Daily Alpha · Substack
· Premarket Alpha · by Buzzberg Research
OpenAI's Jalapeño is a credible custom inference chip, but it reinforces HBM and leading-edge foundry demand rather than removing them.
Themes on this desk
Custom inference silicon
Tae Kim's interview with OpenAI's Richard Ho argues Jalapeño's HBM-per-core design and AI-assisted layout deliver better performance per watt than Nvidia Blackwell on several open-source models, with HBM4 and leading-edge foundry capacity s
OpenAI's Jalapeño beats Nvidia Blackwell on performance per watt in cited benchmarks.
Tae Kim argues OpenAI's Jalapeño is the most promising non-GPU inference chip because its HBM-per-core architecture and AI-assisted design deliver high utilization and performance per watt versus Nvidia Blackwell on several open-source models. If deployed at volume, it could lower inference costs and power use while reducing reliance on merchant GPUs, reinforcing HBM and leading-edge foundry demand.
It supports custom ASIC competition while keeping HBM and leading-edge foundry suppliers in the value chain.
Watch Production deployment, HBM4 orders and independent benchmark replication.
Source →