ByteDance reportedly jointly developing Groq-like AI inference chips with Innostar
Taylor Wilson
ByteDance is teaming up with InnoStar to build a Groq-style inference chip that sidesteps the HBM shortage — its clearest move yet from AI apps into foundational chip hardware.
What problem is this chip trying to solve?
ByteDance and InnoStar are co-developing a Language Processing Unit (LPU) — a chip purpose-built for AI inference — modeled on Groq's architecture.
Unlike GPUs, an LPU integrates compute and memory on the same die using a programmable pipeline. This means → faster inference, lower overhead.
In plain terms = a GPU is a Swiss Army knife; an LPU is a blade built to do one job — inference — and do it faster.
Why InnoStar, and where is the money coming from?
InnoStar is a Chinese integrated device manufacturer specializing in Resistive RAM (ReRAM), a next-generation memory technology.
The company is raising $400 million at a pre-money valuation of roughly $1.5 billion. Investors include ByteDance and Yunfeng Capital (co-founded by Alibaba's Jack Ma).
Ant Group and several state-backed funds have also taken stakes. This means → the project carries dual backing from Big Tech capital and state capital.
Skipping HBM — technical choice or pragmatic workaround?
The chip design reportedly will not use High Bandwidth Memory (HBM) — the dominant high-speed memory in today's AI chips.
The direct reason: a global HBM shortage is driving costs up. Bypassing it reduces supply-chain risk.
This reflects a broader pattern — Chinese AI chip ventures are actively pursuing alternative architectures under sanctions and supply-chain pressure, rather than competing head-on for constrained components.
How far has ByteDance's AI infrastructure push gone?
Its AI chatbot Doubao (豆包) has passed 300 million monthly active users, making it one of China's most popular AI apps.
ByteDance is simultaneously developing its own CPUs for in-house data centers and is considering up to $70 billion in AI infrastructure capex for 2026.
This means → ByteDance's logic is clear: apps → chips → data centers — it wants control at every layer of the stack.
What can we conclude right now?
The project is still at an early stage. Production timelines and foundry partners remain undecided.
Groq's LPU architecture was licensed by Nvidia in late 2025, signaling industry-level validation of the approach.
In plain terms = the direction checks out and the funding is in place, but mass production is a long way off — think "blueprints drawn, ground not yet broken."
Content is for reference only, not financial advice.