OpenAI Benchmarks In-House 'Jalapeño' Chip, Putting Pressure on Nvidia's Inference Margins

OpenAI has revealed initial performance benchmarks for its inaugural in-house semiconductor, codenamed Jalapeño, highlighting an intensifying competitive threat to Nvidia's dominant hold on the artificial intelligence hardware market. The specialized silicon, engineered specifically for AI inference workloads, is designed to reduce operational costs and accelerate model responses as demand for generative AI applications expands.

Developed in collaboration with chipmaker Broadcom, Jalapeño is slated for deployment across OpenAI’s server infrastructure by the end of the year, with second- and third-generation iterations already under development. First announced in June as a ground-up architecture for large language models, the processor aims to deliver lower latency and improved reliability for AI agents. Market analysts speaking to CNBC noted that the introduction of high-efficiency custom silicon directly challenges Nvidia’s highly lucrative inference profit margins, even as the GPU giant maintains its near-monopoly over AI computing overall.

Adrien Sanchez, a technology analyst at Yole Group, told CNBC that Jalapeño demonstrates how hyperscaler-designed processors can now meet or surpass Nvidia’s Blackwell architecture in inference efficiency. While Nvidia continues to control the vast majority of compute capacity alongside deep software lock-in via its CUDA platform, Sanchez emphasized that inference represents the fastest-growing segment of the market. Because OpenAI has historically been one of Nvidia’s largest individual GPU customers, the transition to custom internal hardware significantly alters the dynamics of Nvidia's most critical customer relationship.

Research firm SemiAnalysis conducted on-site testing of Jalapeño at OpenAI’s development facilities, finding that the chip outperformed Nvidia’s Blackwell chips on a performance-per-watt basis across almost every tested configuration. However, SemiAnalysis noted in a published analysis titled "OpenAI Jalapeño: Better Than Nvidia Blackwell" that comparing Jalapeño directly to Blackwell is somewhat incomplete and unfair. Jalapeño integrates next-generation HBM4 high-bandwidth memory, making Nvidia’s upcoming Rubin platform—which also relies on HBM4—a more accurate hardware benchmark. While Nvidia has begun shipping Vera Rubin systems to enterprise clients, Jalapeño currently remains in the engineering sample phase.

Industry experts suggest that the economic benefits of custom application-specific integrated circuits (ASICs) could reshape infrastructure spending across the technology sector. Alexander Harrowell, a senior principal analyst at Omdia, told CNBC that deploying Jalapeño at scale will yield substantial savings across power consumption, cooling requirements, and electrical distribution infrastructure, meaningfully enhancing OpenAI’s underlying unit economics. Harrowell noted that roughly half of global AI infrastructure capital expenditure originates from hyperscale cloud operators capable of building proprietary silicon, marking custom ASICs as the primary long-term threat to Nvidia’s core business model.

Despite the gains made by custom silicon, analysts expect Nvidia to retain its stronghold over compute-heavy tasks such as initial model training and frontier research. TrendForce analyst Fion Chiu told CNBC that while custom processors like Jalapeño will gradually reduce dependence on external vendors for inference, Nvidia GPUs will remain indispensable for intensive workloads due to their flexible programmability, established software ecosystem, and multi-workload versatility.

OpenAI’s hardware push reflects a broader industry movement among tech platforms seeking to curtail hardware expenditures and reliance on third-party suppliers. In recent months, Google showcased updated Tensor Processing Units (TPUs) for training and inference, Meta agreed to a multi-gigawatt deal deploying 1 gigawatt of custom Broadcom-based silicon, and Anthropic pledged more than $100 billion over ten years toward Amazon Web Services infrastructure, including AWS’s proprietary Trainium processors. Simultaneously, specialized semiconductor startups including Cerebras, SambaNova, D-Matrix, Etched, and Fractile are advancing alternative architectures to capture market share.

According to projections from Omdia, custom ASIC deployment volumes are expected to surpass traditional GPUs by 2028. However, revenue transitions will take significantly longer given the vastly higher price points commanded by flagship GPUs. As OpenAI prepares to integrate Jalapeño into its production environment later this year, the shift underscores how the tech industry's largest consumers of compute are rapidly transforming into creators of their own hardware ecosystems.