Qualcomm announces AI chips to compete with AMD and Nvidia

Qualcomm has officially announced its entry into the artificial intelligence chip market, unveiling plans for new AI accelerator chips that aim to rival the current leaders, Nvidia and AMD. This strategic move signifies a significant shift for Qualcomm, which has primarily concentrated on semiconductors for mobile connectivity rather than the expansive data center market. The company revealed two upcoming models: the AI200, expected to launch in 2026, and the AI250, scheduled for 2027. These chips will be designed to fit into a full liquid-cooled server rack, aligning with the offerings from Nvidia and AMD, which provide full-rack systems capable of housing up to 72 GPUs working in unison. Such power is essential for AI labs that require robust computing capabilities for advanced model training and operation. Qualcomm's new data center chips build upon the technology found in its smartphone chips, specifically the Hexagon neural processing units (NPUs). According to Durga Malladi, Qualcomm’s general manager for data center and edge solutions, the company first needed to establish its credibility in other markets before transitioning into the data center sector. The entry of Qualcomm into this competitive landscape comes at a time when demand for AI-focused server infrastructure is surging. A McKinsey report projects that nearly $6.7 trillion will be invested in data centers through 2030, predominantly for systems utilizing AI chips. Currently, Nvidia commands over 90% of this market, with its GPUs being crucial in the training of advanced AI models like those developed by OpenAI. As AI companies, including OpenAI, seek alternatives to Nvidia, Qualcomm's timing could be pivotal. OpenAI has recently indicated plans to source chips from AMD, the second-largest GPU manufacturer, highlighting the industry's shift towards diversification in AI hardware solutions. Major tech firms like Google, Amazon, and Microsoft are also actively developing their own AI accelerators to enhance their cloud services. Qualcomm's chips are primarily aimed at inference tasks—running AI models rather than training them—potentially offering a cost-effective solution for cloud service providers. Each of Qualcomm's full-rack systems is designed to consume approximately 160 kilowatts, a figure that competes with the power requirements of some Nvidia setups. The company plans to sell its AI chips and components individually, catering to large-scale clients who prefer customized configurations. Additionally, Qualcomm envisions the possibility of other AI chip manufacturers, including Nvidia and AMD, leveraging its data center components, such as central processing units (CPUs). Malladi emphasized the company’s strategy to provide flexibility, allowing customers to choose comprehensive solutions or mix and match components according to their needs. While Qualcomm has yet to disclose pricing details for its chips or systems, it has established a partnership with Humain in Saudi Arabia, committing to supply AI inferencing chips for local data centers. This collaboration could see Qualcomm deploying a significant number of systems powered by 200 megawatts. The company asserts that its AI chips will offer advantages in power efficiency, total cost of ownership, and innovative memory management, boasting support for 768 gigabytes of memory—exceeding the specifications of current Nvidia and AMD offerings.