MediaTek closed up 10% in Taipei trading after announcing a $3.5 billion joint development agreement with Nvidia spanning custom AI silicon, automotive compute, and PC integration. The partnership marks MediaTek's first named collaboration with Nvidia at the silicon design level, moving the Taiwanese firm beyond its traditional application processor business into the higher-margin custom chip segment that has defined Taiwan Semiconductor Manufacturing's recent growth.
The agreement structures around three workstreams: integrating Nvidia's GPU technology into MediaTek's automotive system-on-chip roadmap, co-developing AI accelerators for enterprise edge compute, and embedding Nvidia's software stack into MediaTek's PC platform business. MediaTek will license select Nvidia IP blocks rather than reselling finished chips, a distinction that preserves MediaTek's design margins while giving Nvidia distribution into MediaTek's 2.1 billion annual unit shipment base. The $3.5 billion figure represents projected five-year silicon purchases from TSMC under the joint roadmap, not a direct payment between the two companies. MediaTek's existing TSMC allocation expands by approximately 15% starting in the second half of this year to accommodate the new designs.
The automotive portion carries the most near-term weight. MediaTek currently ships 40 million automotive chips annually, almost entirely infotainment and connectivity silicon at $8-$12 average selling prices. The Nvidia collaboration targets the $150-$400 domain controller segment where Qualcomm, Mobileye, and Nvidia itself currently compete. MediaTek's first Nvidia-integrated automotive chip tapes out in the fourth quarter with production starting in late 2026, aligning with the 18-month automotive qualification cycle. Three unnamed Chinese EV manufacturers and one European Tier 1 supplier have already committed to design-ins, according to MediaTek's investor presentation. The custom AI silicon piece addresses the enterprise inference market where $200-$600 ASP chips serve edge deployments in retail, industrial, and telecommunications. MediaTek has no meaningful presence in this segment today; Nvidia's participation provides both the IP foundation and customer credibility MediaTek lacks. First silicon samples in the first quarter of next year.
The market's 10% response reflects two factors. First, MediaTek's smartphone business has contracted 18% year-over-year as Chinese handset demand remains weak, making any growth vector beyond mobile significant for a company trading at 14x forward earnings. Second, the partnership implicitly validates MediaTek's custom silicon capabilities at a time when fabless design houses face existential questions about value capture between hyperscalers designing their own chips and TSMC controlling manufacturing. Nvidia does not partner casually; the company's IP licensing arrangements remain rare outside of its automotive Drive platform. MediaTek gains a credible story that it can compete upstream against Broadcom and Marvell in custom silicon, a $28 billion market growing at 22% annually.
Allocators should track three items: MediaTek's December quarter earnings in late January, where management will detail the TSMC wafer commitment and margin impact; any named automotive design wins beyond the three Chinese OEMs, particularly from legacy German manufacturers who have resisted Chinese silicon; and Qualcomm's response, since the Snapdragon maker announced its own custom AI chip initiative in September without a comparable partnership. MediaTek's custom silicon revenue will stay immaterial through 2025 but could reach $1.2-$1.8 billion by 2027 if the automotive and enterprise roadmaps hold.
Nvidia's stock moved 1.8% on the news. The company now has silicon integration partnerships with MediaTek, Mediatek rival Qualcomm through the PC program, and its own captive automotive business. Jensen Huang is building distribution without diluting his own margin structure, a pattern worth watching as the AI silicon market matures past the current data center sprint.