October 27, 2025 — Qualcomm unveiled the AI200 and AI250, its next generation of AI inference chips aimed at the data center market it has not previously competed in directly.
What happened
The AI200, slated for 2026 availability as individual chips, PCIe cards or liquid-cooled server racks, is built around Qualcomm’s Hexagon Neural Processing Unit and supports up to 768GB of memory per card. The AI250, planned for 2027, adds near-memory computing intended to significantly increase effective memory bandwidth for large-scale inference workloads.
Why it matters
Qualcomm’s entry positions it as a lower-cost, efficiency-focused alternative to Nvidia’s data-center dominance, specifically targeting inference workloads rather than model training, an area expected to consume a growing share of AI compute as more applications move from development into production use.
Security and infrastructure impact
A wider range of viable inference-chip suppliers could reduce the effective cost of AI-powered analytics — including the video and sensor analytics used across physical-security systems — over time, though enterprise buyers will need new due-diligence processes for supply-chain and firmware trust in each new hardware vendor.
