OpenAI's Jalapeño Inference Chip Outperforms Nvidia Blackwell in Early Benchmarks

OpenAI published the first benchmark results for Jalapeño, its custom artificial intelligence inference chip. [1] Tested on the InferenceX benchmark suite, the processor delivered 1.5 to 1.9 times more artificial intelligence work per watt at peak throughput compared to comparison systems. [1] Jalapeño registered both more tokens per user and more throughput per kilowatt than currently available state-of-the-art inference processors. [2]

The chip was tested against three open models: GPT-OSS 120B, DeepSeek R1, and Moonshot AI's one-trillion-parameter Kimi K2.5. [11] Across these models, the processor achieved 1.7 to 3.6 times lower end-to-end latency than Nvidia's GB200 and GB300 systems. [11] At low-latency operating points, the company claims 8.6 to 104.3 times more throughput per kilowatt at the fastest previous time-between-tokens settings for the GB300. [11]

OpenAI developed the chip in close collaboration with Broadcom, advancing from initial team hiring to manufacturing tape-out in approximately 16 months. [7] The 700-watt processor achieved its results operating at lower power requirements than Nvidia's GB200 and GB300 systems, which draw between 1,200 and 1,400 watts. [11] The chip features HBM4 memory, positioning it as comparable to flagship graphics processing units from Nvidia and AMD. [7]

The architecture is designed to minimize delays during the prefill and communication phases of inference processing. [2] By keeping model state explicit and local, the system activates specific combinations of compute, memory, and networking for each inference phase to minimize data movement. [2]

OpenAI utilized its own artificial intelligence models to accelerate Jalapeño's development process. [1] Earlier generations assisted the team in designing the chip, while the company's latest models are accelerating how the chip is optimized and programmed. [1] The hardware is intended to function as a multigenerational platform, enabling the concurrent development of artificial intelligence products, models, processors, and memory. [2]

The company plans to integrate the hardware into its data centers later this year to support its own artificial intelligence models. [9] Richard Ho, OpenAI's head of hardware, estimated that the processor will deploy in very small volumes at the end of 2026, with more significant deployment following in 2027. [2]

Operating at 700 watts allows the company to save money when running data centers, where power represents a key cost. [9] Ho stated that Jalapeño demonstrates performance in both the high-throughput domain, allowing the company to serve many customers more cheaply, and the low-latency domain for faster response times. [9]

For highly interactive workloads, the chip delivered 2.1 to 4.1 times higher performance. [1] For consumers, this combination can translate to faster responses, more responsive Codex sessions, and more reliable access during periods of peak demand. [11] The company stated that developing in-house silicon grants more direct control over model execution and the economics of serving artificial intelligence models. [10]

By combining high throughput and low latency into a single architecture, OpenAI is attempting to rewrite the traditional hardware tradeoff that forces data centers to optimize for one or the other. This efficiency is critical as power consumption becomes the primary bottleneck for AI scaling. Comparing the 700-watt Jalapeño against Nvidia's 1,200-to-1,400-watt GB200 and GB300 systems highlights a strategic shift toward custom, model-agnostic inference silicon. If the company achieves its 2027 deployment targets, it could significantly alter the margin structure of enterprise AI serving by reducing dependency on expensive, power-hungry generic GPUs. What the sources don't address: how Nvidia's next-generation architectures will compare in power efficiency by the time Jalapeño reaches full-scale deployment in 2027.