· 个股论点

英伟达虽面临谷歌TPU等定制芯片竞争,但中期统治力稳固,逢低买入。

涉及标的:

中文翻译

英伟达($NVDA)公布的2026财年Q3营收为570.1亿美元(+62.5% YoY),表现强劲。并指引Q4营收超650亿美元(超预期30亿美元+),以及通过CY 2026年Blackwell/Rubin系列营收超5000亿美元。尽管如此,股价仍下跌12%。现在$NVDA是强力买入吗?答案如下: $GOOGL的TPU项目成为首个对$NVDA GPU构成竞争替代的方案,Anthropic承诺采购超100万颗TPU芯片,Meta据报道正在就数十亿美元的TPU采购进行高级别谈判。沃伦·巴菲特近期也向$GOOGL投资超40亿美元,鉴于伯克希尔对科技股保守的投资立场,这极为罕见。 尽管创下盈利新高,英伟达股价在过去10个交易日中有6天下跌,较10月29日触及5.03万亿美元市值时的历史高点$207.04下跌约12-15%。 分析师反应普遍看多,普遍上调目标价: - Evercore ISI从$261上调至$352 - 美银从$235上调至$275 - 花旗从$220上调至$270 - 高盛从$240上调至$250 - 摩根士丹利从$220上调至$235 但这里有个价值万亿美元的问题:超大规模客户日益增长的定制硅片威胁是否会削弱英伟达在AI领域的统治地位? 与主要捕捉英伟达GPU缺货时的溢出需求的$AMD不同,谷歌的TPU项目代表了根本不同的竞争动态。 TPU v7 Ironwood是首款在性能上与Blackwell持平的非英伟达加速器,提供4.6 petaflops的FP8性能(对比B200的4.5 petaflops),配备192GB HBM3e内存。 Ironwood的架构差异化显著。虽然英伟达最大的集群配置为72个GPU(NVL72),TPU Ironwood可扩展至9,216个芯片组。 客户斩获显著且不断增长: - Anthropic承诺采购超100万颗TPU芯片,价值“数百亿美元”,1GW算力即将上线。 - Meta正在就2026年从谷歌云租赁TPU容量进行高级别谈判,并计划2027年直接采购硬件用于自有数据中心。 - 苹果透露Apple Intelligence基础模型完全在TPU上训练,使用8,192颗TPUv4芯片用于服务器模型。 - Midjourney从GPU转向TPU,推理成本降低65%(从每月200万美元降至70万美元)。 定位微妙。TPU在超大规模推理方面表现出色,在生产级大规模服务中成本性能最高提升4倍(目前)。在训练方面,英伟达优势明显。 对于高度优化的推理任务,TPU架构可能比$NVDA的通用GPU更高效。 然而,我预计下一代英伟达GPU将在许多场景下在推理性能上超越TPU。(类似于LLM之间GPT 5 -> Gemini 3 -> Opus 4.5的迭代超越) 我们看到: 谷歌TPU、AWS Trainium、Meta MTIA、微软Maia和定制超大规模芯片都在扩展,但仍依赖$NVDA。但超大规模客户集体减少对英伟达的依赖,其累积效应是否会削弱英伟达的统治地位? 答案:不会。至少未来两年统治地位稳固。之后如何纯属猜测。 仅看数据:Q3结果证实公司仍是AI不可或缺的基础设施提供商,5000亿美元的订单积压为2026年提供了极高的可见性。 但市场似乎正在定价3年+后的这种微妙现实:长期面临超大规模客户定制芯片的竞争不确定性。 英伟达产能完全售罄,且很可能在下一代芯片中超越TPU性能(推理性能更高且保持通用性)。但定制硅片威胁和44倍市盈率的估值担忧仍是逆风。 无论如何,鉴于出色的超预期财报和未来两年的订单积压,$NVDA目前因恐惧而下跌,似乎是中期强力逢低买入的机会。 你只需要记住这一点: 只要英伟达仍是AI工作负载的行业首选,且TPU和AMD GPU仅在需求超过英伟达供应时填补空白,它就是强力买入标的。 英伟达订单积压已满,AI需求并未放缓。

英文原文

Nvidia ( $NVDA ) reported a blowout Q3 FY2026 revenue of $57.01 billion (+62.5% YoY). And guided $65B+ Q4 ($3B+ beat), and $500B+ USD in Blackwell/Rubin rev through CY 2026. Despite that, the stock dropped 12%. Is $NVDA a strong buy now? Here's the answer: $GOOGL TPU's program emerged as the first competitive alternative to $NVDA GPUs, with Anthropic committing to over 1 million TPU chips and Meta reportedly in advanced negotiations for billions in TPU purchases. Warren Buffet also recently invested $4B+ into $GOOGL, which is extraordinarily rare given Berkshire's conservative stance to tech investments. Despite a record earnings beat, Nvidia's stock has declined in six of the last ten trading sessions and sits roughly 12-15% below its October 29 all-time high of $207.04, when it briefly touched a $5.03 trillion market cap. Analyst reaction was overwhelmingly bullish, with price targets raised across the board: - Evercore ISI raised to $352 from $261 - Bank of America raised to $275 from $235 - Citigroup raised to $270 from $220 - Goldman Sachs raised to $250 from $240 - Morgan Stanley raised to $235 from $220 But here's the trillion dollar question: will the emerging custom silicon threat from hyperscalers reduce NVIDIA dominance in AI? Unlike $AMD, which primarily captures overflow demand when NVIDIA GPUs are unavailable, Google's TPU program represents a fundamentally different competitive dynamic. TPU v7 Ironwood is the first non-NVIDIA accelerator that achieves performance parity with Blackwell, delivering 4.6 petaflops of FP8 performance (versus B200's 4.5 petaflops) with 192GB HBM3e memory. Ironwood's architectural differentiation is substantial. While NVIDIA's largest cluster configuration is 72 GPUs (NVL72), TPU Ironwood scleaes to 9,216 chip pods. The customer wins are significant and growing: - Anthropic committed to over 1 million TPU chips worth "tens of billions of dollars," with 1 gw of compute capacity coming online. - Meta is in advanced negotiations to rent TPU capacity from Google Cloud in 2026, with direct hardware purchases for its own data centers planned for 2027. - Apple revealed that Apple Intelligence foundation models were trained entirely on TPUs using 8,192 TPUv4 chips for server models - Midjourney switched from GPUs to TPUs and reduced inference costs by 65% (from $2M to $700K monthly) The positioning is nuanced. TPUs excel at hyperscale inference with up to 4x better cost-performance for production serving at scale (for now). For training, NVIDIA is the clear advantage. For highly optimized inference tasks, TPU architecture might remain more efficient than $NVDA's general-purpose GPU. However, I'm expecting next-gen Nvidia GPUs to leapfrog TPUs for inference in many scenarios. (similar to how LLMs leapfrog each other GPT 5 -> Gemini 3 -> Opus 4.5) We're seeing: Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia, and custom hyperscaler chips scale up to reliance on $NVDA. But will the cumulative effect on hyperscalers collectively reducing Nvidia's dominance? The answer: No. Not yet. Dominance is secured at least for the next two years. What happens after is only speculation. Just looking at the numbers: Q3 results confirm the company remains the essential infrastructure provider for AI, with a $500 billion order backlog providing exceptional visibility through 2026. But the market appears to be pricing in this nuanced reality 3 years+ from now: long-term competitive uncertainty with custom hyperscaler chips. Nvidia is completely sold out of capacity, and are likely to leapfrog TPU performance in their next generation chips (higher performance for inference while being general purpose). But the custom silicon threat and valuation concerns at 44x earnings remains a headwind. Regardless, $NVDA seems to be a strong mid term dip-buy now on fears, given the exceptional blowout earnings and backlog for the next 2 years. This is the only thing you need to remember: NVIDIA is a strong buy as long as it remains the industry’s first choice for AI workloads, with TPUs and AMD GPUs filling gaps when demand exceeds NVIDIA’s supply. Nvidia is maxed out on backlog, and AI demand is not slowing down.

在 X 上查看原推 ↗