Foundation Model Developers
Build and train large-scale AI models, with an insatiable demand for massive GPU capacity across data centers to support intensive training and continuous model improvements
Sea Cloud 101
Team Sea Cloud
AI Is Now Part of Our Everyday Life
01
02
03
04
AI Products Are Powered by AI Models
LANGUAGE
GPT
Claude
Gemini
Llama
IMAGE
Midjourney
DALL·E
FLUX
AUDIO
ElevenLabs
Whisper
VIDEO
Sora
Veo
GPUs Power Modern AI Training and Inference
Larger models and heavier usage typically require more compute resources
GPUs power most AI training and inference workloads
Compute capacity has become critical infrastructure and an essential resource for the AI era
Who Needs GPU Compute?
AI Research Institutes & Labs
Conduct cutting-edge research and scientific experiments using AI models, with dedicated GPU resources to support complex simulations, data analysis, and demanding computational workloads
Tech Companies & AI Startups
Integrate AI capabilities into their products and services by calling model APIs, delivering innovative experiences that meet evolving user needs without managing their own GPU infrastructure
From Hardware to Compute
Designed for AI Training and Inference
Massive Parallelism
Thousands of cores work together to handle many calculation tasks at the same time
Tensor Cores
Specialized processors for matrix operations make AI training and inference much faster
High-Bandwidth Memory
Ultra-fast memory keeps GPUs supplied with data for faster computing
Primary Uses
TRAINING
GPUs process training data in parallel, calculate errors, and adjust model weights
INFERENCE
GPUs use trained model weights to process new inputs and generate predictions or content
Popular AI GPUs
A100
AmpereThe pioneer of modern AI computing
H100
HopperTransformer Engine revolutionizes AI training
B200
BlackwellFP4 unlocks a new era of AI inference
B300
Blackwell UltraUltra-scale memory and next-level AI performance
How GPUs Become Servers
A GPU Alone Cannot Run AI Workloads
A GPU needs a complete system around it. One or more GPUs work with CPUs, memory, storage and networking to form a GPU server that runs AI training or inference
GPU
A powerful chip capable of massive parallel computing.
GPU Server
A complete system that enables GPUs to run AI workloads.
GPU
Powers AI computations
CPU
Manages system operations
Memory
Fast access to working data
Storage
Stores models and datasets
Networking
Connects servers for data exchange
Rack-Scale System
Connects multiple GPU servers into a high-speed cluster to deliver greater compute power.
GB300 NVL72
72 Blackwell Ultra GPUs
36 Grace CPUs
Vera Rubin NVL72
72 Rubin GPUs
36 Vera CPUs
Data Centers: The Home of Servers
Two Ways to Access GPU Compute
Maximum Control
Bare Metal
Get exclusive physical servers with direct hardware control. No VM virtualization can mean lower latency, higher throughput and better utilization when well tuned.
Full Hardware & Software Control
BIOS & NUMAOS & AI FrameworksDrivers & CUDAGPU, NVMe & Network
MINIMAL OVERHEAD
Cloud Platform Services
Access GPU compute through virtual machines or hosted model APIs, without managing underlying physical hardware.
GPU Virtual Machines
Your cloud GPU computer
No Hardware Hassles
Model APIs
Ready-to-use model outputs
Develop directly with model outputs
More Infra ControlHigher flexibility & customization
Less Infra to ManageFaster and easier to use
Bare Metal: Maximum Performance & Control
Sea-Managed Infrastructure
Co-Managed Infrastructure
Cloud Platform Services: Easy Access to Compute
GPU Virtual Machines
A virtual computer in the cloud with allocated GPU resources, allowing users to train and fine-tune models, run inference workloads, and develop AI applications.
Cloud
Platform
Provision
Configure
Manage
GPU VM
GPU VM
GPU VM
GPU VM
Virtualization Layer
GPU Server PoolA GPU server can be virtualized into one or multiple GPU VMs
Model Inference APIs
Cloud-based model inference service, allowing users to access open-source models, reserve dedicated compute, or deploy their own models.
Cloud
Platform
Model Catalog
Inference
BYOM
Serverless Inference
Use model inference APIs instantly on demand
Dedicated Inference
Use model inference APIs with reserved GPU capacity for guaranteed throughput
Sea’s Ambition: Becoming AI-Native
Over the past decade, these three pillars have built Sea’s strengths
- User Scale
- Hundreds of millions of users across our products
- Business & Data
- Diverse business scenarios with rich real-world data
- Technical Expertise
- Strong engineering and infrastructure capabilities
Together, these strengths provide real-world use cases and a solid foundation
for Sea’s AI development, adoption and continuous improvement.
Sea's AI Journey: Our Three-Layer Foundation
Along our journey, we have built three layers of AI capabilities
Platforms
Empower teams to build AI agents and workflows with internal platforms that connect models, data and tools
Models
Post-train open-source and develop proprietary models to improve business outcomes, while optimizing inference efficiency
Infrastructure
Build and operate large-scale AI infrastructure, optimizing GPU allocation and workload scheduling for reliable, efficient compute
The three layers work together to empower our people, product and service with AI, and deliver real business impact