Marzia Briel

SemiAnalysis published an analysis on DeepSeek, addressing recent claims about its cost and performance. $NVDA The report states that the widely circulated $6M training cost for DeepSeek V3 is incorrect, as it only accounts for GPU pre-training expenses and excludes R&D, infrastructure, and other critical costs. According to their findings, DeepSeek’s total server CapEx is around $1.3B, with a significant portion allocated to maintaining and operating its GPU clusters. The report also states that DeepSeek has access to roughly 50,000 Hopper GPUs, but clarifies that this does not mean 50,000 H100s, as some have suggested. Instead, it’s a mix of H800s, H100s, and the China-specific H20s, which NVIDIA has been producing in response to U.S. export restrictions. SemiAnalysis points out that DeepSeek operates its own datacenters and has a more streamlined structure compared to larger AI labs. On performance, the report notes that R1 matches OpenAI’s o1 in reasoning tasks but is not the clear leader across all metrics. It also highlights that while DeepSeek has gained attention for its pricing and efficiency, Google’s Gemini Flash 2.0 is similarly capable and even cheaper when accessed through API. A key innovation cited is Multi-Head Latent Attention (MLA), which significantly reduces inference costs by cutting KV cache usage by 93.3%. The report suggests that any improvements DeepSeek makes will likely be adopted by Western AI labs almost immediately. SemiAnalysis also mentions that costs could fall another 5x by the end of the year, and that DeepSeek’s structure allows it to move quickly compared to larger, more bureaucratic AI labs. However, it notes that scaling up in the face of tightening U.S. export controls remains a challenge.

Jeremy joining protesters today against the new inheritance tax regulations for farmers