China’s tech giants and AI startups are locked in an escalating price war over AI tokens — the foundational unit of AI processing — and it is reshaping the economics of the country’s entire artificial intelligence industry. Triggered by DeepSeek, the Hangzhou-based AI lab that forced competitors to slash prices by up to 97% in 2025, the battle has now expanded from consumer chatbot pricing into the lucrative enterprise market, where billions of dollars in cloud computing revenue are at stake.
Why It Matters
For foreign companies building AI products or using AI infrastructure in China, the token price war is both an opportunity and a risk. On the opportunity side, AI compute has never been cheaper in China: the cost of processing 1 million input tokens on leading Chinese models has dropped from RMB 14.5 in early 2025 to as low as RMB 0.5 in mid-2026 on some platforms. That is a 97% price collapse in 18 months. If your business trains models, runs inference at scale, or builds AI applications for the Chinese market, your compute bill just got smaller by an order of magnitude.
On the risk side, the price war is concentrating power among a handful of well-funded players. DeepSeek raised $7.4 billion in a funding round announced in July 2026. Tencent launched its Hunyuan 3 model with a free AI-agent feature embedded in its Yuanbao assistant, using free compute as a user-acquisition strategy. Kuaishou secured nearly $3 billion for its AI video unit, Kling AI. Companies that can afford to subsidize compute are building moats that smaller competitors — including foreign entrants — will struggle to cross.
The Details
The AI token — a unit that measures the amount of text an AI model processes — has become the commodity that underpins China’s AI economy. Every ChatGPT-like query, every document summary, every code generation request consumes tokens. And in China, the price per token has fallen faster than anywhere else in the world.
DeepSeek triggered the collapse when it released its V3 model in late 2024 with pricing at roughly 1/20th of OpenAI’s comparable tier. Chinese competitors — Alibaba’s Tongyi Qianwen, ByteDance’s Doubao, Baidu’s ERNIE, and Tencent’s Hunyuan — matched or undercut DeepSeek within months. By May 2025, the price-per-token on China’s leading models had converged to near-zero for consumer use, with providers monetizing through enterprise API volume, cloud bundling, and premium features rather than per-token fees.
Now the war has moved to the enterprise. Caixin’s July 2026 cover story reports that explosive demand for AI compute has forced tech giants and startups into a three-front battle: pricing, computing power, and the lucrative business market. The computing power front is particularly intense. China’s GPU supply remains constrained by U.S. export controls, creating a two-tier market where companies with existing chip stockpiles (ByteDance, Alibaba) can offer lower prices than those dependent on domestic alternatives (Huawei’s Ascend chips). ByteDance is estimated to hold over 100,000 NVIDIA H800 GPUs, giving it a compute cost advantage that rivals cannot easily replicate.
DeepSeek’s $7.4 billion funding round — one of the largest AI fundraises in Chinese history — is explicitly aimed at closing the compute gap. The company plans a major hiring spree and expanded cloud infrastructure. Combined with Tencent’s move to bundle free AI agents into Yuanbao and Kuaishou’s $3 billion AI video bet, the total capital committed to China’s AI infrastructure in the first half of 2026 alone exceeds $14 billion.
What You Should Do
- Re-evaluate your AI compute budget. If your China-based operations use any major LLM API — for customer service chatbots, document processing, marketing content generation, or internal analytics — benchmark current token pricing across providers. The cost difference between the most and least expensive Chinese LLM API for equivalent quality can be 5-10x. Renegotiate contracts based on current market rates, not 2025 pricing.
- Watch concentration risk. If you build your product on a single AI provider’s API, that provider may not survive the price war. The market is consolidating toward 4-5 well-funded players (DeepSeek, ByteDance, Alibaba, Tencent, Baidu). Smaller AI startups without billion-dollar war chests may exit or get acquired. Diversify your API dependencies or build model-agnostic architecture now.
- Track the GPU bottleneck. U.S. export controls on advanced chips remain the binding constraint on China’s AI scale-up. If those controls tighten further — as proposed in the latest U.S. AI diffusion rules — Chinese compute prices could reverse and rise. Foreign companies should maintain at least one non-China compute option for critical workloads as a hedge.
- Look for enterprise-AI entry points. As Chinese AI companies compete on price, they are hungry for enterprise use cases that demonstrate value beyond consumer chatbots. If your company has deep industry expertise — in manufacturing, logistics, healthcare, or finance — pairing that domain knowledge with cheap Chinese AI compute is a genuine market entry strategy.
One Data Point
Remote China market entry support, built around execution.
