OpenAI's Jalapeño: The First Punch at Nvidia
OpenAI just showed off its first custom chip. It's called Jalapeño, built with Broadcom, and it's aimed at one thing only: running models, not training them. Inference. The part where you actually talk to ChatGPT or hand a task to an agent.
The number that matters is cost. Broadcom CEO Hock Tan says it's running roughly 50% cheaper than the AI GPUs everyone uses today, with better performance per watt. And they did it fast. From first design to tape-out in nine months, which both companies are calling maybe the quickest serious ASIC cycle anyone's pulled off. First chips deploy by the end of 2026.
Why does an editor covering agents care about a chip? Because the whole agent economy runs on inference cost. Every agent that loops, every sub-agent you spawn, every long-horizon task that burns ten thousand tokens of reasoning, all of it is inference. If OpenAI cuts that bill in half on its own silicon, the math on what's worth automating changes overnight. Cheaper inference isn't a hardware story, it's an agent story.
And read the timing. Same day Qualcomm went after Nvidia's software moat by buying Modular. OpenAI is going after the hardware. Two punches at the same jaw in 24 hours. The most valuable company in the world has spent a year being everyone's landlord, and everyone's now trying to stop paying rent. Details at https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/
← Back to all articles
The number that matters is cost. Broadcom CEO Hock Tan says it's running roughly 50% cheaper than the AI GPUs everyone uses today, with better performance per watt. And they did it fast. From first design to tape-out in nine months, which both companies are calling maybe the quickest serious ASIC cycle anyone's pulled off. First chips deploy by the end of 2026.
Why does an editor covering agents care about a chip? Because the whole agent economy runs on inference cost. Every agent that loops, every sub-agent you spawn, every long-horizon task that burns ten thousand tokens of reasoning, all of it is inference. If OpenAI cuts that bill in half on its own silicon, the math on what's worth automating changes overnight. Cheaper inference isn't a hardware story, it's an agent story.
And read the timing. Same day Qualcomm went after Nvidia's software moat by buying Modular. OpenAI is going after the hardware. Two punches at the same jaw in 24 hours. The most valuable company in the world has spent a year being everyone's landlord, and everyone's now trying to stop paying rent. Details at https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/
Comments