Self-Host Frontier AI: Running Open-Weight Models on AWS

Inside the event

Self-Host Frontier AI: Running Open-Weight Models on AWS

📅 Date: October 16, 2026 1:00 PM PT (LA Tech Week) ⏱️ Duration: 90 minutes 📍 Location: 2450 Colorado Ave, Santa Monica, CA 90404 🅿️ Parking Information below ⬇️

Open-weight models have caught up. GLM, Kimi, DeepSeek, and Gemma now compete with closed APIs on many benchmarks, and you can run them on infrastructure you control. But getting from downloaded weights to real production traffic is a different problem.

How much GPU memory do you actually need? What happens to capacity as conversations get longer? When does self-hosting beat paying per token, and when does it burn money? And how do you run a new model the day it drops, instead of waiting for official platform support?

Join us for a 90-minute program in three parts. AWS opens with a practical guide to choosing, sizing, and paying for open-weight inference. Radix Ark then goes deep on SGLang: bringing up DeepSeek V4.1 on day zero, and serving diffusion models for image and video at high throughput. We close with a panel Q&A.

Agenda

🎙️ Self-Hosting Open-Weight Models on AWS (AWS · 40 min) - Bailey Cho, Solutions Architect at AWS - Augustin Marcin, Solutions Architect at AWS

🎙️From Day-0 to Beyond Day-0: Bringing Up and Optimizing DeepSeek V4.1 in SGLang (RadixArk · 20 min) - Kevin Mi, Member of Technical Staff at RadixArk

🎙️High-Performance Serving Framework for Offline and Real-Time Diffusion Models (RadixArk · 20 min) - Yihao Wang, Member of Technical Staff at RadixArk

🎙️Panel Q&A (10 min) - All Speaker and Nidhi Vedantam, Account Manager at AWS

What You'll Learn

🧭 API or Self-Host, in Plain Terms — What actually changes when you leave a per-token API: who runs which layer, how the bill behaves, and what your team takes on in exchange for control.

🏗️ The GPU Math That Matters — Model size picks the machine. Conversation length × concurrent users decides how many people that machine can serve. On the same server, 4× longer conversations means 4× fewer at once.

🗺️ A Clear Decision Framework — Five lenses (model quality, cost, latency, data control, ops) and six questions that pick your path, from fully managed Amazon Bedrock to SageMaker, HyperPod, and self-managed EKS/EC2.

💰 The Cost Math Your CFO Will Ask About — Real break-even numbers for per-token APIs vs. self-hosted GPUs, why an idle GPU is the most expensive way to buy tokens, and what commitment and utilization do to the price.

🔧 Day-0 Model Support — How to triage a new release in 10 minutes, and how Radix Ark brings up and optimizes a brand-new architecture in SGLang on day zero.

🎨 Serving Diffusion Models — Throughput and latency techniques for image and video generation, offline and real-time.

🌐 Data Sovereignty & Control — When an API with in-Region processing is enough, and when you need to own the weights, the runtime, and the network path.

Models We'll Cover

Real deployment scenarios with today's leading open-weight models, from a model that fits on a single GPU to a 2.8-trillion-parameter model that needs an 8× Blackwell B300 server:

GLM 5.3 • Kimi K3 • Gemma 4 • DeepSeek V4 & V4.1 • MiniMax M3, plus the newest releases (GLM 5.3 Flash, Qwen3.8) and a quick way to evaluate whatever drops next.

Who Should Attend

• CTOs & Founders — Deciding whether to self-host or stay on an API. The first part of the program is built for you. • ML/AI Engineers — Building inference pipelines and choosing serving frameworks • Platform Teams — Evaluating managed vs. self-managed GPU infrastructure • Teams shipping image or video generation • Anyone who's Googled "how much VRAM do I need for [model]" in the last month

Ideal for teams actively evaluating self-hosted in sovereignty, cost control, customization, or all three.

What You'll Walk Away With

✅ A decision framework for choosing the right deployment path for any open model ✅ A one-page comparison sheet: the seven numbers to collect before you self-host ✅ Hardware sizing guidance grounded in published ✅ Cost comparison and break-even math: per-token APIs vs. self-hosted on AWS ✅ Day-0 bring-up and production serving techniques for LLMs and diffusion models ✅ Direct access to AWS Solutions Architects and Radix Ark engineers during the panel Q&A

Hosted By

AWS Startups — Built by Solutions Architects who deploy open-weight models with startups every week. This isn't slides about what's theoretically possible — it's what we ship in production. ━━━━━━━━━━━━━━━━━━━━━━━━━━

🅿️ Parking Information 📍 Location: 2450 Colorado Ave, Santa Monica, CA 90404

Entrance: Enter the parking structure from Cloverfield Blvd. or 26th St., using the far-right lane as you approach the garage. Pull a parking ticket at the gate to enter.

Where to Park: Park on Level P1, in any spot that is not marked reserved.

Validation: Bring your parking ticket to Amazon Reception to receive a validation chaser sticker.

Exiting: At the exit gate, scan your original parking ticket first, then scan the chaser sticker immediately after ("chasing" the original).

━━━━━━━━━━━━━━━━━━━━━━━━━━ This event is a part of #LATechWeek—a week of events hosted by VCs and startups to bring together the tech ecosystem. Learn more at [www.tech-week.com](http://www.tech-week.com).