Following Google, OpenAI and Anthropic Develop AI Chips
Reducing Reliance on Nvidia and Strengthening Control

[Into the World of AI] Big Tech Moves to Develop Custom Chips: "Now It's Infrastructure Over Models" View original image

OpenAI has unveiled its own AI inference-specialized chip, "Jalapeño," accelerating the company's efforts toward self-reliant AI semiconductor development. The move to develop in-house chips is seen as an attempt to reduce reliance on Nvidia, which supplies graphics processing units (GPUs), and as evidence that big tech companies are focusing on vertically integrating their AI infrastructure.


On June 24 (local time), OpenAI and Broadcom announced the launch of Jalapeño, an AI chip they co-developed, and revealed plans to deploy it in their own data centers by the end of this year. Jalapeño is an application-specific integrated circuit (ASIC) optimized for AI inference—that is, the process by which ChatGPT answers user queries.


OpenAI's decision to develop its own chip is primarily driven by the need to lower its dependence on Nvidia and to cut costs. As ChatGPT usage has soared, OpenAI has been spending enormous sums on acquiring GPUs, making it urgent to improve performance per watt and reduce operating expenses. Given that OpenAI is Nvidia's largest single customer for GPUs, this shift is expected to have a significant impact on the AI accelerator market.


The AI market is currently experiencing rapid expansion in the inference segment, outpacing training. Rather than relying on Nvidia's general-purpose accelerators, OpenAI aims to enhance efficiency by using custom chips optimized for its own services such as ChatGPT and Codex. Notably, OpenAI is not just providing AI models but is working to design and control every aspect of its services, data centers, and chips—focusing on building a comprehensive AI full-stack. OpenAI and Broadcom plan to release the next-generation chip in 2028, with new chips to be introduced annually.


Sam Altman, CEO of OpenAI, visiting South Korea

Sam Altman, CEO of OpenAI, visiting South Korea

View original image

An official from the Institute for Information & Communications Technology Planning & Evaluation (IITP) explained, "As the scale of AI services grows, GPU procurement costs are increasing exponentially, and securing GPUs itself has become a competitive edge. For big tech companies, alternatives are urgently needed," adding, "Designing their own chips also gives these companies an advantage in price negotiations with Nvidia, which is why big tech firms are racing to develop proprietary chips."


Previously, Google led the move away from Nvidia by developing its own chip, the Tensor Processing Unit (TPU). Amazon is also negotiating ways to sell its AI semiconductors not only for internal use but also to data centers of external enterprises. Amazon’s AI accelerator, "Trainium," launched in 2020, and has been adopted by OpenAI, Anthropic, and Uber Technologies via AWS.



Anthropic is also working to build its own chip capabilities. An executive in the AI semiconductor industry commented, "Companies leading the AI model market are rolling up their sleeves to move away from Nvidia. As the AI industry shifts from competition over models to competition over infrastructure, gaining an edge in the full-stack AI platform race will become even more important going forward."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing