Summary of Key Points
OpenAI, in collaboration with Broadcom, has launched the Jalapeño推理 chip, which has sparked discussions about the potential for a shift away from reliance on NVIDIA. However, the truth is that this chip is a custom ASIC (Application-Specific Integrated Circuit) that involved extensive participation from Broadcom itself. The real winners in this scenario are companies like Broadcom and Mellanox, which specialize in designing such customized chips. The AI chip industry is moving towards a more diversified landscape, where NVIDIA still holds a strong position in training tasks, while custom ASICs for inference are gaining prominence. Broadcom and Mellanox play a crucial role as providers of these specialized components. NVIDIA will not disappear from the market, but it will no longer dominate the entire computing power landscape.
Jalapeño Chip: OpenAI’s Dedicated Inference Solution
The Jalapeño is not a “versatile” chip; it performs one specific task: to enable AI applications like ChatGPT to answer user queries more quickly and efficiently. This process is known as “inference,” which is different from the “training” phase of model development.
- Feature 1: Specialized for Inference: It is designed to perform inference tasks more efficiently than NVIDIA’s general-purpose GPUs, making it ideal for use cases where speed and efficiency are critical but not for complex model training.
- Feature 2: Cost Efficiency: There are reports suggesting that the cost of inference operations could be reduced by up to 50% (although this has not been officially confirmed by Broadcom). Since ChatGPT processes billions of requests daily, any cost reduction directly translates into increased profits for OpenAI.
- Feature 3: Rapid Development: The chip was developed and produced in just 9 months, compared to the usual 1.5–2 years for traditional ASICs. It is expected to be deployed on a large scale by the end of 2026, with the first batches going to Microsoft’s data centers.
- Important Note: The Jalapeño was not entirely developed from scratch by OpenAI; Broadcom played a key role in the design process, providing expertise and assistance with manufacturing.
Custom ASICs: The Invisible Designers Behind the Giants
You may be familiar with companies like NVIDIA and AMD, but perhaps not with Broadcom and Mellanox. These two firms are the dominant players in the custom ASIC market.
- What are Custom ASICs? Unlike NVIDIA’s general-purpose GPUs, which can handle a wide range of tasks, ASICs are designed specifically for a particular purpose. For example, Google’s TPU is used exclusively for its search and advertising optimization services, while Amazon’s Trainium serves its cloud infrastructure. Custom ASICs are more tailored to the needs of their respective companies but are not interchangeable.
- Why Don’t Giants Develop Their Own Chips? Chip design requires significant technical expertise and established manufacturing relationships. Companies like Broadcom and Mellanox have decades of experience in this field, enabling them to quickly transform conceptual ideas into functional products. In reality, when giants like Google or Microsoft need custom chips, they often turn to these companies for outsourcing.
Broadcom & Mellanox: The Stable Players in the AI Competition
During times of rapid innovation, it’s often the providers of essential components who profit the most. In the AI industry, Broadcom and Mellanox are positioned as suppliers of specialized computing power.
- Financial Performance: Broadcom reported $8.4 billion in AI-related revenue in Q1 2026, a 106% increase, while Mellanox is expected to generate $11 billion in AI ASIC sales for the year. Together, these two companies control approximately 95% of the global custom ASIC market.
- The Profitability of This Business Model: There’s no need to bet on which AI model will become popular; as long as giants require inference chips, they turn to Broadcom and Mellanox for design and production. This approach reduces risk compared to developing chips in-house.
- Market Reaction: When OpenAI announced the Jalapeño chip, Broadcom’s stock price rose by 2%. Investors are more interested in Broadcom’s ability to supply these critical components than in OpenAI’s own chip capabilities.
The Evolution of the AI Chip Industry
The industry is moving towards a more differentiated structure:
- Current Division of Labor: Taiwan Semiconductor and other manufacturers handle chip production, while NVIDIA focuses on training large models using its CUDA ecosystem.
- Customized Solutions: Companies like Broadcom and Mellanox specialize in designing inference chips for specific use cases.
- Application Requirements: Companies like OpenAI and Google need a combination of NVIDIA’s training capabilities and custom ASICs for inference tasks.
The Shift from Dominance to Collaboration
In the past, AI computing power was dominated by NVIDIA. However, the industry has evolved into a more collaborative landscape:
- New Roles: NVIDIA still dominates training tasks, but its role is shifting towards providing a platform for large-scale model development.
- Competitive Landscape: Other companies are focusing on designing and producing specialized chips to meet the growing demand for inference capabilities.
- Inference Challenges: The focus has shifted from developing the largest models to delivering high-performance solutions at lower costs. A 10% reduction in inference costs can significantly increase profits.
Will NVIDIA’s Dominance Wane?
NVIDIA will not lose its significance in the AI industry, but its dominance is no longer absolute:
- Enduring Value: The need for NVIDIA’s general-purpose GPUs and CUDA ecosystem remains strong, as developers have invested heavily in them over the years. As long as AI models continue to evolve, demand for NVIDIA’s products will remain.
- Changing Dynamics: NVIDIA’s position has changed; it is no longer the sole leader in the computing power market. Instead, it will coexist with other companies that provide specialized solutions.
Conclusion
The launch of the Jalapeño chip by OpenAI represents a shift in the AI chip industry’s power structure. The industry is becoming more fragmented, with various players playing distinct roles. The true winners are companies like Broadcom and Mellanox, which offer essential components that enable companies to innovate more efficiently. In the future, the focus will not be on who can eliminate NVIDIA but on who can secure a stable position within this new competitive landscape.