中國開發者在開放權重AI模型方面的競爭力顯著提升，與全球領先者的差距不斷縮小，同時保持低價優勢。阿里巴巴最新的Qwen3.8-Max模型已躋身Arena基準測試中頂尖的編程模型行列。Silicon Data的混合推理價格指數較五月峰值下降了36%，更低的運算成本改善了搜尋、廣告、電商及客戶服務等業務的經濟效益，也使新的AI應用在商業上變得可行。
儘管前景樂觀，仍需保持謹慎。中國科技業長期以來偏重規模而非盈利，AI領域或不例外。互聯網巨頭面臨來自DeepSeek、Z AI、Moonshot及字節跳動等對手的激烈競爭，同時阿里巴巴、美團和京東在外賣業務上持續價格戰，國內消費需求也依然疲軟。不過目前投資者認為，中國雲端和互聯網股票作為AI成本下降的受益者，重新估值才剛剛開始。
China's rapid advances in low-cost artificial intelligence are bolstering the investment case for its internet giants after years of underperformance relative to chipmakers. Investors increasingly believe that as AI costs fall, industry value will shift toward platform companies owning everyday applications such as search, e-commerce, advertising, and enterprise software. Since late June, the KraneShares China Internet ETF has surged over 20%, roughly doubling the Hang Seng Index's gain, while the Philadelphia Semiconductor Index declined about 11%.
Chinese developers have narrowed the gap with global AI leaders in open-weight models while maintaining competitive pricing. Alibaba's Qwen3.8-Max now ranks among the top coding models on Arena's benchmark. Silicon Data's blended inference-price index has dropped 36% from its May peak, improving the economics of AI-powered services and making new applications commercially viable. Goldman Sachs highlights accelerating token growth and rising capex as reasons to favor China's cloud and data-center sectors, estimating Alibaba Cloud grew 45% in the June quarter.
Caution remains warranted, however. China's tech sector has historically prioritized scale over profitability, and AI may follow the same pattern. The internet giants face mounting competition from rivals such as DeepSeek, Z AI, Moonshot, and ByteDance, while Alibaba, Meituan, and JD.com continue battling a price war in food delivery amid sluggish domestic consumer demand. Nonetheless, investors believe the rotation has room to run as China's cloud and internet stocks are only beginning to be re-rated as beneficiaries of declining AI costs, supported by firm GPU rental rates signaling sustained compute demand.