thenextweb.com web signal

ByteDance Negotiates 50,000-Chip Deal With Iluvatar CoreX

bytedance chips china ai china-ai chips export-controls

Key insights

  • ByteDance's potential 50,000-chip order would be Iluvatar CoreX's first major commercial-scale deal outside its government procurement base.
  • Iluvatar's TianGai-100 chips are targeted at inference rather than training, indicating commercial viability for a specific, high-volume production workload.
  • If finalized, ByteDance would have three confirmed domestic GPU suppliers plus a potential fourth in Baidu's Kunlunxin, spanning Huawei, Cambricon, and Iluvatar.

Why this matters

Iluvatar CoreX moving from near-exclusive government procurement to a potential deal with one of China's largest AI spenders signals that domestically made inference-grade chips have reached a viability threshold for production-scale workloads. ByteDance assembling a multi-vendor domestic GPU stack including Huawei, Cambricon, Iluvatar, and potentially Baidu's Kunlunxin represents deliberate supply-chain diversification under continued US export restrictions, giving Chinese AI leaders structural resilience against further Nvidia access cuts. For practitioners tracking China's AI infrastructure, this demonstrates that the inference tier can now be served by multiple competing domestic vendors, marking a maturation beyond stopgap substitution.

Summary

ByteDance is in talks to buy at least 50,000 AI chips from Iluvatar CoreX, a Shanghai-based GPU maker that until now sold almost entirely to government buyers. The chips are from Iluvatar's TianGai-100 line, which competes with Nvidia's A100 and A800 processors, and would go toward inference workloads for ByteDance's Doubao chatbot rather than model training. Essentially: (ByteDance, Iluvatar CoreX) are moving toward a deal that reshapes both companies' strategic positions. - Iluvatar went public in Hong Kong in January, reporting roughly $148 million in 2025 revenue, with about 90% coming from GPU sales. - The deal would make Iluvatar ByteDance's third domestic GPU supplier, alongside Huawei and Cambricon. - ByteDance is also separately weighing Baidu's Kunlunxin chips, which would add a fourth domestic source. US export restrictions on Nvidia hardware remain the structural driver, pushing China's major AI companies to actively diversify their domestic chip supply.

Potential risks and opportunities

Risks

  • If Iluvatar's TianGai-100 chips underperform at Doubao inference scale, ByteDance could reduce or cancel the order, leaving Iluvatar's post-IPO commercial growth without its anchor customer.
  • Iluvatar's heavy revenue concentration, with roughly 90% of 2025 revenue from GPU sales, means any deal collapse would materially affect its commercial trajectory following the Hong Kong listing.
  • If US export restrictions tighten before domestic chips fully prove out at hyperscale, ByteDance and peers could face a capacity gap during the supplier transition period.

Opportunities

  • Iluvatar CoreX gaining ByteDance as a commercial reference customer could accelerate procurement conversations with other large Chinese AI buyers evaluating inference-grade domestic chips.
  • Cambricon and Huawei, ByteDance's existing GPU suppliers, now face a multi-vendor procurement environment that gives ByteDance leverage to negotiate better pricing and terms across all three suppliers.
  • Chinese AI companies building consumer-facing inference infrastructure now have a new domestic chip source to benchmark, potentially accelerating procurement diversification across the sector.

What we don't know yet

  • Whether Iluvatar's TianGai-100 chips meet ByteDance's performance benchmarks for Doubao inference at scale; no technical performance data was disclosed in the reporting.
  • Final deal terms and volume remain unconfirmed: the article notes details are subject to change and neither company has publicly acknowledged discussions.
  • Whether ByteDance's concurrent consideration of Baidu's Kunlunxin chips will advance to a purchase agreement and on what timeline.