OpenAI Says Its New Chip Outperforms Nvidia's Blackwell As Nvidia Prepares Earnings Release

One day before NVIDIA (NASDAQ:NVDA | NVDA Price Prediction) reports fiscal second-quarter results after today’s close, roughly 4:20 to 4:30 p.m. ET, OpenAI announced its first in-house accelerator beats Nvidia’s Blackwell-generation systems on inference. The chip, codenamed Jalapeño, was co-developed with Broadcom on silicon and networking and with Celestica on systems integration, a product of the October 2025 deal to co-develop 10 gigawatts of custom AI accelerators. The timing coincides with earnings, but that alone does not signal impact on tonight’s numbers.

What OpenAI Claims, and What It Doesn’t

OpenAI’s self-reported benchmarks show Jalapeño delivering 1.5x to 1.9x more AI work at peak throughput versus Nvidia Blackwell-generation systems, with 1.7x to 3.6x lower end-to-end latency and 2.1x to 4.1x faster ultra-low-latency interactive inference across GPT-OSS-120B, DeepSeek R1 and Kimi K2.5. Each rack packs 128 accelerators, 1.7 exaFLOPS of 4-bit compute, 27.5 TB of HBM4 and just under 2 petabytes/sec of memory bandwidth. VP of Hardware Richard Ho said the chip “achieves high throughput and low latency simultaneously, a first in the industry.”

Key caveats: Jalapeño handles inference only; Nvidia’s training dominance remains untouched. The comparison targets Nvidia’s GB200 NVL72 and GB300 NVL72 racks that launched in 2024 and 2025, not the upcoming Vera Rubin platform. Tests excluded speculative decoding, and OpenAI did not disclose system-level power, so this measures throughput, not performance per watt. Deployment trickles out later in 2026, with volume production in 2027. AMD’s and Nvidia’s latest racks still deliver 1.46x to 2x more compute and up to 12% more memory than OpenAI’s rack, at roughly 85% of its memory bandwidth. That’s a tradeoff.