 hours daily for the past week, it will automatically remind you that "monthly packages are more cost-effective." This intelligent cost management is rarely seen among general cloud vendors.
4) Pre-configured Development Environment
Product Description:
The computing rental service pre-integrates mainstream AI frameworks and development tools. Users can start Jupyter Notebook or VS Code remote development environments with one click without manually installing drivers, CUDA, or Python environments. API-based computing invocation is also supported, making it easy to integrate into CI/CD workflows.
Pre-configured Environment Content:
- Base Environment: Ubuntu 22.04 + CUDA 12.1 + cuDNN 8.9
- AI Frameworks: TensorFlow 2.15 / PyTorch 2.1 / JAX 0.4 / PaddlePaddle 2.5
- Development Tools: Jupyter Lab / VS Code Server / SSH
- Additional Components: Git / Docker / Conda / MLflow / Weights & Biases
👉 Problems Solved:
- Complex environment deployment → Out-of-the-box, from login to training start < 5 minutes
- Long development cycles → No need to debug driver version conflicts, focus on algorithms rather than operations
For data scientists and algorithm engineers, the most painful thing is not designing models, but configuring environments. "CUDA version incompatibility," "PyTorch and TensorFlow conflicts," "missing some system library"... These trivial issues often consume 30% of development time. Magicsoft's pre-configured environment installs and tests all common frameworks and dependencies in advance, and supports user-defined custom images — you can package your company's internal private libraries and configurations into images that automatically load on each startup. This way, new employees can start running models on their first day, instead of spending two days installing drivers.
5) Distributed Training Support
Product Description:
For large-scale training tasks, single-machine computing power is often insufficient. Magicsoft's computing rental platform natively supports multi-machine, multi-card distributed training, automatically handling task splitting, node communication, and gradient aggregation. Users only need to specify the number of nodes required, and the platform handles the underlying scheduling.
Distributed Training Workflow Diagram:
User submits training script (specifying node count)
↓
Platform automatically allocates N GPU nodes
↓
Initialize distributed communication (NCCL / RPC)
↓
Data parallelism / Model parallelism automatic splitting
↓
Training execution + real-time log aggregation
↓
Training complete, resources automatically released👉 Problems Solved:
- Low training efficiency → Multi-machine linear acceleration, 64-card cluster speedup ratio > 0.85
- Inability to fully utilize resources → Task queuing + automatic scheduling, idle card utilization < 5%
The barrier to distributed training is actually quite high: you need to configure SSH passwordless login, set environment variables, synchronize code and data, handle node failures... Magicsoft encapsulates all this complexity. You only need to write training scripts like single-machine code (PyTorch DDP or TensorFlow MirroredStrategy), and the platform will automatically recognize and launch distributed tasks. If a node crashes during training, the platform automatically removes that node and recovers from the nearest checkpoint, without you needing to resubmit the task. This "zero-modification distributed training" experience is one of our core competitive advantages.
■ Usage Mode (Three Steps to Get Started)
- Step 1: Register and recharge
- Step 2: Select GPU specifications and duration
- Step 3: One-click environment startup, begin training
The entire process is fully self-service and completed online, without sales involvement or signing paper contracts. We provide trial quotas — new users receive 10 hours of free computing (A100 cards) upon first registration, allowing you to experience at zero cost.
■ Typical Application Scenarios
✔ AI Enterprises
- Model training and fine-tuning (LLM / multimodal / vertical industries)
- AI product development and testing (inference service stress testing, A/B testing)
- Short-term computing burst needs (competitions, paper experiments, quarterly reports)
An AI voice company has several "sprint periods" each year, such as participating in international evaluation competitions or completing model optimization before client demos. During these periods, computing demand may be 5 times normal. If they purchased equipment themselves, most resources would be idle normally; if they used public cloud entirely, costs would be too high. Magicsoft Computing Rental allows them to elastically scale during sprints while keeping only basic resources normally, perfectly matching business peaks and valleys.
✔ Startup Teams
- Low-cost validation of AI products (rapid MVP launch)
- Iterate models after receiving user feedback
- Consider self-built computing after securing funding
We served a 3-person AI startup team that only had tens of thousands of yuan in initial funding. Buying an A100 server would cost most of their budget and take two weeks to arrive. They chose Magicsoft Computing Rental, paying just a few yuan per hour, totaling less than 2000 yuan per month, and successfully got their product prototype running, securing angel round investment. This is the value of computing rental for startups: preventing good ideas from being killed by expensive hardware.
✔ Web3 Projects
- Computing network deployment (Filecoin storage nodes, Livepeer video transcoding)
- Node operation and testing (public chain validation nodes, testnets)
- Short-term computing needs (airdrop activities, stress testing)
Web3 project teams often need to deploy large numbers of nodes temporarily for airdrop activities or stress testing, with nodes no longer needed after activities end. Magicsoft Computing Rental can launch hundreds of instances at once, releasing them with one click after activities end, at costs far lower than long-term hardware ownership.
✔ Enterprise Business
- Peak period computing resource supplementation (Double 11 recommendation systems, financial report season risk control)
- Data analysis and batch processing tasks (ETL, data cleaning, feature engineering)
An e-commerce platform needed an additional 100 A100 cards during Double 11 to support real-time recommendation and advertising bidding model inference. Their own data center only had 50 cards, with a gap of 50 cards. Magicsoft Computing Rental provided elastic resources, interconnecting with the local cluster through dedicated lines to form a unified computing pool, automatically releasing after Double 11 ended — perfectly solving the "only busy a few days a year" computing challenge.
■ Core Value (Why Choose Magicsoft Computing Rental)
| Value Dimension | Specific Benefits |
|---|---|
| Zero Investment Startup | No hardware procurement, no data center, no operations team required |
| Ultimate Flexibility | Use on demand, stop when finished, scale from 1 to 1000 cards in minutes |
| Optimal Cost | Avoid resource idle waste, comprehensive cost 40%~70% lower than self-built (short-term scenarios) |
| Rapid Launch | From registration to running first training task < 10 minutes |
| Professional Optimization | Pre-configured AI/Web3 environment + distributed training support, efficiency far exceeds general cloud |
Computing rental is not Magicsoft's "low-end alternative," but a core product line running parallel with computing hardware. It serves enterprise needs at different stages and scenarios. We even support a "rent first, buy later" model: after validating model effects during the rental period and stabilizing business, you can convert the rental configuration directly into a hardware procurement plan, with previous rental fees partially deductible from the purchase price. This flexibility is unique in the market.
■ Overall Computing Product Advantages (Computing Hardware + Computing Rental Dual-Wheel Drive)
Magicsoft's computing system is not a single product, but a complete ecosystem covering the full enterprise lifecycle.
Stage Evolution Path:
- Startup Phase (validating ideas) → Computing Rental (low cost, zero barrier)
- Growth Phase (stable business) → Hybrid Architecture (rental + some self-built)
- Maturity Phase (large-scale production) → Self-built Computing Center (hardware procurement + private deployment)
1) Complete Computing Ecosystem Closed Loop
From hardware → cloud computing → platform → applications, forming an end-to-end capability system. Enterprises at any stage can find matching solutions at Magicsoft, with smooth migration without starting from scratch.
2) Support Multiple Deployment Modes
| Deployment Mode | Description | Applicable Scenarios |
|---|---|---|
| On-Premise Private | Hardware deployed in enterprise data center | High compliance, long-term stable needs |
| Public Cloud Computing | Magicsoft managed platform | Elastic needs, short-term projects |
| Hybrid Cloud Architecture | On-premise + cloud unified scheduling | Balancing security and elasticity |
3) Cover Full Enterprise Lifecycle
| Stage | Recommended Solution | Characteristics |
|---|---|---|
| Startup Phase | Computing Rental | Zero assets, low risk |
| Growth Phase | Hybrid Architecture | Core business self-built, elastic part rented |
| Maturity Phase | Self-built Computing Center + Rental Supplement | Optimal TCO while retaining elasticity |
4) For AI + Web3 Dual Scenarios
Whether training large models or running blockchain nodes, Magicsoft's computing system has been deeply optimized, with performance far exceeding general clouds.
5) Standardization + Rapid Delivery
Combined with Magicsoft's "30-day launch" system (one week for requirements, one week for design, one week for development, one week for launch), computing resources can be ready simultaneously with project initiation, not delaying business progress.
■ Frequently Asked Questions (FAQs)
| Question | Answer |
|---|---|
| Is there any performance difference between rented and self-built computing? | No difference. We use the same hardware as self-built solutions, and the network is specially optimized, with distributed training performance even better than ordinary self-built environments. |
| How is data security guaranteed? | Each rental instance is network isolated; user data does not persist by default (optional persistent storage available). After tasks complete, instances are destroyed and data cannot be recovered. Supports encrypted transmission. |
| Can environments be retained long-term? | Yes. You can save configurations as custom images for direct use next time. Persistent storage can also retain data. |
| Is API invocation supported? | Yes. The platform provides REST API that can be integrated into your CI/CD or automation workflows. |
| What are the advantages compared to AWS/Azure? | ① Pre-configured AI/Web3 environment, out-of-the-box; ② Deep optimization for distributed training; ③ Lower cost (20%~40% cheaper for equivalent configurations); ④ Flexible business models such as "rent first, buy later." |
■ Next Steps (CTA)
📌 Experience Magicsoft Computing Rental Now:
- ✅ New users receive 10 hours of free A100 computing
- ✅ Self-service activation through online console, no manual review required
- ✅ Scan QR code to join technical exchange group for best practices
👉 Let computing become your business's "tap water" — turn on when needed, turn off when not.
Computing rental is not a temporary measure, but an advanced computing consumption model. Magicsoft makes high-performance computing within reach through productization, platformization, and servitization. Whether you are an AI entrepreneur, Web3 developer, or enterprise IT leader, you can find value suited to you from computing rental. Welcome to experience immediately.