ByteDance reportedly develops AI chip, in talks with Samsung

ByteDance is reportedly developing an AI chip and is in talks with Samsung to manufacture it, according to sources familiar with the matter.

The company plans to produce at least 100,000 units this year, aiming to ramp up to 350,000, mainly for AI inference.

Negotiations include access to scarce memory chip supplies amid global shortages.

ByteDance aims to receive sample chips by end-March.

Its chip efforts began around 2022 as part of a broader AI strategy.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

ByteDance plans about $23 billion in AI capital spending, pushed by heavy in-house use. Analysts cited in a Financial Times-sourced report say its services burn nearly as many daily computing tokens as Google 1.
ByteDance executives have said AI bills are too high, so the company wants cheaper, more efficient computing. That goal sits behind a self-designed chip tuned for its AI inference workloads 2.
Samsung offers more than foundry work, since it can fabricate chips designed by others and help secure related components.
Talks include access to tight memory chip supplies that AI systems rely on.
Samsung is increasing capacity for High Bandwidth Memory (HBM)4, a next-generation memory used in next-generation AI accelerators. That could boost its role in AI chips, though forecasts that it will lead the market remain projections rather than confirmed outcomes 3, 4.

The chip effort responds to cost pressure, plus US-China tech rivalry.
ByteDance spends billions leasing overseas data centres to legally use Nvidia’s most advanced hardware for training AI models, plus serving customers outside China. The spending underscores limits on its hardware access 1.
With in-house silicon (internally designed chips) in products like the Doubao chatbot, ByteDance can rely less on foreign technology while tuning performance for its own services 1.
Similar work by Alibaba and Baidu speeds up a separate Chinese AI hardware stack, which may offer options beyond Nvidia and split today’s largely unified AI infrastructure market.