OpenAI has revealed that its new AI processing chip, named "Jalapeño" performs better than Nvidia's current-generation chips in certain tests. This marks progress for OpenAI in developing its own AI chips to reduce reliance on major chip manufacturers.
Benchmark tests show Jalapeño excels in two areas: AI performance per unit of power and response speed. The chip was compared against Nvidia's GB300, a leading chip on the performance ranking used in this evaluation.
Richard Ho, head of OpenAI's chip development team, told Bloomberg that the tests were conducted using a public benchmark system comparing AI chip performance.
OpenAI plans to begin deployingJalapeñoto support its AI models within this year, as part of efforts to build its own AI infrastructure from chips to processing systems.
This move comes amid intensifying competition in the AI chip industry, with many tech companies developing in-house chips for internal use, while Nvidia continues to dominate the AI chip market.
Richard Ho stated that if OpenAI can scale up Jalapeño production as planned, the new chip is likely to significantly reduce the company's costs.
However, he noted that the extent of cost reduction depends on how much Jalapeño production can be increased and whether anticipated cost savings materialize. OpenAI's goal is to use its own chips for more AI workloads.
OpenAI developed Jalapeño in partnership with Broadcom, a semiconductor company specializing in custom chips for various clients. The collaboration was announced last year, and in June, they revealed that Jalapeño was developed in an exceptionally short timeframe compared to similar chip projects.
Typically, different chips excel at different functions: some handle many AI tasks simultaneously, while others focus on quick response times. OpenAI claims Jalapeño performs well in both aspects.
Going forward, OpenAI will decide which AI models are best suited to run on Jalapeño to ensure clients benefit from choosing chips optimized either for cost reduction or performance enhancement.
According to OpenAI's report, lab tests show Jalapeño delivers strong High Throughput performance, meaning it can handle large volumes of work efficiently, enabling OpenAI to serve more customers at lower cost.
Jalapeño also excels in Low Latency performance, reducing the delay from user input to AI response—benefiting customers who prioritize quick interaction.
Another advantage is Jalapeño's ability to operate effectively at relatively low voltage, consuming about 700 watts. This characteristic helps OpenAI lower data center operating costs since electricity and power are major expenses in AI services.
Additionally, Jalapeño is not designed for AI model training—a strength of Nvidia chips—but is optimized for inference tasks, which involve running AI models after training, such as answering user prompts or processing commands.
Simply put,Jalapeñofocuses on deploying trained AI models rather than the computationally intensive process of creating or training models from scratch.
Jalapeño's speed performance enables OpenAI to handle certain tasks that previously required chips with different memory architectures and designs.
Currently, OpenAI also relies on Cerebras Systems technology for some AI models, but Cerebras chips are better suited for very small models, whereas Jalapeño supports larger AI models.
Richard Ho added that OpenAI still requires massive computational power, which is why the company maintains agreements with multiple computing providers.
OpenAI also revealed that Jalapeño was tested using both its own small open-source AI models and models from other companies, including DeepSeek and Moonshot AI, with findings showing thatJalapeñohad a clear advantage when running Moonshot AI's Kimi model, the largest among those tested by OpenAI.
Furthermore, internal tests showed Jalapeño performs well with some large, advanced OpenAI AI models not yet publicly released. These results suggest that the larger and more complex the workload, the greater Jalapeño's value and advantage.
OpenAI noted in a company post that Jalapeño's architecture may be particularly well suited to large-scale, high-compute AI workloads.
Another interesting point is that OpenAI used its own AI to accelerate the chip development process and is already progressing on a second-generation chip.
OpenAI expects to reach the Tape Out stage—the final critical design phase for chips—within the next few months, after which actual chip manufacturing will commence.
Moreover, OpenAI is not stopping at the first and second generations but has begun conceptualizing a third-generation chip aimed at reducing AI infrastructure costs, as the company invests in building infrastructure globally.
Richard Ho stated that developing its own chips helps OpenAI reduce energy and infrastructure costs. "We are at a cost and energy level that will lower infrastructure expenses, and this is just the first step," Ho said.
Although OpenAI is developing its own chips and comparing Jalapeño to Nvidia's, it still views Nvidia as a key supplier and plans to continue using a large number of Nvidia chips. Developing in-house chips does not mean OpenAI will stop relying on Nvidia.
The bigger picture is not that OpenAI is moving away from Nvidia, but that it is creating its own alternatives to better control costs, energy use, and AI infrastructure amid rapidly growing compute demand.
For technology news and digital innovation insights from ThairathMoney to help you achieve "Good Finance, Good Life," visithttps://www.thairath.co.th/money/tech_innovation
Follow the Facebook page: Thairath Money at this link -https://www.facebook.com/ThairathMoney