ByteDance, the Chinese tech giant behind TikTok, is quietly moving into one of the most strategic corners of the global tech race: building its own artificial intelligence chip. According to Reuters, the company is in talks with Samsung Electronics to manufacture an in-house processor designed specifically for AI workloads, a move that underscores both ByteDance’s growing AI ambitions and the mounting pressure on Chinese firms to secure advanced computing power amid tightening U.S. export controls.

While ByteDance has publicly denied reports of an in-house chip project, multiple signals—from hiring patterns to procurement plans—suggest the effort is real, urgent, and expensive.

What is ByteDance building, and why does it matter?

At the center of ByteDance’s strategy is a custom AI chip optimized for inference tasks, the stage where trained AI models are deployed to generate outputs, rather than trained from scratch.

Inference chips are critical for:

• Running recommendation algorithms
• Powering large language models at scale
• Serving real-time AI features to millions of users

For ByteDance, whose platforms rely heavily on AI-driven content ranking, inference efficiency directly translates into lower costs and faster performance.

Why custom chips are a big deal

Designing a proprietary chip allows companies to:

• Reduce dependence on NVIDIA’s increasingly scarce GPUs
• Optimize hardware for their own software stack
• Control long-term costs as AI workloads scale

This playbook isn’t new, but it’s becoming unavoidable.

How many AI chips does ByteDance plan to produce?

According to people familiar with the matter, ByteDance has ambitious production targets for its AI chip, internally codenamed SeedChip.

Reported production goals include:

• At least 100,000 units planned for production this year
• Sample chips expected by the end of March
• Capacity ramp-up to 350,000 units over time

If achieved, that would place ByteDance among the most aggressive first-time entrants into custom AI silicon, particularly for a company without a legacy semiconductor business.

Why Samsung is central to the talks

The discussions with Samsung Electronics reportedly go beyond fabrication alone.

Two key components of the negotiations:

1. Manufacturing the AI inference chip
2. Securing access to memory chips, which are in extremely tight supply

Memory—especially high-bandwidth memory (HBM)—has become one of the biggest bottlenecks in AI infrastructure globally. With hyperscalers and cloud providers racing to build data centers, supply is stretched thin.

For ByteDance, tying chip manufacturing and memory access together could be a strategic hedge against future shortages.

Why ByteDance is under pressure to move fast

The urgency behind SeedChip is not just competitive; it’s geopolitical.

U.S. export controls are reshaping the AI landscape

Washington has imposed sweeping restrictions on the sale of advanced AI chips to China, limiting access to cutting-edge NVIDIA processors. While workarounds and downgraded chips exist, supply is uncertain and politically fragile.

For Chinese tech companies, this has accelerated a push toward self-reliance in semiconductors.

ByteDance is not alone:

• Alibaba recently launched its Zhenwu chip for large-scale AI workloads
• Baidu already sells AI chips externally and plans an IPO for its chip unit, Kunlunxin

Compared to these rivals, ByteDance is a late mover—but potentially a fast follower.

What makes ByteDance’s approach different?

Unlike Alibaba and Baidu, which have leaned into both training and inference chips, ByteDance appears to be starting with inference.

That choice is strategic.

Why inference-first makes sense

• Inference chips are easier to design than training chips
• They deliver faster returns by cutting operating costs
• They align closely with ByteDance’s core business of content delivery

This mirrors the path taken by Amazon, which introduced Inferentia before expanding deeper into AI silicon.

How long has ByteDance been working on this?

Although the project is only now coming into public view, ByteDance’s chip ambitions are not new.

Key timeline points:

• 2022: ByteDance begins hiring chip and semiconductor talent
• 2023: The company establishes Seed, a unit focused on large language models and AI adoption
• 2024–2025: SeedChip emerges as a hardware extension of this broader AI push

Despite ByteDance’s denial of specific reports, the long-term investment pattern suggests a deliberate, multi-year strategy.

How much is ByteDance spending on AI overall?

The scale of ByteDance’s AI investment is striking—even by Big Tech standards.

Reported AI spending plans:

• Over ¥160 billion ($22 billion) is earmarked for AI-related procurement this year
• More than half is allocated to NVIDIA chips
• The remainder is split between infrastructure and advancing in-house chip development

This dual-track approach—buying NVIDIA hardware while building its own—reflects both pragmatism and caution. Custom chips take years to mature, and ByteDance can’t afford downtime in the meantime.

Why this matters beyond ByteDance

SeedChip isn’t just a corporate cost-cutting exercise. It’s part of a broader shift in how AI power is built and controlled.

The bigger picture:

• AI is becoming vertically integrated—from models to hardware
• Geopolitics is accelerating fragmentation in the chip supply chain
• More companies are deciding they can’t rely on one supplier, no matter how dominant

NVIDIA remains the undisputed leader in AI chips. But the rise of in-house silicon—from Silicon Valley to Beijing—signals a future where AI hardware is more diverse, more customized, and more politically entangled.

What to watch next

Several milestones will determine whether SeedChip becomes a cornerstone or a footnote:

• Delivery and performance of sample chips by March
• Whether Samsung formalizes a manufacturing agreement
• How quickly ByteDance can scale production
• Any regulatory response tied to cross-border chip partnerships

TL;DR

• ByteDance is developing an in-house AI inference chip, codenamed SeedChip
• Talks with Samsung include manufacturing and access to scarce memory chips
• Initial production could reach 100,000 units, scaling to 350,000
• The move is driven by cost, competition, and U.S. export controls
• It’s part of a ¥160 billion ($22B) AI investment push