OpenAI's Jalapeño chip crushes Blackwell on perf-per-watt — and it'll keep buying Nvidia anyway
OpenAI unveiled Jalapeño, the first chip it designed in-house, with benchmarks that mop the floor with Nvidia's flagship Blackwell — and then turned around, patted Nvidia on the shoulder, and said: don't worry, I'll keep buying.
Three Key Facts
More work per watt. In standard test suites, Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput, and cut end-to-end latency by 1.7 to 3.6 times versus Nvidia's GB200 and GB300. For inference — where the power bill is the whole game — that's a real cost cut.
There's an unfair asterisk. SemiAnalysis flagged the comparison as "not quite like-for-like": Jalapeño uses newer HBM4 memory, while the Blackwell in the ring is still on older memory. The truly matched opponent is Nvidia's next-gen Rubin platform, which also runs HBM4 — and that fight hasn't happened yet.
Clear-eyed on capacity and positioning. OpenAI plans a small-scale deployment by the end of 2026, with a larger rollout in 2027. And it said the quiet part out loud: the custom chip is about saving money on inference; for the hard war of training frontier models, it will keep buying Nvidia GPUs.
WangDou's Take
This is Silicon Valley playing chess — slap Nvidia across the face with a custom chip while sliding the other hand into its order book. OpenAI knows exactly what it's doing: inference is the electricity bill you pay every day, so every watt saved matters, so build your own; training is the occasional big battle where the chip-building trench is too deep, so keep paying Jensen his protection money. The "slap" is really a negotiation — flash a 1.9x number today, carve two more points off the price at the next purchase. As for that HBM4 asterisk, everyone gets it: beating an old-memory card with new memory isn't a clean win, but nobody prints that line on the keynote slide. The chip war isn't about who's stronger anymore — it's about who can look strong and place an order at the same time.
Source: CNBC、SemiAnalysis