OpenAI's Jalapeño inference ASIC, revealed at Hot Chips, is built first and foremost for efficiency, according to hardware VP Richard Ho. In interviews with Tom's Hardware, Ho said the chip's main goal is to squeeze more useful compute out of power-constrained data centers, not to beat Nvidia on raw benchmarks. Both sources agree that efficiency—not performance—was the driving force behind the design.
Ho stressed that Jalapeño is programmable and general-purpose, pointing to benchmarks showing it running open-source models like DeepSeek R1 and Kimi K2.5, which OpenAI engineers got working in just two months. This was meant to counter the perception that the chip is hard-coded for OpenAI's own models. However, Ho also said the company will have its hands full meeting internal compute demand for "a good long time," leaving the door open to external sales but making clear that is not the near-term plan.
The two sources—a news article and a full interview transcript—are consistent on all key points. They also note that while public benchmarks compared Jalapeño to Nvidia's Blackwell, OpenAI has run internal tests against the newer Vera Rubin platform and says results look strong, though those numbers are not published. For now, the ASIC remains an internal tool aimed at making OpenAI's inference more efficient, with any broader rollout contingent on future supply and demand.```json // Wait, the user asked for strict JSON with exactly those keys. I'll output valid JSON only, no code fences or extra text. Let me redo properly.{