The AI Inference Economy
AI Inference Economy @ Brazil — GPU Access and the Real Cost of Inference
A private, in-person evening at Oracle for early-stage technical founders building AI products.
A short, direct conversation about the least glamorous part of building with AI: getting compute, understanding what it actually costs, and making architecture decisions knowing that inference economics can kill a product long before it finds product-market fit.
Capacity is limited to 30 founders.
Applications are required.
Agenda
💡 6:30 PM — Oracle: Inference in Production - What we see from the infra side (Amanda Machado, AI Solution Engineer @ Oracle);
💬 7:00 PM — Canastra Talks: Inference Economics from First Principles. (Larissa Bomfim, MP @ Canastra, talks with Thiago Tergolino, MS in Math from IMPA and Co-Founder & CEO @ stealth mode);
🥂 7:30 PM — Networking Session. Drinks, bites, and conversations about the AI inference economy with tech founders, researchers, and AI providers.
Who This Is For
This evening is for early-stage founders and technical leads who:
Are running AI workloads in production, or about to
Write the code and own the architecture decisions
Want to understand inference economics before it becomes a margin problem
Are facing real cost, latency, and reliability tradeoffs
Prefer an honest technical conversation to a panel of institutional talking points
The Theme
"You have the money and still can't get the GPU." People have been saying that in San Francisco for months now, and it's started showing up here.
The shortage stopped being general and became specific. The bottleneck now sits in top-end parts, where advanced packaging and high-bandwidth memory are the gating input, and hyperscalers pre-committed to that supply years in advance. Power has become a harder constraint than silicon: a cluster's timeline is now set by its grid interconnection date, not by chip delivery.
On the other side of the same equation sits inference. Plenty of teams ship the product, put it in production, and only then discover that cost per call eats the entire margin. Running that math beforehand is not trivial, and almost nobody talks about it publicly.
So that's the conversation. Small room, people who've already run into this, no pitching.
Hosts
*By registering, you agree to share your data with Canastra Ventures, sponsors and our partners. Additionally, by registering, you authorize us to send you the latest news, company announcements, and other communications from Canastra Ventures and partners.