OpenAI Challenges NVIDIA: Its Own AI Chip Jalapeño Shows Impressive Efficiency

OpenAI, a leading player in artificial intelligence, has announced a major breakthrough in the development of its own hardware infrastructure. At the Hot Chips conference, the company presented its in-house chip for AI workloads called Jalapeño, which, it says, outperforms NVIDIA’s flagship solutions in efficiency, including the GB300 system with its impressive 1400 W power draw. This signals OpenAI’s strategic move toward building its own AI-optimized hardware base, which could significantly shape the future of the industry.

Efficiency That Changes the Game

Developed in collaboration with Broadcom, the Jalapeño chip delivered impressive results in tests conducted using SemiAnalysis’ publicly available InferenceX package. According to the released data, the 700 W version of the chip proved far more efficient than its competitors. In particular, on the metric of throughput per kilowatt of power (kW), Jalapeño outperformed NVIDIA’s GB200 and GB300 systems by 1.5 to 1.9 times. Even more impressive were the end-to-end latency results, where OpenAI’s chip delivered performance 1.7 to 3.6 times better than NVIDIA’s equivalents.

These figures are especially notable given that the 700 W version of Jalapeño was compared with NVIDIA accelerators consuming 1200 W and 1400 W respectively. This points to significantly better energy efficiency in OpenAI’s design, a critical factor for deploying large-scale AI systems in data centers. The company plans to begin integrating its own chips into its data centers by the end of this year, underscoring its confidence in the technology.

Testing was carried out on three powerful models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 from Moonshot AI, totaling around 1 trillion parameters. OpenAI particularly emphasizes Jalapeño’s advantages in low-latency operation modes. In these scenarios, with optimization of the time between token generations, throughput per kilowatt increased by a remarkable 8.6 to 104.3 times compared with GB300.

Technical Details and the Competitive Landscape

OpenAI carefully normalized the results by taking into account each accelerator’s stated TDP (Thermal Design Power). Interestingly, during actual testing, Jalapeño’s measured sustained power draw remained significantly below its nominal rating, reaching 550 W or less. This is another indication of its exceptional efficiency. Even when comparing total useful power per accelerator (1.18 kW for Jalapeño versus 2.55 kW for GB300), the gap remains substantial. However, in multi-token prediction modes, which NVIDIA more often uses in production environments, Jalapeño’s peak efficiency advantage narrows to about 1.5 times.

It is worth noting that Jalapeño was not compared with NVIDIA’s Vera Rubin system, a key competitor in the high-end computing segment. The main prediction benchmark tests were based on a single token, which is more typical for latency testing.

SemiAnalysis, according to OpenAI engineers, confirmed that Jalapeño outperforms all tested comparable solutions from NVIDIA, AMD, and Google. Each Jalapeño chip consists of a compute die complemented by six stacks of HBM4 memory with a total capacity of 216 GB and bandwidth of 15.4 TB/s. For comparison, GB300 has 288 GB of HBM3E memory at a nominal power of 1400 W. This means that per watt of nominal power, OpenAI’s chip includes roughly 50% more memory, a major advantage for workloads that require large amounts of data.

Memory Market Challenges and Future Developments

One of the key challenges in developing modern AI accelerators is ensuring sufficient memory bandwidth. At present, HBM (High Bandwidth Memory) is one of the scarcest components in the semiconductor industry. Major manufacturers such as Samsung, SK hynix, and Micron have already sold out their HBM production capacity through 2027. This creates significant constraints for all market participants.

The shortage is so severe that, according to rumors, NVIDIA is considering reduced configurations of its upcoming Rubin Ultra system with only 192 GB of memory. Kwak Noh-jung, CEO of SK Hynix, predicts that 2027 could become the most difficult period due to a chronic memory shortage.

Despite these challenges, OpenAI is actively investing in its own hardware development. According to Bloomberg, the second stage of Jalapeño’s development is in its final phase, and the chip is expected to be fully ready within a few months. Moreover, the company is already working on the third generation of its AI accelerators, reflecting a long-term strategy to build an infrastructure of its own, independent of third-party suppliers. This strategic step could give OpenAI a significant competitive edge, allowing it to optimize hardware directly for its needs and accelerate innovation in artificial intelligence.