OpenAI and Broadcom Launch Custom Jalapeño Chip to Slash AI Costs and Challenge Nvidia

In a move that could fundamentally rewrite the economics of the artificial intelligence industry, OpenAI and Broadcom have unveiled Jalapeño, a custom-designed intelligence processor architected specifically for large language model inference. The newly introduced chip represents OpenAI’s first official step into custom silicon, marking a major milestone in the company’s transition toward a full-stack platform that spans from consumer products and frontier models down to the physical hardware that runs them. Co-developed in a breakneck timeline, Jalapeño is designed to drastically lower the energy consumption and financial costs associated with running massive AI workloads at scale[1][2]. By designing its own application-specific integrated circuit, OpenAI aims to challenge the industry's near-total dependence on general-purpose graphics processing units, bringing specialized physical infrastructure directly in line with its long-term algorithmic roadmap[3][2].

The development of Jalapeño highlights a highly compressed design cycle that semiconductor experts consider unprecedented for high-performance advanced silicon. OpenAI and Broadcom, alongside system integration partner Celestica, managed to take the chip from a blank-slate design to manufacturing tape-out in only nine months[1][4]. This rapid turnaround was achieved through close hardware-software co-development and, notably, by utilizing OpenAI's own advanced generative models to accelerate parts of the design and optimization process[5][4]. With the designs finalized, the custom chips are manufactured by Taiwan Semiconductor Manufacturing Company, the world’s leading advanced chip foundry[6]. Engineering samples of the silicon are already active in laboratories, successfully running active machine learning workloads at production-target frequencies and power levels[1]. In particular, the startup confirmed that the silicon is currently running internal testing on its GPT-5.3-Codex-Spark model, demonstrating that the chip is fully capable of managing cutting-edge, complex architectures[1][7].

Unlike the versatile graphics processing units designed by Nvidia, which were originally built for graphics rendering and are primarily used to train massive AI models, Jalapeño is an application-specific integrated circuit tailored exclusively for inference[6][2]. As AI models have transitioned from the training phase to mass deployment, the computational focus of the industry has shifted rapidly. Running millions of daily queries on general-purpose GPUs introduces a severe efficiency penalty due to excessive data movement and an imbalance between compute power, high-bandwidth memory, and on-chip networking[1][2]. Jalapeño addresses this bottleneck directly by optimizing the balance between memory bandwidth and compute efficiency, thereby minimizing data movement during model execution[1][2]. While OpenAI is still finalizing performance measurements, early laboratory testing indicates that the chip delivers a substantial performance-per-watt advantage over the current state-of-the-art[8][1]. According to leadership from Broadcom, early versions of the processor have shown the potential to deliver cost savings of roughly fifty percent compared to standard graphics processing units[6], putting its performance on par with Nvidia's powerful Blackwell chips and Google’s proprietary tensor processing units[6].

The introduction of Jalapeño is part of an expansive, multi-generation infrastructure roadmap designed to scale OpenAI’s operations to a magnitude previously unseen in the tech sector[9]. OpenAI and Broadcom have announced a strategic collaboration to deploy custom AI accelerators at a staggering gigawatt scale[9]. The roadmap envisions deploying up to ten gigawatts of custom AI accelerators over multiple generations, with rack deployments scheduled to begin in the latter half of the year and scaling progressively through the end of the decade[10][2]. Canadian electronics manufacturer Celestica has been tapped to construct the customized server boards and rack systems[6]. To connect thousands of these processors into a cohesive, high-performance computing cluster, the infrastructure will leverage Broadcom's highly regarded Tomahawk networking silicon[1][11]. These massive hardware clusters will eventually be integrated into data centers operated by Microsoft and other key infrastructure partners, allowing OpenAI to seamlessly transition its user-facing services onto its own homegrown silicon[1][6].

The strategic implications of this partnership extend far beyond OpenAI's internal balance sheets, threatening the pricing power and market dominance of Nvidia[2]. For years, major tech firms and AI startups have grappled with severe shortages of high-end graphics processing units, leading to delays in product rollouts and inflated operational costs[3][12]. By developing proprietary silicon, OpenAI joins an elite group of technology giants—including Google, Amazon, and Meta—that have successfully fielded custom chips to bypass the Nvidia supply bottleneck[3]. Because running inference is projected to account for upwards of eighty percent of all AI compute spending by the end of the decade[2], even a moderate increase in hardware efficiency can result in billions of dollars in savings. For Broadcom, the partnership solidifies its position as the premier design partner for custom application-specific integrated circuits, expanding its market cap and reinforcing its relevance in the generative AI era[3][7].

Ultimately, the unveiling of Jalapeño signals a broader maturation of the artificial intelligence ecosystem, shifting the competitive battlefield from algorithmic breakthroughs to raw physical efficiency. As frontier models become more integrated into daily life and enterprise workflows, the ability to deliver low-latency, low-cost intelligence at scale will determine which platform dominates the next era of computing[6][12]. By vertically integrating its operations, OpenAI is attempting to ensure that its path to artificial general intelligence is not constrained by third-party hardware supply chains or prohibitively expensive energy bills[3]. Jalapeño represents a bold statement of intent, proving that the software pioneer is equally capable of innovating in the highly complex, capital-intensive world of advanced semiconductor manufacturing[8][2].