ByteDance in Talks with Iluvatar CoreX for AI Chip Procurement

ByteDance, the parent company of TikTok, is reportedly in advanced negotiations with Chinese AI chip startup Iluvatar CoreX to secure domestic GPU supplies for its expanding AI infrastructure, marking a significant acceleration in China’s push to replace NVIDIA hardware amid tightening U.S. export controls.

Photo by Brett Jordan on Unsplash

ByteDance Diversifies Its AI Chip Supply Chain

ByteDance is pursuing a multi-vendor AI chip strategy to reduce its overwhelming dependence on NVIDIA GPUs, according to sources familiar with the matter. The company is among the world’s largest consumers of AI accelerators, operating massive data centers to power recommendation algorithms for TikTok, Douyin, and its growing portfolio of large language models.

Iluvatar CoreX (天数智芯), founded in 2018 and headquartered in Shanghai, has emerged as a frontrunner in China’s domestic GPU race. The company’s flagship BI-V150 GPU is designed for AI training workloads and is positioned as a direct alternative to NVIDIA’s A100. Sources indicate that ByteDance has been conducting extensive testing of Iluvatar’s silicon across several of its AI model training pipelines since early 2025.

The negotiations come as U.S. export restrictions — tightened in October 2023 and further refined in 2024 — have effectively cut off Chinese companies from accessing NVIDIA’s H100, H200, and Blackwell-series GPUs. While NVIDIA developed the China-specific H20 chip to comply with regulations, even that reduced-capability product has faced additional scrutiny, leaving Chinese tech giants scrambling for viable domestic alternatives.

The Chinese AI Chip Landscape

ByteDance is evaluating multiple domestic chip suppliers beyond Iluvatar CoreX. The company has also been testing hardware from Enflame Technology (燧原科技), Cambricon (寒武纪), Hygon (海光信息), and Moore Threads (摩尔线程). This multi-supplier approach reflects a broader industry consensus that no single Chinese GPU maker can yet match NVIDIA’s full-stack ecosystem, but a portfolio of alternatives can collectively reduce dependency.

Why Iluvatar CoreX Stands Out

Several factors make Iluvatar CoreX an attractive partner for ByteDance. The company’s BI-V150 delivers competitive floating-point performance for AI training tasks and supports mainstream deep learning frameworks. Unlike some competitors that focus primarily on inference, Iluvatar has prioritized the training segment — which is where ByteDance’s most intense compute demand lies.

Iluvatar has also invested heavily in software compatibility, developing a CUDA-compatible software layer called IKUDA that allows developers to port existing CUDA workloads with minimal code changes. This software portability is critical for ByteDance, which runs vast codebases built on NVIDIA’s CUDA ecosystem. Without a credible migration path, even superior hardware would face adoption barriers.

Photo by Luka Borazan on Unsplash

Strategic Implications

A formal deal between ByteDance and Iluvatar CoreX would be one of the most significant domestic GPU procurement agreements in China’s tech sector to date. ByteDance’s endorsement would serve as a powerful validation signal, potentially attracting other enterprise customers who have been hesitant to commit to unproven Chinese silicon.

The move also reflects a pragmatic shift in how Chinese hyperscalers approach the chip supply problem. Rather than waiting for a single domestic “NVIDIA killer,” companies like ByteDance are building resilient, multi-vendor supply chains that combine chips from several Chinese GPU makers alongside legacy NVIDIA inventory acquired before export bans took full effect.

However, significant challenges remain. Chinese GPUs still lag behind NVIDIA’s latest offerings in raw performance, power efficiency, and — most critically — software ecosystem maturity. The CUDA moat that NVIDIA has built over 15+ years cannot be replicated overnight, and even CUDA-compatible layers like IKUDA require ongoing engineering investment to maintain parity with NVIDIA’s rapidly evolving software stack.

CII Analysis

This development underscores a critical inflection point in the China AI industry. ByteDance’s willingness to engage seriously with domestic GPU suppliers — despite known performance gaps — signals that geopolitical risk has become a primary factor in procurement decisions, outweighing pure technical benchmarks.

For investors tracking China’s semiconductor ambitions, the ByteDance-Iluvatar talks represent a concrete demand signal that could catalyze the domestic AI chip ecosystem. If ByteDance commits to significant volumes, it would provide Iluvatar with the revenue base and real-world feedback loop needed to accelerate product iteration — a virtuous cycle that has historically been the missing ingredient for Chinese chip startups.

Our assessment: the decoupling of China’s AI infrastructure from U.S. chip supply chains is accelerating faster than most Western analysts projected. While Chinese GPUs are not yet performance-competitive with NVIDIA’s best, they are becoming “good enough” for a growing range of workloads — and the strategic imperative to adopt them is overriding technical reservations. Companies and investors should monitor domestic GPU adoption rates as a leading indicator of China’s AI infrastructure independence timeline.

For more analysis on China’s AI supply chain developments, see our coverage of the China AI Industry and China Semiconductor Industry.