OpenAI built its own AI chip to cut its dependence on Nvidia
OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom AI chip — built for inference (running already-trained models, not training them) and, by OpenAI's own benchmarks, beating Nvidia's Blackwell chips on performance per watt in most tests.
Open in the XNewsAi app →Why this matters
For the 806
Nvidia chips are the single biggest cost and bottleneck in building AI today. If OpenAI's own chip works as claimed at real scale, running AI gets cheaper and less dependent on one supplier — which matters for how fast AI can grow and what it costs everyone downstream.
What we know
- Jalapeño is OpenAI's first custom chip, co-developed with Broadcom, reaching tape-out in about nine months.
- It's designed for inference — serving already-trained models — not for training new ones.
- OpenAI's own published benchmarks show it beating Nvidia's Blackwell-generation systems on performance-per-watt in nearly all tested scenarios.
What we don't know
- How the chip performs at full production scale and cost, outside OpenAI's own published benchmarks.
- How much of OpenAI's total compute this actually replaces, versus supplementing continued heavy Nvidia purchases.
Sources
- PrimaryOpenAI ↗
OpenAI's own published benchmark results for the Jalapeño inference chip.
- ReportingCNBC ↗
Independent reporting on competitive implications for Nvidia.
- ReportingTechCrunch ↗
Independent reporting on the chip's development timeline and deployment plans.