All the focus around Jalapeño seems centered on performance vis-a-vis Nvidia, but the engineering velocity behind it is truly remarkable. Some highlights from the @hotchipsorg talk: 1. Building custom chips used to require massive armies of engineers and multi-year timelines. OpenAI went from initial RTL code to tapeout in just 9 months using XLS (a Rust-like language for hw from Google) and AI-driven co-design. 2. An AI loop took a raw, barely functional (DeepSeek) kernel running at 0.31% efficiency and autonomously optimized it to 88.94% of the chip's physical limit in under 2 days. 3. The AI-generated code ran up to 1.8x faster on critical blocks than implementations hand-tuned by world-class kernel engineers. 4. AI did more than just write software for the chip. It optimized the physical hardware circuits before tapeout. Again, this isn't zero-sum or a net-negative for Nvidia. Custom ASICs unlock absurd serving efficiencies for specialized workloads and will expand the total compute pie for the entire industry.