Why ByteDance Is Betting Its AI Future on Custom CPUs—On Two Architectures at Once

• Arm: The established architecture favored by cloud hyperscalers for its power efficiency and rich software ecosystem.
• RISC-V: An open-source, royalty-free architecture that offers complete customization and avoids the licensing fees and geopolitical restrictions tied to proprietary ISAs.

This dual-track strategy buys ByteDance time and leverage. It can assess which architecture best fits its longer-term performance, cost, and supply-chain needs while signaling to incumbent vendors that it has credible alternatives. Multiple external partners are expected to assist with chip design and securing foundry capacity .

Separately, ByteDance is developing a custom AI inference ASIC reportedly codenamed "SeedChip." Engineering samples were targeted for March 2026, and the company is in talks with Samsung Electronics about manufacturing up to 350,000 units. The SeedChip is designed specifically for AI inference workloads—not training—and represents a parallel effort to reduce Nvidia dependence at the accelerator level .

Why ByteDance Is Betting on In-House Silicon

Several forces are pushing ByteDance toward custom silicon:

Spiraling supplier costs. Intel and AMD have raised server CPU prices by 10 to 35 percent per quarter, a pace that is unsustainable for a company scaling its AI infrastructure at ByteDance's speed .

Prolonged supply shortages. Chip supply constraints have directly limited ByteDance's ability to expand its AI footprint .

Geopolitical risk and export controls. US restrictions on advanced GPU and chip technology exports to China have forced ByteDance to shift a larger portion of its budget toward domestic and controllable silicon sources .

Unprecedented AI infrastructure scale. ByteDance raised its 2026 capital expenditure plan to more than 200 billion yuan (approximately $30 billion), an increase of at least 25 percent over an earlier 160-billion-yuan draft. A significant slice of that budget is allocated to proprietary chip development .

Challenges and Realities

Designing server-class CPUs from scratch is notoriously difficult and capital-intensive, and doing so on two architectures doubles the engineering complexity .

Foundry access remains a critical bottleneck. Samsung is a potential manufacturing partner for ByteDance's SeedChip ASIC, but the availability of advanced process nodes for Chinese-designed chips remains geopolitically sensitive. Discussions between ByteDance and Samsung also include access to memory chip supply, underscoring how integrated the supply-chain challenge has become .

Even if ByteDance's custom CPUs succeed, they will complement rather than fully replace merchant silicon for years. ByteDance still plans to spend roughly $14 billion on Nvidia chips in 2026 alone, a figure that highlights just how deeply embedded Nvidia's GPUs are in the company's training and inference workflows .

Where ByteDance Fits in the Bigger Picture

ByteDance's move mirrors a well-established hyperscaler playbook. Amazon built its Graviton Arm-based CPUs and Trainium/Inferentia accelerators. Google developed its TPUs. Microsoft created the Cobalt CPU. In each case, the goal was the same: optimize cost and performance at scale by controlling the silicon .

What makes ByteDance's approach distinctive is its multi-architecture bet and the urgency driven by geopolitics. While US hyperscalers gravitated toward Arm for their custom CPUs, ByteDance is keeping RISC-V as a parallel track—partly for architectural flexibility, partly as a hedge against future restrictions on proprietary chip IP .

The broader market is moving in the same direction. Custom ASIC-based AI server shipments are projected to capture 27.8 percent of the market in 2026, with shipments growing 44.6 percent year-over-year—nearly three times the growth rate forecast for merchant GPUs .

ByteDance is not relying solely on in-house designs to diversify its silicon supply. In late May 2026, the company struck a deal with Qualcomm to purchase millions of application-specific integrated circuits (ASICs) for running AI agent software, making it one of the first major customers for Qualcomm's AI-focused data-center silicon . Combined with discussions around domestic Chinese AI chips and the SeedChip ASIC, ByteDance is assembling a multi-pronged chip strategy that spans in-house CPUs, custom accelerators, and strategic partnerships.

The Long Road Ahead

Custom CPU programs take years to move from design to volume deployment. ByteDance's dual Arm and RISC-V tracks are still in the evaluation stage, and the company has not publicly disclosed a timeline for production silicon. In the near term, the company's AI infrastructure will continue to rely heavily on Nvidia GPUs and a growing fleet of partner ASICs. Over the medium to long term, however, ByteDance's custom silicon investments could reshape its cost structure and give it the kind of hardware independence that only the largest hyperscalers have achieved so far.