2025-11-11 · 供应链分析

新云财报解读及反驳伯里GPU贬值论，强调新云逻辑强劲。

涉及标的： $AMZN $CIFR $CRWV $IREN $META $ORCL $WULF

中文翻译

今日新云(Neocloud)财报 + 为何迈克尔·伯里(Michael Burry)错了： $WULF - 财报超预期(ER)表现积极，来自520兆瓦(HPC)算力的高性能计算(HPC)容量总合同价值(TCV)达170亿美元。在德州成立240兆瓦合资企业(JV)用于数据中心(DC)扩张 + 50亿美元融资。不像$CIFR与AWS的财报超预期那样特别，只是确认（如交易融资）和站点开发。 $CRWV - 因第四季度远期收入较低，财报超预期表现短期看空。 556亿美元积压订单（其中224亿美元来自OpenAI）。小幅盈利超预期13.6亿 vs 12.9亿（同比129%），EBITDA利润率（61%）。股价因数据中心延迟下跌，这损害了2025财年（50.5-51.5亿 vs 共识52.9亿）。运营利润率下调，正如我预测的那样，因为债务融资产生的利息（~10亿+利息/年）侵蚀了利润率，而其他新云没有同样的问题。话虽如此，这是一次性问题，鉴于市场是前瞻性的，我们可能会看到任何抛售在下个季度后被买回。 _ 至于迈克尔·伯里关于GPU贬值的争论：他完全错了，但在低估这一点上部分正确。他的普遍主张：芯片在2-3年产品周期上的资本支出(Capex)不应导致使用寿命的延长 -> 在谷歌的案例中，他们7年前的TPU正以100%的利用率运行。重复一遍，是7年前。 Amin Vahdat表示，谷歌目前生产中有七代TPU硬件，且“七到八年前的TPU拥有100%的利用率”。 “谷歌称TPU需求超过供应，声称8年前的硬件迭代‘100%利用率’” - 数据中心动态 -> 在英伟达(Nvidia)的案例中，A100（PCIe和SXM4变体）仍在运行。它们于2020年推出。现在是2025年。旧型号保持其价值并仍然提供收益。它们不会仅仅因为是补充新模型的附加组件就被核销。英伟达的A100可以划分为隔离实例以最大化利用率，旧一代加速器仍用于推理或低优先级任务进行货币化，而不是像他所声称的那样被核销。它们实际上具有残值，5年后仍能高价转售。他在保质期问题上完全错了，但在某些超大规模云服务商如何拉伸利用率概念以证明激进资本支出方面可能是正确的。这是一个微妙的论点，但新云逻辑依然完好且比以往任何时候都更强。

英文原文

Neocloud Earnings Today + Why Michael Burry is Wrong: $WULF - ER was positive, $17B TCV from 520 MW HPC capacity. 240 MW JV in Texas for DC expansion + $5B financing. Nothing special like $CIFR ER with AWS, just conf (eg. deal finances) and site dev. $CRWV - ER was bearish near term due to lower forward revenue Q4. $55.6B backlog ($22.4B from OpenAI). Slight earnings beat 1.36B vs. 1.29B (129% Y/Y), EBITDA margins (61%). Stock dropped on datacenter delay, this hurts FY 2025 (505-5.15B vs. 5.29B consensus). Operating margins was lowered, which I predicted due to debt financing with interest (~1B+ interest/year) that cuts into margins, while other Neoclouds don't have that same issue. That being said this was a one-off issue, and we'll likely see any sell-off bought back after next quarter since markets are forward looking. _ As for Michael Burry arguing on GPU deprecation: he's completely wrong, but partially right about understating it. His general claim: capex of chips on a 2-3 yr product cycle should not result in the extension of useful lives -> in Google's case, their TPUs from 7 years ago are being run at 100% utilization lol. Repeat 7 years ago. Amin Vahdat, said that Google currently has seven generations of its TPU hardware in production, and that the “seven- and eight-year-old TPUs have 100 % utilization.” "Google says TPU demand is outstripping supply, claims 8yr old hardware iterations have “100% utilization” -Data Center Dynamics -> in Nvidia's case A100 (PCIe and SXM4 variants), are still in operation. They were launched in 2020. It's 2025 now. Older models keep their value and still deliver equity. They don't just get nuked since it's an add-on to compliment newer models. Nvidia’s A100 can be partitioned into isolated instances to maximize utilization, and older-gen accelerators are still monetized for inference or lower-prio tasks rather than getting nuked as what he claims. They actually have residual value and still re-sale for a lot 5 years later. He's completely wrong here on shelf-life, but might be correct in how some hyperscalers stretch the idea of utilization to justify aggressive capex. It's a nuanced argument but Neocloud thesis remains in tact and stronger than ever.

在 X 上查看原推 ↗