中国AI模型美国市场占比飙升至46%，DeepSeek V4峰谷定价7月15日上线 | Chinese AI Models Capture 46% of US Market, DeepSeek V4 Peak-Valley Pricing Launches July 15
中国AI模型美国市场占比飙升至46%，DeepSeek V4峰谷定价7月15日上线
据OpenRouter统计，美国企业调用中国AI模型的Token占比已从2025年上半年的4.5%飙升至2026年7月的46%，不到一年翻了十倍。同日，DeepSeek公布V4正式版定价细则，引入行业首创的”峰谷机制”。
十倍增长背后的逻辑
4.5%到46%，这个数字震惊了整个行业。超过80%的美国AI初创企业正在使用中国开源模型。原因很简单：
- 性能差距缩小：中国头部开源模型与OpenAI、Anthropic顶尖模型的技术差距约6-9个月，对大多数应用场景来说已经够用
- 价格碾压：定价低60%-90%，在token消耗量大的场景下，这意味着每月节省数千甚至数万美元
- 开源优势：可自部署、可微调、数据不出企业——这对有合规要求的美国企业来说反而是一个加分项
智谱GLM-5.2上线首周日均Token调用量增长27倍、客户数增长80倍。DeepSeek-V4-Flash连续七周位居全球调用量榜首。这些数字说明，中国开源模型正凭借极致性价比大规模替代付费闭源模型。
DeepSeek V4：峰谷定价的行业首创
DeepSeek于7月8日公布V4正式版定价细则，7月15日全量上线。核心数据：
- 上下文：原生128K，可扩展至1M
- 推理吞吐：较V3提升约40%
- 单位token成本：下降30%
最引人注目的是”峰谷机制”：
- 高峰时段（工作日9:00-12:00、14:00-18:00）：API价格翻倍
- 低谷时段（夜间+周末）：输入价低至0.5元/百万tokens
这是一种类似电力市场的定价策略。通过价格杠杆引导用户错峰使用，平滑算力负载。对于非实时任务（如批量数据处理、夜间训练），低谷价格极具吸引力。
0.5元/百万tokens意味着什么
对比一下主流模型的价格：
- GPT-5.6 Luna：1美元/百万tokens（约7.2元）
- DeepSeek V4低谷价：0.5元/百万tokens
- 差距：14.4倍
即使对比国产模型，这个价格也是地板级别。DeepSeek显然在用亏本价抢占市场份额，赌的是规模效应最终摊薄成本。
全球AI市场格局重塑
中国AI模型在美国市场的崛起不仅仅是商业竞争——它在重塑全球AI供应链。当美国企业开始依赖中国开源模型，技术脱钩的逻辑就出现了裂缝。
但这个趋势也面临不确定性。美国政府对AI技术的出口管制正在加强，如果中国开源模型被纳入管制范围，46%的市占率可能是一个不可持续的高点。
下周展望
7月15日DeepSeek V4上线，7月17日谷歌Gemini 3.5 Pro发布。三款旗舰模型（GPT-5.6、DeepSeek V4、Gemini 3.5 Pro）将在同一周内正面交锋。大模型价格/能力战进入白热化阶段。
Chinese AI Models Capture 46% of US Market, DeepSeek V4 Peak-Valley Pricing Launches July 15
According to OpenRouter, the token share of Chinese AI models called by U.S. enterprises has surged from 4.5% in H1 2025 to 46% in July 2026 — a tenfold increase in less than a year. On the same day, DeepSeek announced V4 pricing details with an industry-first “peak-valley mechanism.”
The Logic Behind 10x Growth
From 4.5% to 46% — this number has shocked the entire industry. Over 80% of U.S. AI startups are using Chinese open-source models. The reasons are simple: the performance gap between top Chinese open-source models and frontier models from OpenAI/Anthropic is only 6-9 months, which is sufficient for most use cases; pricing is 60-90% lower; and open-source allows self-deployment and fine-tuning, which is actually a plus for U.S. companies with compliance requirements.
Zhipu GLM-5.2 saw daily token usage grow 27x and customer count grow 80x in its first week. DeepSeek-V4-Flash has topped global usage for seven consecutive weeks.
DeepSeek V4: Industry-First Peak-Valley Pricing
DeepSeek V4 launches July 15 with native 128K context (extendable to 1M), 40% throughput improvement over V3, and 30% lower per-token cost. The most notable feature is the “peak-valley mechanism”: peak hours (weekdays 9:00-12:00, 14:00-18:00) double the API price, while off-peak hours (nights + weekends) drop input price to 0.5 RMB per million tokens.
This is similar to electricity market pricing — using price leverage to shift usage to off-peak hours and smooth compute load. For non-real-time tasks like batch data processing, the off-peak price is extremely attractive.
At 0.5 RMB/million tokens, DeepSeek is 14.4x cheaper than GPT-5.6 Luna’s $1/million tokens. DeepSeek is clearly subsidizing market share with below-cost pricing, betting that scale will eventually reduce costs.
Reshaping the Global AI Market
The rise of Chinese AI models in the U.S. market isn’t just commercial competition — it’s reshaping the global AI supply chain. When U.S. companies start depending on Chinese open-source models, the logic of tech decoupling develops cracks. But this trend faces uncertainty: if U.S. export controls extend to Chinese open-source models, 46% market share may be an unsustainable peak.