Baidu: DeepSeek Has Implications for Chinese Companies

Key Morningstar Metrics for Baidu

• Fair Value Estimate: $157.00
• Morningstar Rating: ★★★★★
• Morningstar Economic Moat Rating: Wide
• Morningstar Uncertainty Rating: High

Based on DeepSeek’s initial claims about lowering computing costs to less than 5% of ChatGPT’s costs while keeping similar performance for its reasoning model R1, we believe this could be a wakeup call not just for US-based hyperscalers, but also for Chinese artificial intelligence firms such as Baidu BIDU. While it doesn’t render Ernie, Baidu’s AI model, irrelevant, R1′s freely distributed format makes monetizing Baidu’s AI business harder in the personal cloud segment. Our fair value estimate for Baidu is unchanged at $157 per share. We haven’t factored in much growth from the personal AI cloud business, and DeepSeek shouldn’t affect our valuation.

While we still have questions about the specific costs of DeepSeek, we believe R1′s open-source format is driving innovation and will force companies to rethink how to best monetize AI or capital expenditure. Ernie is a closed-source model like ChatGPT and requires time to train and build out. Currently, Baidu hasn’t monetized personal AI materially, as most of its AI revenue still comes from the enterprise cloud or advertising. Therefore, we don’t believe DeepSeek will hit near-term revenue, but its introduction to the industry could lower Baidu’s total addressable market, should DeepSeek’s claims about its costs be true.

CEO Robin Li said in April 2024 that closed-source models have better reasoning power and cost advantages. We’re not trying to make a technical case for either, but we believe Ernie may be behind, given that other models are either making headlines or are in partnership talks. While we’re uncertain who the eventual winners will be, we believe that large language models are becoming commoditized, given that ByteDance, Tencent TCEHY, and Alibaba BABA have their own models, not to mention other open-source models such as Zhipu, which is also built off Llama and uses the same language as DeepSeek.