Reuters exclusive: Qualcomm accelerates its data‑center push, twin AI chips to enter mass production next year, shares surge 20%
On October 27, 2025, U.S. chip giant Qualcomm held an "AI Day" event at its San Diego headquarters and formally announced two chips designed to accelerate AI in data centers — Ares‑1 and Ares‑2 — planned for mass production and large‑scale commercial deployment in the first half of 2026. The move marks Qualcomm's strategic shift from smartphone chip leader to a core player in AI infrastructure, directly targeting NVIDIA's dominant position in data‑center GPUs. Qualcomm's stock spiked an intraday 20.3%, its largest single‑day gain since 2023, adding $38 billion in market value overnight; the market sees the launch as a new turning point in the AI hardware arms race.
The Ares series is fabricated on a 5nm+ process and uses a self‑developed NPU (neural processing unit) + CPU + GPU heterogeneous architecture optimized for large‑model training, inference and edge AI. Official figures claim the Ares‑1 can reach 2.1 PFLOPS in FP8 at a power draw of 350W; the flagship Ares‑2 integrates eight NPU clusters, delivers up to 5.8 PFLOPS, supports NVLink‑level interconnect and can scale to a 1024‑card supercomputing cluster. Qualcomm emphasized the chips have a 38% higher performance‑per‑watt than NVIDIA's H100 and claim a 45% reduction in total cost of ownership when training a Llama‑70B model.
"Over the past decade we defined the AI experience on mobile; over the next decade we will redefine AI infrastructure in the cloud," Qualcomm CEO Cristiano Amon said at the event. He said Qualcomm has reached strategic agreements with Microsoft Azure, Google Cloud and Oracle, and that Ares chips will be first deployed on Azure AI supercompute nodes in Q2 2026 to support training of GPT‑5‑class models. Amon also said OpenAI has signed a letter of intent to test Ares‑2 in its Sora 3 video‑generation cluster, with a potential full migration expected in 2027.
The move directly challenges NVIDIA's roughly 90% share of the AI data‑center market. Backed by its H100/B200 GPU series, NVIDIA's AI chip revenue for fiscal 2025 was expected to top $120 billion, but Qualcomm's entry could alter that concentration. Morgan Stanley analyst Joseph Moore commented: "Ares from Qualcomm is not a simple imitation but a system‑level approach — it does not sell single cards but offers full rack solutions," including liquid cooling, power management and a software stack. Qualcomm's in‑house QAI Runtime runtime aims to support PyTorch, TensorFlow and ONNX, lowering migration costs for developers.
On the technical side, Ares uses a chiplet architecture, integrates HBM3e high‑bandwidth memory (up to 192GB per card) and CXL 3.0 interconnect, and supports memory pooling to improve multi‑card coordination. Compared with NVIDIA's NVSwitch, Qualcomm's Q‑Link interconnect is claimed to have 22% lower latency and 1.8TB/s bandwidth. Qualcomm also announced it will open‑source the QAI SDK, courting partners such as Hugging Face and LangChain, with the stated goal of building the world's third‑largest AI developer platform by the end of 2026.
Market response was intense. Within an hour of the announcement, Qualcomm's intraday trading volume reached about eight times normal levels as retail and institutional investors piled in. Wall Street banks raised price targets: Goldman Sachs lifted its target from $185 to $240, and JPMorgan issued an "overweight" rating and predicted AI chips could account for 30% of Qualcomm's revenue by 2027. Meanwhile, NVIDIA's stock slid about 4.1% intraday, wiping roughly $110 billion off its market value in the largest single‑day drop this year.
Qualcomm's AI push has been prepared over time. In 2024 Qualcomm acquired Israeli AI chip startup Decima, gaining sparse‑computation patents; in early 2025 it secured a 500,000‑wafer 5nm capacity agreement with TSMC to ensure Ares supply. Amon said Qualcomm has invested more than $3 billion in AI R&D and plans to add $5 billion over the next three years to "make performance per watt the industry standard."
Challenges remain. NVIDIA's CUDA ecosystem has deep roots and strong developer lock‑in; although Qualcomm has seen PC success with Snapdragon X Elite, the data center is a new battleground. Bernstein analyst Stacy Rasgon warned: "Qualcomm must prove Ares is not just a 'cheaper GPU' but a 'better AI system,' otherwise customers will be reluctant to bear the migration risk."
China is another major variable. Qualcomm said Ares will be produced simultaneously in mainland China and could form a three‑way competitive landscape with Huawei's Ascend and Cambricon. But geopolitical risks loom — the U.S. Commerce Department is reportedly evaluating whether to place Ares on the Entity List on grounds it could be "used for military AI training."
Qualcomm's AI breakout reflects a new configuration in the global chip industry: from mobile to cloud, from consumer to enterprise, AI is reshaping the value of every silicon chip. As NVIDIA's market value approached $4 trillion, Qualcomm's actions signal that the data‑center AI chip game is only beginning.
Watch points:
- 2026 Q1: First customer test reports for Ares‑1
- 2026 Q2: Azure AI supercompute nodes go live
- 2027: Will OpenAI's Sora 3 abandon NVIDIA for Qualcomm or move to in‑house silicon?
Whether Qualcomm's Ares can become the "iPhone moment" for AI chips remains to be seen, but the market's initial verdict was clear — the stock rocketed.