· 供应链分析

AI算力需求指数级增长抵消GPU迭代贬值,NVDA客户优质,非泡沫崩盘。

涉及标的:

中文翻译

答案很微妙。 主要看两个因素: 1. GPU 变得更节能。 2. 大语言模型(LLM) 在容量/能效上更高效。 在 LLM 方面,我们看到像 DeepSeek 这类模型在处理不需要高精度的任务(如回答烹饪食谱或知识库查询)时极其高效。 然而……随着计算力的增加,准确率(尤其是复杂研究问题)也在提升。Elon 和 Magnificent Seven 意识到了这一点,所以他们正在扫货市场上的所有 GPU 以创造超级智能。这也是为什么 Anthropic 和 Google 正在建设耗资 400 多亿美元的数据中心,用于运行需要更多算力进行批判性思维(如 Genesis 任务)的更高级 Opus 和 Gemini 模型。 在 GPU 方面,每一代新 GPU(例如 H100 -> B200)在能效和每美元性能上都有显著提升。例如,Blackwell B200 是 Hopper H100 的 30 倍。 如果基于这个假设,那么到 2027/2028 年,市场上将出现大量过时的低效 H100 和 B200,导致二手 GPU 市场崩盘。 但是:这是假设我们没有看到对新 AI 能力的指数级需求(我们很可能会看到,且正在发生)。正因为这种指数级需求,今天旧模型(如 7 年前的 TPU 和 2020 年的 GPU)仍被用于低优先级的推理任务。 $NVDA 的订单已积压数年,人们正在购买 $AMD 的 GPU 和 $GOOGL 的 TPU 来构建任何新增产能。 至于思科类比,思科的客户是互联网泡沫时期无盈利能力的公司。$NVDA 的客户是 $META、$AMZN、$GOOGL、$MSFT,这些是世界上最盈利的公司。所以最坏的情况我们可能看到回调,而不是互联网泡沫式的崩盘。

英文原文

Answer is nuanced. So two factors: 1. GPUs get more power efficient. 2. LLMs get more capacity/power efficient. For the LLMs case, we're seeing that on deepseek type models be extremely efficient on stuff that don't require much accuracy. Basic stuff like responding to questions about cooking recipes, or knowledge-base stuff. However... accuracy increases, especially with complex research questions, scaled with compute. And people like Elon + mag7 realize this, which is why they're just buying up all the GPUs on the market to create superintelligence. And why antrhopic/google is building $40B+ datacenters for more advanced opus and gemini models that require more compute for critical thinking (eg. Genesis Mission) For the GPUs case, every new generation of GPU (e.g., H100 -> B200) offers dramatic improvements in power efficiency and performance per dollar. eg. Blackwell B200 is 30x than the Hopper H100. If we go off that assumption, then there would be a massive useless supply of less-efficient H100s and B200s in 2027/2028 creating a used GPU market crash. HOWEVER: This is if we don't see an exponential demand for new AI capabilities (which we likely will, and what we're seeing now). Because of this exponential demand, TODAY, older models are still used (eg. TPUs from 7 years ago and GPUs from 2020), for lower inference task in lower priority inference tasks. $NVDA is backlogged for years and people are buying GPUs from $AMD /TPUs from $GOOGL to build out any new capacity. As for Cisco analogy, Cisco's customers were .com bubble companies with no profitability. $NVDA's customers are $META, $AMZN, $GOOGL, $MSFT the most profitable companies in the world. So worst case scenario we might see a correction, not a .com bubble crash.

在 X 上查看原推 ↗