Deepseek疑虛報實際成本, Nvidia股價漲 9%

摩根大通分析師Harlan Sur 和花旗分析師Christopher Danley 在分別給投資者的報告中表示，由於DeepSeek 使用了一種名為「蒸餾」的過程——換句話說，它依賴Meta (META) 的開源Llama AI 模型來開發其模型——低這家中國新創公司提到的支出（用於訓練其最新 V3 模型的支出不到 60 億美元）並未完全包含其成本。

蘇爾寫道：“我們認為，在得出結論之前驗證這些成本至關重要。”

Danley 補充道：“鑑於 Deepseek 是基於雲端服務供應商 [Meta] 的利用，而人工智慧仍處於起步階段，我們傾向於人工智慧支出持續強勁增長的論點。”

即便如此，Raymond James 分析師Srini Pajjuri 在周一給投資者的報告中寫道，DeepSeek「顯然無法獲得像美國超大規模企業那樣多的計算能力，並且以某種方式設法開發出一種看起來極具競爭力的模型」。

JPMorgan analyst Harlan Sur and Citi analyst Christopher Danley said in separate notes to investors that because DeepSeek used a process called “distillation” — in other words, it relied on Meta’s (META) open-source Llama AI model to develop its model — the low spending cited by the Chinese startup (under $6 billion to train its recent V3 model) did not fully encompass its costs.

“We believe it is crucial to validate these costs before drawing conclusions,” Sur wrote.

Danley added: “Given Deepseek is based on leveraging cloud service providers [Meta] and AI is still in its infancy, we lean towards the argument of continued strong growth in AI spending.”

Even so, DeepSeek “clearly doesn’t have access to as much compute as US hyperscalers and somehow managed to develop a model that appears highly competitive,” Raymond James analyst Srini Pajjuri wrote in a note to investors Monday.


https://finance.yahoo.com/news/nvidia-stock-begins-recovery-after-deepseek-ai-frenzy-prompted-near-600-billion-loss-134240328.html

---------------------------