Baidu’s margin expansion via custom silicon
Baidu’s AI cloud revenue grew 184% year-over-year during its Q1 2026 earnings call. This growth brought AI cloud revenue to RMB 8.8 billion ($1.3 billion), which is over half of the company’s total quarterly revenue of RMB 26 billion ($3.8 billion). Without the surge in AI-related sales, Baidu’s total quarterly revenue would have declined. CEO Yanhong Li says this growth is above broader market trends. CFO Haijian He states that Baidu’s GPU cloud services provide higher margins than traditional CPU cloud offerings because of demand, supply chain constraints, and increased pricing power. Baidu uses its Kunlunxin AI chips to optimize costs and improve its customer mix, which supports margin expansion. This chip development helps as enterprises move beyond model training into inference workloads. Dou Shen, president of Baidu’s AI Group, says enterprises prioritize stability at scale, compatibility with mainstream models, and support for large-scale cluster deployment over peak chip performance. Customers care more about migration costs and friction than pure performance. Baidu continues to deploy its Apollo Go robotaxis and expand its Digital Human business to integrate these services into everyday life.
Alibaba slashes Qwen pricing
Alibaba Cloud reduces prices on its visual language model Qwen-VL by up to 85% to win more business from the enterprise segment, which is a demonstration of how competition among China’s technology giants to win more business for their nascent artificial intelligence products is intensifying. The company previously reduced prices on its Qwen AI model by as much as 97% in May to boost demand. Alibaba has deployed Qwen models to over 90,000 enterprise users. Alibaba was named an Emerging Leader of Generative AI Model Providers in the 2025 Gartner Innovation Guide. The Qwen family includes over 100 open-weight models, such as the Qwen3-Omni multimodal model, the Qwen-Audio model, and the Qwen3-Coder variant. The Qwen3-Omni model accepts text, images, video, and audio as input and generates both text and speech in real time. Qwen models also include the Qwen-Image foundation model for image generation and the QwQ model for reasoning tasks. Qwen3 models use 36 trillion tokens and support 119 languages and dialects. You should watch how these price cuts among Chinese tech giants like Alibaba, Tencent, and ByteDance change the market. DeepSeek provides a different competitive pressure, having released its R1 reasoning model in January 2025. DeepSeek’s V3 was trained for approximately $5.6 million in compute, which is much less than the $100 million estimated for comparable Western models. DeepSeek-R1’s reinforcement learning phase cost an additional $294,000, according to a peer-reviewed Nature paper published in September 2025. DeepSeek uses a Mixture-of-Experts architecture with 671 billion total parameters but only 37 billion active at a time. DeepSeek’s R1 API was 27x cheaper than OpenAI’s o1 model at launch, and by early 2026, the cost gap grew to over 140x. DeepSeek remains available under the MIT license for free use and self-hosting. DeepSeek tends to be computationally slower than its rivals.
API pricing for Qwen models
Alibaba’s pricing for models in the Singapore region shows different costs between different generations. Qwen3.8-Max lists at $2 input and $6 output, which results in a $3.00 blended price. This is a 1.07x increase over the older Qwen-Max model, which has a $2.80 blended price. Qwen3.7-Max lists at $2.50 input and $7.50 output, creating a $3.75 blended price. The Qwen3.7-Plus model lists at $0.40 input and $1.60 output, which is a $0.70 blended price. Qwen3.5-Plus lists at $0.40 input and $2.40 output, which is a $0.60 blended price. Qwen-plus lists at $0.40 input and $1.20 output, a $0.60 blended price.
| Model | Region | Input (per 1M) | Output (per 1M) | Blended Price |
|---|---|---|---|---|
| Qwen-Max | Singapore | $1.60 | $6.40 | $2.80 |
| Qwen3.8-Max | Singapore | $2.00 | $6.00 | $3.00 |
| Qwen3.7-Max | Singapore | $2.50 | $7.50 | $3.75 |
| Qwen-plus | Singapore | $0.40 | $1.20 | $0.60 |
| Qwen3.7-Plus | Singapore | $0.40 | $1.60 | $0.70 |
In the Beijing region, the price gap is larger. Qwen-Max lists at $0.345 input and $1.377 output, whereas Qwen3.8-Max lists at $1.65 input and $4.951 output. This is a 4.1x blended gap between the two models in China. DeepSeek’s V4.1 Flash off-peak lists at $0.15 input and $0.60 output, a $0.263 blended price. Kimi K3.1 Pro lists at $2 input and $10 output, which is a $4.00 blended price. The Qwen3-Omni model remains competitive across the board, yet will the extreme price reductions on older models permanently erode the value of the latest flagship releases?
