TSMC will fab Qualcomm's Dragonfly C100 data center CPUs and AI300 AI accelerators on 3nm/2nm nodes. The chips will use CoWoS or SoIC advanced packaging, per a media report cited by @dnystedt.
Key facts
- Dragonfly C100 is a Qualcomm data center CPU
- AI300 is an AI accelerator chip
- TSMC 3nm/2nm nodes targeted
- CoWoS or SoIC advanced packaging planned
- Qualcomm previously cancelled Centriq 2400 server CPU in 2018
TSMC is expected to manufacture Qualcomm’s Dragonfly C100 data center CPUs and AI300 AI accelerator chips on 3nm/2nm nodes, using CoWoS or SoIC advanced packaging, according to a media report cited by @dnystedt. The report, which references unnamed sources, marks a deepening of the foundry relationship between the two companies.
Key Takeaways
- TSMC to fab Qualcomm's Dragonfly C100 and AI300 chips on 3nm/2nm nodes.
- The move challenges NVIDIA in data center AI, but timelines and performance remain undisclosed.
What the Dragonfly C100 and AI300 Are
The Dragonfly C100 is Qualcomm’s push into server-class CPUs, targeting data center workloads where the company has historically been absent. The AI300 accelerator is an AI inference and training chip, competing with NVIDIA’s H100/B200 and AMD’s MI300X. Both chips leverage TSMC’s most advanced nodes — 3nm (N3 family) for the CPU and likely 2nm (N2) for the AI accelerator — to maximize transistor density and power efficiency.
Advanced Packaging as a Differentiator
The mention of CoWoS (Chip-on-Wafer-on-Substrate) or SoIC (System-on-Integrated-Chips) packaging is critical. CoWoS is the standard for high-bandwidth memory integration, used by NVIDIA and AMD for their AI GPUs. SoIC, a more advanced 3D stacking technology, could allow Qualcomm to integrate compute dies with HBM or SRAM more tightly, reducing latency and power. The choice of packaging will determine memory bandwidth — a key constraint for AI accelerators.
Competitive Positioning
If the report holds, Qualcomm will join the small club of companies with access to TSMC’s bleeding-edge nodes — alongside Apple, NVIDIA, AMD, and Intel. The move comes as Qualcomm’s Nuvia-based CPU cores (acquired in 2021) are being adapted for server use, and as the company seeks to diversify beyond mobile and automotive. The AI300 accelerator directly challenges NVIDIA’s dominance in data center AI, though Qualcomm has not disclosed performance targets or volume commitments.
Unknowns
The report provides no timeline for tape-out or production, nor pricing or expected performance. Qualcomm has not officially confirmed the Dragonfly C100 or AI300; the company’s last public data center CPU play, the Centriq 2400, was cancelled in 2018 after a brief run. Investors will watch for any official announcement at upcoming events like Computex or Hot Chips.
What to watch
Watch for Qualcomm’s official confirmation at Computex 2026 (June) or Hot Chips (August). Any disclosed performance benchmarks versus NVIDIA’s B200 or AMD’s MI400 will determine the competitive threat. Also track TSMC’s N2 capacity allocation — Qualcomm will compete with Apple and AMD for scarce 2nm wafers.
[Updated 26 Jun via dck_news]
Meta has signed on as a customer for Qualcomm's data center CPUs, according to a report [per Data Center Knowledge]. Qualcomm also disclosed a second unnamed hyperscaler customer and aims to build a $15 billion data center chip business by 2029. The AI300 accelerator will be available via a new Qualcomm AI Infrastructure Platform, integrating networking and software.
