OpenAI's first custom chip Jalapeño: 9-month tape-out, AI-accelerated design
thehiphopswami · x · 2026-09-22
Richard Ho, VP of Hardware at OpenAI, details Jalapeño, the company's first custom AI accelerator, on The Data Exchange podcast.
Key points:
- Goal: control more of the hardware stack to lower inference costs and boost performance
- Unusual combination of high throughput and low latency; blank-slate architecture bringing memory and compute closer together
- The chip taped out in just nine months, with AI tools compressing parts of the development cycle and dramatically boosting expert team productivity
- Covers whether CUDA is still a moat, Triton, speculative decoding, RL, and agents driving inference demand
- Roadmap from gen one to gen three with HBM4 and new memory; power, not chips, is becoming the data center bottleneck
Related event: OpenAI Hardware VP Details Jalapeño, Its First In-House AI Chip(2 posts)→
More from Companies & People
- VentureTwins maps the B2A economy: dozens of agent infra and app companies in first series installment — venturetwins · 2026-11-12
- How DoorDash's GenAI platform team lets engineers build with AI quickly and safely — charles_irl · 2026-11-12
- Databricks launches autonomous AI coworker Genie: 91% on enterprise tasks with half the tokens — Yuchenj_UW · 2026-11-12
- Anthropic Lead SWE makes ~10x a GS-13's pay — called a national security blunder — harris_edouard · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- Is Jensen Huang's bad behavior more commonplace than it seems? TechCrunch investigates — TechCrunch AI · 2026-11-12