ByteDance’s Doubao AI model usage rises over 10x

ByteDance reported that usage of its Doubao large language models rose over 10x in 2025, driven by increased enterprise adoption in China.

The company said average daily token usage for Doubao exceeded 50 trillion this month, up from 4 trillion in December 2024.

ByteDance, based in Beijing, is known for TikTok and operates the Volcano Engine cloud platform.

Volcano Engine president Tan Dai said more than 100 corporate clients have consumed over 1 trillion tokens in total, though no specific clients were named.

IDC data cited by the company indicated Volcano Engine accounted for about 49% of China’s AI model-focused public cloud services in 2025.

Major clients include Mercedes-Benz, BYD, China Merchants Bank, and the education platform Yuanfudao.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Missing revenue clarity behind the 10x usage surge

ByteDance says usage hit 50 trillion daily tokens with 10x growth, yet gives no split between paid enterprise demand and free or in-house traffic [Article]. Pricing differs, as Doubao’s coding agent (an AI assistant for programming) runs 9.9 yuan (~$1.30) on promo or 40 yuan monthly standard 1, about 62.7% below China’s industry average 2. Tokens are chunks of text an AI model reads or writes when processing requests.
It also says 100+ corporate customers used over 1 trillion tokens, yet gives no per-token price or revenue mix [Article]. Without the share of that traffic that earns money versus subsidized promo or internal use, monetization stays murky.

API compatibility creates migration opportunities for developers

Doubao models work with OpenAI API protocols 3. The Seed-Code model (a coding-focused Doubao variant) also supports Anthropic’s API 4 2. Anthropic, a U.S. AI company behind the Claude assistant, changed service rules in September to block Claude for subsidiaries of Chinese firms 1. Developers now seek options.
Tool vendors and API-integration platforms can sell migration help to move Chinese teams from restricted Western AI APIs to Doubao. Doubao offers a 256K token context window (the amount of text the model can consider at once) 4. It posts a 78.8% SWE-Bench score (a software-engineering benchmark for automated bug fixing), on par with Anthropic’s Claude Sonnet 2, which gives coding parity at lower cost.