OpenAI Unveils Its Own Jalapeño Chip: An Accelerator 1.5–2 Times More Efficient Than Nvidia
On August 25, 2026, OpenAI unveiled test results for its custom inference accelerator, Jalapeño. On the public InferenceX benchmark, systems using the new chip delivered 1.5–1.9x more computations per watt at peak throughput and reduced response latency by 1.7–3.6x compared to systems based on Nvidia GB200 and GB300. The chip, rated at 700W nominal power, consumed up to 550W under tested loads, with a block of 128 units reaching 1.7 exaflops in 4-bit precision. OpenAI plans to deploy Jalapeño in its infrastructure by late 2026, marking the first generation of a multi-year platform.
Jalapeño was developed in collaboration with Broadcom (for the die and networking) and Celestica (for boards and racks) over nine months. It is specifically designed for OpenAI's own predictable, high-volume inference workloads around its language models, computational kernels, and data movement, unlike Nvidia's general-purpose accelerators. The initiative is backed by an agreement with Broadcom to deploy 10 GW of OpenAI's custom accelerators between late 2026 and 2029.
While OpenAI will continue relying on external suppliers like Nvidia for training cutting-edge models and parts of inference, the company aims to control the processor architecture for its largest daily computational stream. This shift addresses the economics of inference: by building chips as internal components, OpenAI avoids paying the market premium associated with Nvidia's high-margin commercial GPUs, directly lowering the cost per query.
This move follows a trend where major AI players (e.g., Anthropic with Google TPUs) transition to custom silicon as their inference volume becomes predictable. However, specialization risks reducing flexibility for future AI architectures and creates a single point of failure with manufacturing partners. The challenge for OpenAI will be ensuring Jalapeño remains competitive through multiple future model generations.
cryptonews.ruHace 1 hora(s)
