Key Takeaways

  • OpenAI developed Jalapeño, a custom inference chip built with Broadcom that cuts end-to-end latency by 1.7x to 3.6x compared to Nvidia GB200 and GB300 systems.
  • Jalapeño delivers 1.5x to 1.9x more inference throughput per watt, beating Nvidia's current Blackwell and future Rubin architectures on efficiency.
  • OpenAI plans to deploy 10 gigawatts of custom accelerator capacity through 2029, multiplying its active compute capacity by five times.
  • OpenAI and Anthropic have bought and rented tens of thousands of Mac Minis and Mac Studios to run reinforcement learning loops for desktop computer-use agents.

The Inference Pivot: Jalapeño vs. Nvidia

For two years, the AI hardware story was simple: buy every Nvidia GPU you can afford. OpenAI just changed that math.

Hardware analyst Dylan Patel analyzed OpenAI's custom silicon design, codenamed Jalapeño. Developed in partnership with Broadcom, the chip targets dedicated inference rather than broad training workloads. As John Coogan explained, “Dylan Patel says OpenAI's jalapeno is spicy. Usually first generation chips aren't competitive, but OpenAI is beating Nvidia Blackwell and even Reuben.”

The efficiency gains solve OpenAI's largest operational bottleneck: the power grid. Coogan noted that the design delivers “roughly 1.5 to 1.9x more useful inference throughput per watt than Nvidia GB200, GB300 systems while simultaneously cutting endto-end latency 1.7 to 3.6x.” When you run hundreds of millions of user queries every day, a 2x boost in throughput per watt directly cuts datacenter construction costs and power purchase requirements.

This silicon is not a side project. OpenAI mapped out “10 gigawatts of open AI designed accelerator systems manufactured by Broadcom or in partnership with Broadcom deployed from the second half of 2026,” Coogan shared. Jordi Hays pointed out the scale: “put it into context, that's roughly five times more compute than the whole company currently has operational.”

Why Labs Are Buying Thousands of Mac Minis

While OpenAI builds custom datacenter silicon for backend inference, front-end agent training is happening on standard consumer hardware. Both OpenAI and Anthropic have deployed tens of thousands of Apple Mac Minis and Mac Studios inside their clusters.

“OpenAI has reportedly bought tens of thousands of Mac minis and Mac Studios for reinforcement learning and training computer use agents,” Coogan said.

Training an agent to control software requires running authentic operating system sessions. Emulating thousands of desktop environments on standard Linux server instances introduces latency, API quirks, and high virtualization overhead. Mac Minis offer a low-power, self-contained box with unified memory and a native desktop environment. The labs hook these physical machines into reinforcement learning loops where agents practice clicking buttons, playing games, filling out spreadsheets, and debugging applications directly on macOS.

Instead of treating the operating system as an abstract API, the leading labs are treating real personal computers as the training gym for agentic software.

What to Do With This

If you build software agents, stop relying purely on synthetic browser sandboxes. Audit your agent testing pipeline this week. Set up a physical or native OS test environment that matches your target user's desktop, and measure how often your agent fails on real operating system quirks like window management, native file dialogues, and permission popups.