Qualcomm is an American semiconductor company best known for the Snapdragon chips that power most Android phones, laptops, and wearables. It doesn’t manufacture its own chips — it designs them and pays a semiconductor foundry such as TSMC to fabricate them — but its designs sit inside a huge share of the world’s mobile devices. Now Qualcomm is pushing into new territory: the AI inference chips that run AI models inside data centers, putting it in direct competition with Nvidia.

From wireless patents to Snapdragon

Qualcomm was founded in 1985 in San Diego, California, by Irwin Jacobs, Andrew Viterbi, and several colleagues from the wireless company Linkabit. Its early business was built on CDMA, a wireless communication standard the company patented and licensed to phone makers worldwide — a licensing model that still generates roughly a fifth of its revenue today. In its 2025 fiscal year, Qualcomm reported about $44 billion in total revenue, with the large majority coming from designing and selling chips, chiefly the Snapdragon system-on-a-chip: a single piece of silicon that bundles a processor, a graphics chip, a wireless modem, and a dedicated AI accelerator called the Hexagon Neural Processing Unit (NPU). Samsung, Xiaomi, and dozens of other Android phone makers use Snapdragon chips, and a growing line of Windows-on-Snapdragon laptops has carried the same chip family into PCs.

That NPU is why Qualcomm already has a foothold in AI. Every modern Snapdragon-powered phone or on-device AI laptop uses it to handle tasks like photo processing, voice recognition, and small language models locally, without sending data to a server. It’s a very different kind of AI chip from the ones that train or run today’s largest models — smaller, far more power-efficient, and built to run on a battery rather than a data-center power supply.

The push into data centers

That changed in October 2025, when Qualcomm announced the AI200 and AI250: its first chips designed specifically for data-center AI inference — running an already-trained model to answer a user’s request, rather than training a new one. The AI200, due in 2026, ships as individual chips, PCIe cards, or full liquid-cooled server racks built around the same Hexagon NPU family used in phones, scaled up to support up to 768GB of memory per card and rack configurations drawing around 160 kilowatts of power. The AI250, planned for 2027, adds “near-memory” computing designed to move data around inside the chip more efficiently — a common bottleneck when running very large models at scale.

Qualcomm is pitching the chips mainly on cost: a lower total cost of ownership per server rack than Nvidia’s GPUs, aimed at companies that want cheaper capacity to run existing models rather than raw horsepower to train new ones.

Why it matters

Qualcomm’s move matters because the AI industry has spent the past few years worried about depending on a single supplier for the chips that power it. A credible second or third source for AI inference hardware — backed by a company with four decades of chip-design experience and its own foundry relationships — gives cloud providers and AI labs more room to negotiate on price and more resilience if one supplier hits a bottleneck. Whether Qualcomm’s chips actually dent Nvidia’s dominant market share will depend on real-world performance once the AI200 starts shipping at scale.

In the news

Qualcomm’s data-center ambitions are already showing up in real deals: see our report on Qualcomm’s new Dragonfly chip family and its first major data-center customer.