Nvidia Commits $3.5 Billion to MediaTek for Custom AI Chip Integration

The agreement centers on licensing Nvidia's GPU architectures and CUDA software to MediaTek for system-on-chip products aimed at data center and edge AI workloads. Company filings and joint statements indicate the first products are slated for 2026, with MediaTek supplying reference designs to hyperscalers seeking alternatives to fully in-house ASICs.

This move responds to documented shifts where Google, Amazon, and Microsoft have deployed their own tensor processing units to reduce reliance on merchant GPUs. Primary earnings transcripts show Nvidia's data center revenue grew 126 percent year-over-year in fiscal 2024, yet custom silicon now accounts for an estimated 15-20 percent of new AI accelerator deployments according to industry supply-chain data.

By opening CUDA compatibility to MediaTek's foundry relationships, Nvidia lowers the switching cost for customers while retaining software control. The arrangement also positions the firm to navigate U.S. export restrictions on advanced chips to China through a Taiwanese intermediary with established manufacturing channels.

Next steps hinge on whether major cloud providers adopt the resulting reference platforms in volume; initial design wins are expected to appear in 2025 earnings reports from both companies.