Qualcomm has just announced that it will launch the AI200 chip next year, followed by the AI250 in 2027. These accelerators are designed specifically for artificial intelligence in data centers, where they focus on inference, meaning running already trained AI models. The company, which until now has focused mainly on chips for mobile devices and wireless networks, is now targeting a sector dominated by Nvidia with a market share of more than 90%. Qualcomm shares jumped 11% following the announcement.
Durga Malladi, Qualcomm's general manager for data centers and edge computing, explained that the chips are based on Hexagon NPU technology, which the company has already successfully deployed in smartphones. "We first proved ourselves in other areas, and then it was easy to bring it to data centers," Malladi said at a press conference. Qualcomm offers these chips in fully equipped liquid-cooled rack systems that can contain up to 72 units working as a single computer. Each such rack consumes 160 kilowatts of power, which is comparable to some systems from Nvidia.
Memory and power consumption advantages
One of the main advantages of the AI200 and AI250 chips is their support for 768 gigabytes of LPDDR memory per card, exceeding the offerings from Nvidia and AMD. This high capacity makes it possible to process massive AI models more efficiently, with lower power consumption and overall operating costs. Qualcomm claims that its solution delivers up to 10 times more bandwidth than existing alternatives, accelerating inference for large language models.
The company is focusing on inference rather than model training, where Nvidia dominates. For example, OpenAI, which uses Nvidia GPUs to train its GPT models in ChatGPT, recently announced plans to purchase chips from AMD and potentially acquire a stake in the company. Qualcomm sees an opportunity in the fact that inference accounts for a large portion of cloud operations and is offering its chips as a cheaper and more energy-efficient alternative. Malladi added that Qualcomm also sells the chips separately, allowing clients such as hyperscalers—large cloud service providers—to design their own racks. He even suggested that Nvidia or AMD could purchase certain components from Qualcomm, such as its CPUs.
Market estimates
Qualcomm has already secured a customer in the form of Saudi Arabian company Humain, with which it entered into a partnership in May. Humain has committed to deploying Qualcomm systems in data centers with a total capacity of up to 200 megawatts. This represents a massive deployment in a region where demand for AI infrastructure is growing.
According to a McKinsey estimate, a total of $6.7 trillion (approximately CZK 157 trillion) will be invested in data centers by 2030, with most of it going toward systems equipped with AI chips. Nvidia has a market capitalization of more than $4.5 trillion (about CZK 106 trillion), thanks to GPU sales, but competition is growing. In addition to AMD, Google, Amazon, and Microsoft are also developing their own AI accelerators for their cloud services. Qualcomm is trying to capture a share of the market by offering flexible solutions that allow customers to "mix and match" components as needed.
The future in data centers
The AI200 and AI250 are designed for rack-scale deployment, meaning that an entire server rack operates as a unified system. Qualcomm has not disclosed the prices of the chips, cards, or complete racks, nor the exact number of NPUs in a single rack. However, it emphasizes the advantages in power consumption and a new approach to memory management, which should reduce overall costs for cloud operators.
This move by Qualcomm brings a breath of fresh air to a market that is currently attracting the attention of the world's largest technology companies. With growing demand for AI, such innovations are expected to help make advanced models more widely available for a broader range of uses.
Sources: qualcomm.com and cnbc.com
