5 reasons ByteDance is building custom CPUs

ByteDance is designing its own data-center CPUs to cut AI infrastructure costs and reduce chip risk.

ByteDance is moving into in-house server CPUs for one big reason: the company says its AI buildout is getting too expensive and too exposed to outside suppliers. Reuters reported that Intel and AMD data-centre chip prices have risen 10% to 35% in recent quarters, while ByteDance’s 2026 AI-infrastructure budget reportedly climbed 25% to about 200 billion yuan, or $29.4 billion.

Item	Architecture	Why it matters
ByteDance Arm track	Arm	Proven server path
ByteDance RISC-V track	RISC-V	Lower licensing and control exposure
Intel server CPUs	x86	Current supplier, higher prices
AMD server CPUs	x86	Current supplier, higher prices

1. Rising CPU prices are squeezing the budget

ByteDance is not treating custom silicon as a side project. It is reacting to a procurement problem that has become expensive enough to affect group-level economics, especially as AI infrastructure spending keeps rising.

According to the Reuters reporting cited by TNW, Intel and AMD have raised data-centre-grade processor prices by 10% to 35% in successive quarters. For a company with a reported 2026 AI-infrastructure budget of about 200 billion yuan, that kind of increase changes the math fast.

• Reported budget growth: 25%
• Reported AI-infrastructure budget: about 200 billion yuan
• Reported quarterly price increases: 10% to 35%

2. Arm gives ByteDance a proven server route

One of ByteDance’s two design tracks is based on Arm, which is the safer bet if the goal is to get a production CPU into data centres without inventing a new software ecosystem. Arm-based server chips already power major cloud platforms.

Amazon’s Graviton, Microsoft’s Cobalt, and Google’s Axion show that Arm can scale in real-world cloud operations. That makes Arm the practical option for ByteDance if it wants a cleaner path from design to deployment.

• Arm server CPUs are already in production at hyperscalers
• Lower ecosystem risk than a newer ISA
• Good fit for general-purpose server workloads

3. RISC-V offers more control inside China

The second track uses RISC-V, an open-source instruction-set architecture first developed at Berkeley. ByteDance’s interest in it is not just technical. It is also about licensing, sourcing, and political exposure.

RISC-V is less proven at server scale than Arm, but it is increasingly attractive in China because it avoids some of the licensing and export-control issues tied to Arm’s UK-headquartered, SoftBank-owned IP. Beijing has also been backing RISC-V as part of its chip-sovereignty push.

• Royalty-free ISA
• Lower dependence on foreign IP holders
• Strategic fit with China’s autonomy goals

4. Export controls make outside chips harder to rely on

ByteDance is building against a political backdrop that makes imported chips harder to count on. US export controls have tightened, and Chinese firms are under more pressure to reduce dependence on US-origin technology where possible.

The company has also been told by Beijing’s National Development and Reform Commission to avoid US-origin capital in funding rounds without clearance. That does not just affect financing. It reinforces the same direction of travel in hardware planning.

Design risk + supply risk + policy risk = stronger case for in-house CPUs

5. ByteDance is already expanding beyond CPUs

The CPU program is part of a wider chip strategy, not a one-off experiment. ByteDance recently reached an agreement with Qualcomm to supply millions of ASICs for AI data-centre inference, while Qualcomm also helps bring ByteDance’s own ASIC design toward production.

That matters because it shows ByteDance is splitting chip work by job: custom ASICs for inference, custom CPUs for general server use. In other words, it is building a broader silicon stack for AI infrastructure rather than relying on a single vendor class.

• Qualcomm deal covers millions of ASICs
• ASICs target AI inference
• Custom CPUs target data-centre general compute

How to decide

If you care most about near-term deployment, Arm is the cleaner bet because it already has a strong server record. If you care most about control and political insulation, RISC-V is the more interesting track, even if it is less mature at scale.

For ByteDance, the answer is not either-or. It is both, because rising x86 prices, export controls, and a much larger AI budget make diversification look less like an option and more like a requirement.