Burry误读NVIDIA GPU图表实为租金预测,作者仍不看空
Burry says NVIDIA's GPU "resale value" chart is really a rent forecast. I checked the source: he's right. I still don't think it's a reason to short NVIDIA.
提供对Burry观点的深度拆解与量化验证,包含具体的租金盈亏平衡测算逻辑,适合关注AI算力产业链估值与风险的投资者研究参考。
TL;DR:
TL;DR(太长不看):
- The values on NVIDIA's slide are Silicon Data's discounted rent, not resale prices, and they move with rent.
- But the memory shortage holding rent up is backed by the buyer, not just the sellers, and runs past his September 2027 puts.
- The lessors mostly pay off their loans inside their contracts, and where they don't, the credit market is charging for it.
- NVIDIA幻灯片上的数值是Silicon Data的折现租金,而非转售价格,且它们随租金波动。
- 但支撑租金高企的内存短缺背后有买家支持,而不仅仅是卖家,这超出了他2027年9月到期看跌期权的范围。
- 出租方大多在合同期内还清贷款;若未还清,信贷市场会为此收取溢价。
I own NVIDIA, and The Big Short was the first movie I watched when I started investing, so I read Burry's October 1 post properly instead of scrolling past.
我持有NVIDIA股票,《大空头》是我开始投资时看的第一部电影,所以我认真读了Burry 10月1日的帖子,而不是直接滑过。
His target is one slide from NVIDIA's September investor deck. It shows A100, H100 and B200 values far above a five-year depreciation curve. The footnote cites Silicon Data, and Silicon Data's own methodology says those values are a discounted cash flow of projected rent over an eight-year life. Nobody actually sold a used B200 for $72,695.
他的目标源自NVIDIA 9月投资者演示文稿中的一张幻灯片。该幻灯片显示A100、H100和B200的价值远高于五年折旧曲线。脚注引用了Silicon Data,而Silicon Data自身的方法论指出,这些价值是基于八年使用寿命的预期租金现金流折现值。实际上并没有人将二手B200以72,695美元的价格卖出。
The clearest evidence: Silicon Data's H100 value was $14,976 in November 2025 and $22,068 in June 2026. Same chip; its rent went from $1.99 to $2.74 an hour. And the last time a GPU shortage eased, H100 rent fell 44% in five months (Sept 2024 to Feb 2025). On all of that, Burry is right.
最清晰的证据:Silicon Data的H100估值在2025年11月为14,976美元,在2026年6月为22,068美元。同一款芯片;其租金从每小时1.99美元涨至2.74美元。而上一次GPU短缺缓解时,H100租金在五个月内下跌了44%(2024年9月至2025年2月)。基于以上所有因素,Burry是正确的。
Where I disagree:
我的分歧点在于:
- His clock. His puts expire September 2027. The evidence that the memory shortage lasts comes from the buyer, not just the sellers. NVIDIA's CFO lowered her own Q4 margin guidance because of memory prices and said supply stays a bottleneck through January 2028. NVIDIA's purchase commitments went from $119B to $279B in one quarter, mostly memory. Samsung says a new fab takes more than three years to produce.
- The cushion. B200 rent is $5.86/hr. Assuming 40% of rent goes to operating costs (close to CoreWeave's 59% adjusted EBITDA margin), this is how far it can fall before a new B200 stops paying for itself in six years:
- 70% utilization, ignoring the cost of money: 65% (to $2.04)
- 50% utilization, ignoring the cost of money: 51% ($2.85)
- 70% utilization, money at 10% a year (about what CoreWeave's riskiest loan costs): 52% ($2.81)
- 50% utilization, money at 10% a year: 33% ($3.93). That's the case that would make me wrong.
- His target. His 1968 analogy is about leasing companies that borrowed against machines and had to keep re-renting them. Today's version is the neoclouds. Nebius's customers prepay 50% to 60% of capex. CoreWeave's main loans are repaid inside each customer contract. Its one exception, DDTL 5.5, runs about five years against customer contracts averaging three, so the lenders carry renewal risk. The CFO called it the first of its loans "to include shorter duration customer contracts" (Q2 call, Aug 11). It's rated Ba2/BB+ and priced at SOFR + 5.50%, and CoreWeave's 5-year CDS is about 8.2%. In 2006 this kind of risk was rated investment grade and cheap to insure. This time it's priced.
- 他的时间线。他的看跌期权于2027年9月到期。证明内存短缺将持续的证据来自买家,而不仅仅是卖家。NVIDIA的首席财务官因内存价格下调了她自己第四季度的利润率指引,并表示供应瓶颈将持续到2028年1月。NVIDIA的采购承诺在一个季度内从1190亿美元增至2790亿美元,大部分用于内存。三星表示,新建晶圆厂需要超过三年才能投产。
- 缓冲空间。B200的租金为5.86美元/小时。假设40%的租金用于运营成本(接近CoreWeave 59%的调整后EBITDA利润率),这是在新B200六年内无法自偿之前,租金能下跌的空间:
- 70%利用率,忽略资金成本:可下跌65%(至2.04美元)
- 50%利用率,忽略资金成本:可下跌51%(至2.85美元)
- 70%利用率,年利率10%的资金成本(约等于CoreWeave最危险贷款的利率):可下跌52%(至2.81美元)
- 50%利用率,年利率10%的资金成本:可下跌33%(至3.93美元)。这才是可能让我出错的情况。
- 他的目标。他在1968年的类比涉及那些以机器为抵押借款并不得不反复续租的租赁公司。今天的版本是“新云”(neoclouds)。Nebius的客户预付50%至60%的资本支出(capex)。CoreWeave的主要贷款在每个客户合同内偿还。其例外情况DDTL 5.5的期限约为五年,而客户合同的平均期限为三年,因此贷款人承担续约风险。首席财务官称这是其贷款中“首次包含较短期限客户合同”的一笔(8月11日的第二季度电话会议)。该贷款评级为Ba2/BB+,定价为SOFR + 5.50%,而CoreWeave的5年期信用违约互换(CDS)约为8.2%。在2006年,这类风险的评级为投资级且保险成本低廉。这次则已计入定价。
NVIDIA isn't risk-free. Its worst-case exposure to the companies it helps finance is $164.5B (guarantees, AI cloud capacity backstops and leases), about 37% of FY28 consensus EBITDA, though none of the $105B Ohio guarantee is live before his puts expire.
NVIDIA并非没有风险。它对其协助融资的公司最坏情况下的敞口为1645亿美元(包括担保、AI云容量后备支持和租赁),约占2028财年一致预期EBITDA的37%,尽管在其看跌期权到期前,1050亿美元的俄亥俄州担保尚未生效。
The bigger risk to my position isn't rent. NVIDIA trades at about 16.5x forward earnings, and if the market prices the end of the shortage early, the way it already prices Micron at 6.1x, the stock could fall 36% with rent intact. Price isn't one of my break conditions, so if that happened with the business intact, I'd hold and add. Whether the multiple holds depends on what demand looks like after the shortage, and that's my next post.
对我持仓而言更大的风险并非租金。NVIDIA的交易价格约为远期收益的16.5倍,如果市场提前对短缺结束进行定价,就像它目前将美光科技(Micron)定价为6.1倍那样,股价可能会下跌36%,而租金保持不变。价格并非我的止损条件之一,因此如果发生这种情况且业务保持完整,我会持有并加仓。估值倍数能否维持取决于短缺结束后需求的情况,这将是我的下一篇帖子。
What would prove me wrong: Silicon Data's B200 rent index below $2.85 for three straight months before September 2027 (roughly where a new B200 stops paying for itself in six years at 50% utilization, or at 70% with 10% debt), or Micron, Samsung or SK hynix saying memory supply and demand balance in 2027. I start reviewing below $3.93, where half-empty clusters funded at 10% stop paying.
证明我错误的情况:Silicon Data的B200租金指数在2027年9月之前连续三个月低于2.85美元(大致相当于一台新的B200在50%利用率下六年停止自我盈利的位置,或在70%利用率及10%债务情况下),或者美光、三星或SK海力士表示内存供需在2027年达到平衡。我在低于3.93美元时开始审查,此时资金按10%杠杆借入的空置集群停止自我盈利。
Full post with charts and sources: https://darrenleung1.substack.com/p/burry-is-right-about-the-gpus-hes
包含图表和来源的完整帖子:https://darrenleung1.substack.com/p/burry-is-right-about-the-gpus-hes
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力