OpenAI's in-house 'Jalapeño' chip beats Nvidia Blackwell on inference benchmarks
OpenAI presented the first real, public benchmarks for its in-house "Jalapeño" inference chip at Hot Chips 2026, showing a 1.5x to 4.1x advantage in performance-per-watt and latency over Nvidia's current Blackwell chips on SemiAnalysis's InferenceX benchmark.
A real, important caveat analysts are raising directly
Blackwell is Nvidia's current-generation chip, not its most recent announced architecture — Nvidia's upcoming Vera Rubin platform hasn't yet been benchmarked against Jalapeño. Comparing a new challenger chip against an incumbent's current, rather than next, generation is a real, meaningful distinction worth keeping in mind before treating this as a definitive "OpenAI beats Nvidia" result.
Why OpenAI building its own chip is a genuinely significant strategic move
OpenAI has been one of Nvidia's largest real customers — a major AI lab investing in its own custom silicon reflects a genuine, broader industry pattern of large AI companies seeking to reduce dependence on any single chip supplier, directly tying to the same memory and pricing pressures covered elsewhere in AI hardware news this week.
Benchmark results presented by the chip's own developer, even at a credible technical venue like Hot Chips, are worth treating as a real, first data point rather than a fully independent, neutral verdict — genuinely useful context, but not the final word until independent testing catches up.
Source: the-decoder.com