·
AI Infrastructure & Inference
Woosuk Kwon
Co-founder & CTO |Inferact
Co-created vLLM, the most adopted open-source LLM inference engine (2,000+ contributors). Invented PagedAttention for efficient KV-cache management, reducing GPU memory waste from 60-80% to under 4%. Ph.D. from UC Berkeley under Ion Stoica. Inferact launched Jan 2026 with $150M seed at $800M valuation (a16z, Lightspeed, Sequoia).
GitHub
Open Source Projects
4 repositories · 195 total stars
Languages
Tech Stack
Python75%Shell25%
Episodes
Podcast Appearances
2#
Title
Date
Dur.
Research Papers
3 papersBiography
Deep Research
Woosuk Kwon is the co-founder and CTO of Inferact, an AI inference infrastructure startup that launched in January 2026 with $150M in seed funding at an $800M valuation led by Andreessen Horowitz and Lightspeed. He co-created vLLM, the most widely adopted open-source LLM inference engine with over 2,000 contributors, and invented PagedAttention, an algorithm inspired by OS virtual memory that eliminates 60-80% of GPU memory waste in KV caches. Kwon earned dual B.S. degrees in Computer Science and Mathematical Sciences from Seoul National University (ranked 1st of 134 students) and a Ph.D. in Computer Science from UC Berkeley under Ion Stoica with a 4.0 GPA, completing his dissertation in December 2025. Before founding Inferact, he was a Research Scientist at Google DeepMind (2024-2025) and a Member of Technical Staff at Thinking Machines Lab (May-Nov 2025).
LLM Inference OptimizationKV Cache Memory ManagementPagedAttentionGPU SchedulingSpeculative DecodingModel PruningDistributed SystemsOpen Source AI InfrastructureServerless AI ServingSystems for Machine Learning
Timeline
1 Paper1 total
Key Contributions
vLLM
The most widely adopted open-source LLM inference and serving engine, achieving 24x throughput over HuggingFace Transformers and 3.5x over TGI. Over 2,000 contributors and used by major AI labs globally.
PagedAttention
Novel attention algorithm inspired by OS virtual memory paging that partitions KV caches into non-contiguous blocks, reducing memory waste from 60-80% to under 4% and enabling copy-on-write memory sharing across requests.
Inferact
AI inference infrastructure company commercializing vLLM with a serverless managed offering, observability, and multi-hardware support. Raised $150M seed at $800M valuation from a16z, Lightspeed, Sequoia, and Databricks.
Jenga (SOSP 2025)
Memory management system for serving LLMs with heterogeneous hardware, extending PagedAttention principles to mixed GPU environments.
SkyPilot (NSDI 2023)
Intercloud broker for sky computing that optimizes workload placement across multiple cloud providers for cost and performance.
Learned Token Pruning (KDD 2022)
Method for pruning tokens in Transformer models to accelerate inference while maintaining accuracy.
Nimble (NeurIPS 2020)
Lightweight and parallel GPU task scheduling framework for deep learning workloads.
Notable Quotes
“
We see a future where serving AI becomes effortless. Today, deploying a frontier model at scale requires a dedicated infrastructure team. Tomorrow, it should be as simple as spinning up a serverless database. The complexity doesn't disappear; it gets absorbed into the infrastructure we're building.
“
vLLM taught me a tough lesson: to keep the GPU fully utilized, we need to pay close attention to everything happening on the CPU.
“
vLLM improves the throughput of popular LLMs by 2-4x with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca.
12 sources(click to expand)
Woosuk Kwon - Personal WebsiteWoosukKwon - GitHub ProfileEfficient Memory Management for LLM Serving with PagedAttention (arXiv)vLLM: Easy, Fast, and Cheap LLM Serving with PagedAttention (blog)vLLM GitHub RepositoryInferact launches with $150M funding to commercialize vLLM (SiliconANGLE)a16z announces $150M seed for Inferact (X)Woosuk Kwon | Sequoia CapitalvLLM: An Efficient Inference Engine for LLMs - Ph.D. Dissertation (UC Berkeley)Woosuk Kwon - DBLP BibliographyPyTorch Conference 2024 - vLLM TalkUC Berkeley Sky Computing Lab - vLLM Project Page
Research generated March 19, 2026
AI Infrastructure & Inference/Woosuk Kwon
All Profiles