ByteDance is building its own CPUs on Arm and RISC-V to feed its AI infrastructure

At a glance:

• ByteDance is developing custom data-center CPUs using both Arm and RISC-V architectures.
• The move is driven by 10–35% price hikes from Intel and AMD and growing US export controls.
• The company's 2026 AI infrastructure budget has reportedly surged to approximately 200 billion yuan ($29.4bn).

A dual-track strategy for silicon sovereignty

ByteDance, the parent company of TikTok, is aggressively pursuing a dual-track semiconductor strategy to power its rapidly expanding AI infrastructure. According to a Reuters report citing sources familiar with the matter, the company is developing central processing units (CPUs) based on two distinct architectures: the industry-standard Arm and the open-source RISC-V instruction-set architecture. This parallel approach allows the company to evaluate which design provides the best long-term performance and stability for its massive data-center requirements.

The shift toward in-house silicon is no longer merely a theoretical optimization for ByteDance; it has become a balance-sheet imperative. The company's 2026 AI-infrastructure budget has reportedly grown by 25%, reaching an estimated 200 billion yuan (approximately $29.4 billion). As the scale of procurement grows, the financial impact of rising external hardware costs becomes increasingly difficult to ignore, pushing the company toward vertical integration.

Commercial pressures and geopolitical headwinds

The primary catalysts for this move are a combination of aggressive vendor pricing and tightening geopolitical restrictions. ByteDance currently relies heavily on Intel and AMD for its server-CPU footprint, but both incumbents have reportedly raised prices for data-center-grade processors by between 10% and 35% in recent successive quarters. This pricing spiral creates a significant procurement gap that ByteDance aims to close by designing its own silicon.

Beyond the economics, the geopolitical landscape is forcing ByteDance's hand. While Arm-based server CPUs follow a proven path—already utilized by Amazon's Graviton, Microsoft's Cobalt, and Google's Axion—the RISC-V track offers a strategic hedge. RISC-V is an open-source, royalty-free architecture that is increasingly favored within China because it sidesteps the licensing and export-control risks associated with Arm’s UK-headquartered, Softbank-owned intellectual property. Beijing has actively endorsed RISC-V as a pillar of strategic autonomy as it hardens its chip-sovereignty posture through 2026.

Expanding the custom silicon ecosystem

ByteDance's semiconductor ambitions are part of a much broader, multi-pronged program. Earlier this week, the company reached an agreement with Qualcomm to secure millions of application-specific integrated circuits (ASICs) for AI data-center inference. Under this deal, Qualcomm will also assist ByteDance in bringing its own proprietary ASIC designs through to full production, signaling a sophisticated partnership model that blends off-the-shelf components with custom logic.

This internal push is also occurring under heavy regulatory scrutiny. The company has been instructed by Beijing’s National Development and Reform Commission to reject US-origin capital in funding rounds without explicit clearance. Furthermore, new travel restrictions on senior AI talent have begun to impact ByteDance staff, alongside employees from other major Chinese AI players like DeepSeek, Moonshot, and StepFun. These factors are collectively driving ByteDance to build a self-sufficient ecosystem that is less reliant on Western technology and capital.

The manufacturing hurdle and market implications

While the design phase is well underway, the ultimate success of ByteDance's program may depend on the production-foundry side. High-performance custom CPUs require leading-edge fabrication, typically at 4nm nodes or below. While TSMC remains the global leader for such advanced nodes, US export controls on advanced manufacturing for Chinese customers create a significant bottleneck. China's domestic leader, SMIC, has reached 7nm in production but currently lags behind TSMC by roughly two process nodes.

If ByteDance succeeds, it will represent another major defection from the x86 ecosystem. The trend of hyperscalers—such as AWS, Microsoft, and Google—moving toward custom Arm silicon has already shifted a meaningful share of the market away from Intel and AMD. ByteDance's entry into this space could accelerate a feedback loop: higher x86 prices drive more custom silicon adoption, which in turn reduces the volume base for x86 vendors, making it harder for them to amortize fab costs and potentially leading to even higher prices.