ByteDance Develops Groq-Inspired AI Inference Chip

TikTok parent ByteDance is racing ahead with its own custom artificial intelligence hardware, developing a new inference-focused chip modeled after the architecture of Groq’s specialized Language Processing Units (LPUs), according to multiple reports citing industry sources.

Credit: ByteDance

The move underscores the Chinese tech giant’s determination to reduce reliance on U.S. chipmakers amid ongoing export restrictions and skyrocketing demand for AI computing power.

It forms part of a broader, multi-billion-dollar push to secure its AI infrastructure as the company prepares for a massive expansion of agent-based AI products.

Sources told The Information that ByteDance is partnering with Chinese chipmaker InnoStar Semiconductor on the project.

The new chip draws inspiration from Groq’s LPUs, which are purpose-built for AI inference (the process of running trained models to generate responses) rather than the more computationally intensive training phase.

Groq’s architecture emphasizes deterministic, high-throughput processing optimized for low latency and cost efficiency, making it particularly attractive for high-volume applications like recommendation engines and conversational AI.

ByteDance, which powers platforms including Douyin (the Chinese version of TikTok) and its Coze AI agent platform, faces enormous inference demands from hundreds of millions of daily users.

The company has reportedly invested in InnoStar and is leveraging the partner’s expertise in memory technologies, such as Resistive Random-Access Memory (RRAM), to sidestep restrictions on high-bandwidth memory (HBM) chips that are critical for many advanced AI systems but subject to U.S. export controls.

Fabrication is expected to occur at more mature process nodes, such as TSMC’s 28nm, which face fewer restrictions.

This inference chip project is not an isolated effort. Earlier this week, Reuters reported that ByteDance is also developing its own custom central processing units (CPUs) to support its AI rollout, with plans to deploy them in its data centers and servers.

The company is exploring both Arm-based and open-source RISC-V designs.

ByteDance reportedly unveiled a new central processor and AI system incorporating Groq-derived technology as early as March.

With a chip design team now numbering around 1,000 people, the company is pursuing multiple parallel silicon initiatives as it weighs capital expenditures of up to $70 billion in 2026 for AI infrastructure.

The push comes as surging global chip prices and persistent supply shortages have complicated expansion plans for even the largest tech players.

By investing in proprietary hardware, ByteDance aims to gain greater control over costs and supply chains while aligning with Beijing’s broader push for technological self-sufficiency.

The development arrives months after Nvidia acquired Groq in a roughly $20 billion deal in late 2025.

Notably, that transaction included non-exclusive licensing of Groq’s inference technology, allowing other companies, including potential competitors, to build upon the architecture.

ByteDance has also struck major deals with other chip suppliers.

In recent weeks, it reached an agreement with Qualcomm to purchase millions of custom AI data center chips, marking a significant win for the U.S. firm’s expansion into AI infrastructure.

While ByteDance continues to purchase Nvidia hardware where possible, homegrown alternatives provide insurance against tightening U.S. export controls and help address the unique economics of its inference-heavy workloads.

The new inference chip remains in early development stages, and commercial deployment timelines remain unclear.

Transitioning from prototypes to large-scale production, particularly with novel memory technologies like RRAM, could take time and faces technical hurdles.

Nevertheless, the initiative highlights ByteDance’s transformation from a social media and content powerhouse into a formidable player in the global AI hardware race.

As AI agents and real-time generative applications proliferate, the ability to run models efficiently and affordably at scale could prove a decisive competitive advantage.

ByteDance has previously emphasized its commitment to investing heavily in AI research and infrastructure to drive innovation across its platforms.

This story is developing. Updates will follow as more details emerge on timelines, performance specifications, and deployment plans.