Today, OpenAI took to X to show the fruits of its in-house chip effort. The new OpenAI Jalapeño Intelligence Platform is built with Broadcom in a big win for its design side. OpenAI is seeing massive demand for model serving, so it is investing in building its own silicon, much like we have seen traditional hyper-scalers do in the past.
OpenAI Jalapeño Intelligence Platform Shown Powered by Broadcom #
Here is an interesting one. OpenAI said that using its model, it was able to tape out in just 9 months when designing its own accelerator.
If you look closely, at the photo of Sam Altman and Hock Tan, you can see an accelerator with HBM stacks on either side. Between the wafer shot and the package, it looks like there are six spaces for HBM, but that might be off. OpenAI says that the silicon is already running GPT-5.3-Codex-Spark in the lab. They also disclosed that they are using Broadcom Tomahawk for their networking silicon and that Celestica is providing the board and rack-scale infrastructure.
Final Words #
OpenAI said that it will go into the architectural details and performance over the coming months. That makes this a bit of a stay-tuned announcement. Still, we are in an age where, instead of largely seeing startups build next-gen NVIDIA competitors, it is really large companies, hyper-scalers, and large-scale AI serving companies. What a fun time in the industry.
In other news, I am at Qualcomm’s Investor Day 2026, where the company said it is going to roll out its investor plans. Full disclosure that Qualcomm paid for some of my travel. We will have more from the company later today as the first big announcement dropped in the form of Qualcomm’s acquisition of Modular to bolster its software platforms. More to come on STH later today and the Substack.