ByteDance, the parent company of TikTok, is reportedly advancing plans for an in-house AI inference chip, codenamed SeedChip, as part of efforts to bolster its AI infrastructure.
According to sources, the company is negotiating with Samsung Electronics for chip production and memory supply, aiming to reduce reliance on Nvidia amid global supply constraints and U.S. export controls. However, ByteDance has denied the accuracy of reports on the project, while Samsung declined to comment.
Project Details and Production Targets
The SeedChip project focuses on AI inference — processing trained models for real-time applications — rather than training. Development began in 2022 with talent acquisition, escalating in 2023 with the formation of a dedicated "Seed" team for AI models. ByteDance anticipates receiving sample chips by late March 2026, with initial production targeting at least 100,000 units in 2026. Volumes could scale to 350,000 units, supporting internal AI services like content recommendation and generation.
Negotiations with Samsung and Memory Challenges
Discussions with Samsung encompass not only chip fabrication but also critical high-bandwidth memory (HBM) supplies, which have become a bottleneck in AI infrastructure. HBM shortages are limiting scalability even when compute chips are available, shifting the industry's narrowest constraint from GPUs to memory. This deal could provide ByteDance with a strategic edge in securing components amid surging global demand.
Massive AI Investments and Nvidia Dependency
ByteDance's AI procurement budget is set to surpass 160 billion yuan ($22 billion) in 2026, with over half earmarked for Nvidia chips, including H200 models. Despite this, the in-house chip push aims to diversify supply chains and control costs, aligning with broader efforts to mitigate U.S. restrictions on advanced tech exports to China.
Following the Global Trend
ByteDance's initiative mirrors strategies by tech giants worldwide:
- Google: Tensor Processing Units (TPUs) for AI workloads.
- Amazon: Trainium for training and Inferentia for inference.
- Microsoft: Maia accelerators.
- Alibaba: Recently launched Zhenwu chip for large-scale AI.
- Baidu: Kunlunxin chips, with plans for external sales and listing.
This trend underscores the transformation of the AI race into a hardware arms race, where companies build custom silicon and supply chains to optimize performance, reduce costs, and ensure scalability.
Company Responses and Uncertainties
ByteDance has refuted the reports, stating the information on the in-house chip project is "inaccurate" without further elaboration. Samsung has not commented. These developments come as Chinese firms accelerate domestic chip efforts amid geopolitical tensions.
Also read:
- Analyzing Ogilvy Social Lab's 2026 Social Trends Report: Key Insights and My Recommendations
- Meta's Wild Patent: Digital Resurrection to Keep You Posting from the Grave
- Who Will Benefit Most from AI? The World in 2028–2029 and Regional Divides
Conclusion
If realized, ByteDance's AI chip could enhance its competitive position in the AI landscape, reducing Nvidia dependency and addressing memory bottlenecks. As the AI boom intensifies, the focus on proprietary hardware highlights a shift toward integrated ecosystems, where control over the full stack — from chips to applications — becomes paramount.
