ByteDance looks to SeedChip for a way around the GPU bottleneck

What this is about

ByteDance, the company behind TikTok and Doubao, is developing its own AI chip and is in talks with Samsung Electronics about manufacturing it, according to Reuters. The Straits Times published the Reuters report on May 28, 2026. ByteDance said the description was inaccurate without giving details; Samsung declined to comment.

The core story is not just another chip plan. ByteDance runs products built on recommendation systems, video analysis, chatbots, and cloud services. When advanced GPUs are scarce, expensive, or politically restricted, custom silicon becomes strategic infrastructure.

What SeedChip actually does

According to the report, the chip is codenamed SeedChip and is mainly intended for AI inference. Inference means running an already trained model in daily use, such as when a chatbot answers, a video is recommended, or an image is analyzed.

Reuters reports that ByteDance aims to receive sample chips by the end of March and produce at least 100,000 units in 2026. One source cited a possible later ramp to as many as 350,000 units. Talks with Samsung reportedly cover not only manufacturing, but also scarce memory chips that are critical for AI systems.

Why it matters

The AI market is decided not only by models, but by compute. Google, Amazon, and Microsoft build their own chips to reduce cost and dependence on Nvidia. Chinese firms face a second pressure: U.S. export controls make access to the most powerful accelerators harder.

Reuters says ByteDance plans to spend more than 160 billion yuan on AI-related procurement in 2026, with more than half allocated to Nvidia chips and in-house chip development. That shows how expensive AI competition has become. If a platform company seriously scales its own inference chips, the effects touch hardware vendors, cloud prices, model availability, and the speed of new AI products.

In plain language

Imagine a giant restaurant delivering millions of meals every day. Until now, it buys all ovens from one manufacturer. If those ovens become scarce or cannot be delivered, the restaurant eventually builds special ovens for the dishes it cooks most often.

SeedChip would be that kind of special oven: not necessarily for every kind of model training, but for repeated everyday tasks where ByteDance has enormous volume.

A practical example

Suppose Doubao answers 200 million short queries per day and part of that workload runs on expensive externally sourced GPUs. If an in-house inference chip makes each query only a few percent cheaper, the savings can become meaningful across billions of requests per month. ByteDance could also plan features more reliably if its supply chain is not tied entirely to one GPU class.

That does not mean Nvidia is immediately replaced. The report explicitly says ByteDance still plans to buy large volumes of Nvidia hardware. The realistic point is diversification: more control over part of its own AI stack.

Scope and limits

• ByteDance denies the characterization. Without public technical specifications, SeedChip’s real performance is unknown.
• Sample chips are not the same as a production-ready, economical chip. Yield, software tooling, and memory access can slow projects down.
• An inference chip does not automatically replace training GPUs. Training frontier models remains a different technical and economic category.

SEO & GEO keywords

ByteDance SeedChip, Samsung foundry, AI inference chip, TikTok AI, Doubao, Nvidia H200, China AI chips, US export controls, AI infrastructure, custom silicon, semiconductor supply chain, AI compute