ByteDance is reportedly designing its own AI chips with TSMC

Waiting in line for over $2 billion worth of throttled graphics cards will make any tech giant reconsider its hardware strategy.

TikTok owner ByteDance is reportedly working with Taiwan Semiconductor Manufacturing Company (TSMC) to design two custom AI chips—one for training and another for inference—aiming for mass production by 2026, according to The Information.

The initiative is still in the design phase and plans are not finalized, but the motivation is obvious: weaning itself off pricey Nvidia GPUs and dodging tightening U.S. trade restrictions.

A $2 billion backlog sparks a DIY push

ByteDance has not been shy about spending cash on standard hardware. The company reportedly ordered more than 200,000 of Nvidia's H20 chips—a GPU tailored specifically to skirt U.S. export controls.

That order is worth more than $2 billion, and ByteDance is still waiting for delivery. The Information estimates that shifting to in-house silicon manufactured by TSMC could eventually save the company billions of dollars.

ByteDance is not the only Chinese player taking the custom route. Baidu has designed its own Kunlun 3 AI chip, which is also scheduled to enter production at TSMC soon.

Doubao's boom demands cheaper compute

ByteDance's hardware urgency is fueled by the explosive growth of its consumer and enterprise AI products.

Last August, the company launched its Doubao chatbot, which quickly surpassed Baidu's Ernie Bot in downloads and iOS monthly active users inside China, according to Bloomberg data.

In May, ByteDance rolled out enterprise-grade Doubao foundation models at aggressively low price points, triggering a domestic price war that forced rivals to slash their model fees by as much as 97%.

However, turning custom blueprints into physical chips remains precarious. Earlier this month, Washington unveiled another round of export controls targeting critical technologies, with rules updated on an annual basis. Chinese firms can only manufacture at TSMC as long as their designs stay within those moving regulatory boundaries.

Why it matters

For enterprise developers relying on Chinese cloud APIs, in-house inference chips could make hyper-cheap model pricing sustainable. ByteDance triggered a massive race to the bottom this year, but subsidizing API calls on overpriced Nvidia hardware is financially brutal over the long run.

Across the broader market, this underscores how the battleground for Chinese tech champions has shifted from raw software features to physical hardware independence. Relying on modified, lower-spec Nvidia chips leaves domestic champions exposed to sudden regulatory bans.

We think ByteDance's push into both training and inference silicon is ambitious, but relying on TSMC as the fabrication partner represents a glaring vulnerability. Designing competitive accelerators in-house is hard enough, but your production line remaining entirely dependent on annual U.S. policy revisions makes 2026 feel like an eternity away.