2025-11-25 · 个股论点

分析$NBIS自建数据中心、GPU贬值、融资稀释及过度建设四大风险。

涉及标的： $AMD $APLD $AWS $CRWV $GOOGL $IREN $NBIS $NVDA $ORCL $QLCM

中文翻译

关于风险的好问题，我也喜欢讨论下行风险。 1. 超大规模云厂商(Hyperscalers)在5年内完全自建数据中心，包括GPU（$GOOGL的TPU，$AWS的Tranium）、能源和选址。这类似于高通(Qualcomm)与苹果的关系。苹果使用$QLCM，然后建立垂直整合——一旦完成就抛弃客户。因此在此期间建立公司自身护城河很重要（$NBIS在此期间通过Cursor、Shopify等企业客户进行全栈建设并扩展子公司。完全依赖一两个超大规模云厂商合同的公司表现不佳）。 2. $NBIS、$IREN、$CRWV和$ORCL的GPU贬值风险。任何进行全栈服务的提供商都有更高的利润率、更高的收入和更快的爬坡速度。但这涉及在GPU上花费数十亿美元，而这些GPU在未来4年内价值大幅缩水。我一直认为GPU多年后仍有价值，例如TPU仍在使用7-8年，2020年的$NVDA GPU仍高价转售并用于不同层级的推理（低优先级），并与新模型并行运行。 3. 信贷收紧+稀释建设支出过多->需要更多稀释以建设和购买GPU。我们已经看到$4B可转债+$2.5B ATM发行。$NBIS应该足够，但始终存在利润率低于预期的风险（如$ORCL建设），然后公司需要更多稀释。在更难融资的市场中（如$APLD垃圾债），可能会遇到$CRWV的问题，被迫以糟糕的利率融资，侵蚀利润率。 4. 过度建设如果LLM软件有突破，例如Claude Opus 4.5使用DeepSeek类型的轻量级推理，那么GPU利用率降低，我们会看到数据中心、$NVDA、$AMD等随AI浪潮下跌。这些可能是$NBIS的四个主要风险。每家公司不同，对于$IREN我会指出其IaaS层的软件编排以优化利润率，但$NBIS已经掌握这一点。

英文原文

Good question about risk, I enjoy talking about the downsides too. 1. Hyperscalers completely their own datacenter buildout with GPUs (TPUs with $GOOGL, Tranium with $AWS), energy, locations in 5+ years. This is the analogy of Qualcomm with Apple. Apple just uses $QLCM, then builds it's own vertical integration -> once it's done it sacks the customer. Hence why it's important to build up a company's own moat during the time ( $NBIS is doing full-stack with its own enterprise clients like cursor, shopify, etc during this time and scaling subsidiaries. Companies fully dependent on one or two hyperscaler contracts don't do well). 2. GPU deprecation for $NBIS, $IREN, $CRWV, and $ORCL. Any provider doing full stack has higher margins, higher revenue, higher ramp. But that comes into spending billions on GPUs and the GPUs not too much value over the course of the next 4 years. I've always maintained GPUs are still valuable many years later, eg. TPUs still are used 7-8 years later, 2020 $NVDA GPUs are still resold high and are used different stacks of inference (lower prio) but run alongside newer models. 3. Credit tightening + dilution Too much spend on buildout -> needs to dilute more to build + buy GPUs. We've already seen that with the $4B convertible + $2.5B ATM offering. $NBIS should have enough, but there's always the risk margins are lower than expected eg. $ORCL buildout, then the company needs to dilute more. In a market where it's harder to raise eg. $APLD junk bonds, it might run into the $CRWV problem where they are forced to raise money with terrible interest rates cutting into the margins. 4. Overbuildout If there's a breakthrough in LLM software, for example, Claude Opus 4.5 using deepseek type lightweight inference, then less utilization of GPUs then we see data centers, $NVDA, $AMD, and others go down with the AI ship. Those are probably the four main risks for $NBIS. This is different for each company, for $IREN I would point to software orchestration for their iaas layer for margin optimization but for $NBIS they have that down.

在 X 上查看原推 ↗