<p>大卫·泰珀是华尔街最成功的投资经理之一。他的阿帕卢萨管理对冲基金自 1993 年成立以来，年化总回报超过 28%。这一回报远远超过了过去 32 年中标准普尔 500 指数约 10.6% 的年化回报。</p>
<p>泰珀以购买接近破产公司的不良债务而闻名。事实上，阿帕卢萨在 1990 年代被视为垃圾债券投资精品公司。这种逆向投资策略通常也延伸到他的股票投资组合中。</p>
<p>尽管如此，他并不是那么固执于逆流而上，以至于不愿意购买像人工智能（AI）这样的明显趋势中的股票。像 <strong>英伟达</strong>（NVIDIA）和 <strong>超微半导体</strong>（AMD）这样的 AI 股票在过去几年中价值飙升。泰珀在这些股票上赚了不少钱。</p>
<p>但他最近开始出售这些股票，转而投资另一家 AI 芯片制造商，可能是在对抗这两家大型 GPU 制造商。</p>
<p><img src="https://imageproxy.pbkrs.com/https://g.foolcdn.com/image//query-b3A9cmVzaXplJnVybD1odHRwczovL2cuZm9vbGNkbi5jb20vZWRpdG9yaWFsL2ltYWdlcy84MjU0ODkvZ2V0dHlpbWFnZXMtYWktY2hpcC1vbi1hLWNpcmN1aXQtYm9hcmQuanBnJnc9NzAw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://g.foolcdn.com/image//query-b3A9cmVzaXplJnVybD1odHRwczovL2cuZm9vbGNkbi5jb20vZWRpdG9yaWFsL2ltYWdlcy84MjU0ODkvZ2V0dHlpbWFnZXMtYWktY2hpcC1vbi1hLWNpcmN1aXQtYm9hcmQuanBnJnc9NzAw"/></p>
<p>图片来源：Getty Images。</p>
<h2>AI 革命背后的基础设施</h2>
<p>图形处理单元（GPU）是计算机芯片或系统，已被证明在处理用于训练和运行大型语言模型的数据方面极为高效。GPU 被设计用于进行训练 AI 算法所需的计算，并且可以并行运行这些过程，使其比使用串行处理的标准 CPU 更高效。</p>
<p>英伟达长期以来一直是 GPU 的领导者，追溯到它们主要用于高端游戏视觉的时代（因此 G 代表 “图形”）。随着大型语言模型的处理需求呈指数级增长，英伟达的领先 GPU 系统需求激增。</p>
<p>即使在 2023 年和 2024 年强劲增长之后，英伟达的数据中心收入在上个季度同比增长了 73%。凭借强大的经营杠杆，公司盈利大幅上升，投资者也给予了回报。现在，它是全球最有价值的公司，市值超过 4 万亿美元。</p>
<p>但 AMD 开始在追赶英伟达方面取得进展。该公司的 MI400 芯片预计将在明年推出，可能提供比英伟达当前的 Blackwell 系列芯片更好的性价比。虽然到那时英伟达将进入其下一代 Vera Rubin 平台，但 AMD 正在为英伟达最大的客户提供一个可行的替代方案，这可能会阻止其价格大幅上涨。</p>
<p>AMD 的股票表现远不及英伟达。在 2024 年初达到峰值后，该股在今年 4 月暴跌超过 60%。第一季度本可以是购买该股票的绝佳机会，尤其是对于希望对抗英伟达持续主导地位的逆向投资者来说。</p>
<p>但泰珀在第一季度出售了他在 AMD 的全部股份，这一头寸最初是在 2023 年第二季度建立的。他还继续削减他在英伟达的持股，留下的仅为他在 2023 年中期为阿帕卢萨持有的 3% 的股份。相反，他正在押注另一家对 AI 数据中心 GPU 主导地位构成日益威胁的芯片制造商。</p>
<h2>泰珀正在购买的 AI 芯片制造商</h2>
<p>虽然 GPU 极为灵活，能够处理各种任务，但许多开发尖端 AI 能力的最大公司正在研发定制硅，以比耗电的 GPU 更高效地处理特定任务。这些应用特定集成电路（ASIC）对 GPU 构成了重大威胁，因为像 <strong>Meta Platforms</strong> 和 <strong>Alphabet</strong> 的谷歌这样的超大规模公司正在设计更先进的芯片，能够处理 AI 训练和推理。</p>
<p>ASIC 的能力正在扩展。Meta 表示，其定制芯片历史上处理机器学习 AI，现在正在扩展到训练大型语言模型，最初是从机器学习算法开始，然后转向 AI 推理。谷歌在其自有芯片设计上训练了其大型语言模型 Gemini，并在 4 月发布了其首个用于 AI 推理的张量处理单元（TPU）。</p>
<p>帮助 Meta 和谷歌设计其 ASIC 的公司是 <strong>博通</strong>（Broadcom）。除了 ASIC 业务外，博通还是领先的网络芯片制造商。网络是 AI 数据中心的重要组成部分，因为良好的网络性能确保所有数据能够快速高效地传输到昂贵的 GPU 或 ASIC 上。这些企业在这些芯片上花费数十亿美元，因此他们不希望这些芯片闲置太久。</p>
<p>博通还有一个企业软件业务，由虚拟机软件 VMWare 主导。</p>
<p>也就是说，博通相比于英伟达甚至 AMD（后者也有强大的 CPU 业务）提供了更为多元化的芯片制造商。这可能是泰珀在第一季度对该公司进行小额投资的原因，因为它是 AI 芯片的领先竞争者，同时通过其 VMWare 业务提供了一些下行保护。</p>
<p>不过，博通的股票价格昂贵。其前瞻性市盈率接近 40。这与英伟达持平，略低于 AMD。如果 ASIC 设计随着时间的推移在数据中心占据更多市场份额，该公司的潜在上行空间可能更大。考虑到 ASIC 与 GPU 相比的潜在效率提升，这在长期内似乎是可能的。</p>
<p>但投资者可能不得不接受比英伟达或 AMD 更缓慢而稳定的增长。因此，值得关注博通的股票，看看它是否在价格回落到更具吸引力的水平之前，跟随泰珀进入该股票。</p>

英伟达

博通

<p>亿万富翁投资者大卫·泰珀已出售他在英伟达和 AMD 的股份，而选择投资于博通，这是一家专注于人工智能芯片的公司。泰珀的对冲基金 Appaloosa Management 历来专注于困境资产，但现在正转向人工智能趋势。尽管英伟达仍然是 GPU 领域的领导者，但博通凭借其定制 ASIC 芯片正在获得关注，这些芯片在人工智能任务中越来越受到青睐。这一转变反映了泰珀的逆向投资策略，因为他在不断发展的人工智能芯片市场中寻求机会</p>

<p>David Tepper is one of the most successful investment managers on Wall Street. His Appaloosa Management hedge fund has produced gross annualized returns of more than 28% since its inception in 1993. That far outpaces the <strong>S&amp;P 500</strong>'s annualized return over the last 32-plus years of about 10.6%.</p>
<div class="lb-trans"><p>大卫·泰珀是华尔街最成功的投资经理之一。他的阿帕卢萨管理对冲基金自 1993 年成立以来，年化总回报超过 28%。这一回报远远超过了过去 32 年中标准普尔 500 指数约 10.6% 的年化回报。</p>
</div><p>Tepper is best known for buying distressed debt from companies close to bankruptcy. In fact, Appaloosa was considered a junk bond investment boutique in the 1990s. That contrarian approach often extends to his stock portfolio as well.</p>
<div class="lb-trans"><p>泰珀以购买接近破产公司的不良债务而闻名。事实上，阿帕卢萨在 1990 年代被视为垃圾债券投资精品公司。这种逆向投资策略通常也延伸到他的股票投资组合中。</p>
</div><p>That said, he's not so set on swimming against the current that he won't buy stocks that are part of an obvious trend like artificial intelligence (AI). AI stocks like <strong>Nvidia</strong>(NVDA -0.42%) and <strong>Advanced Micro Devices</strong>(AMD -2.16%) have soared in value over the last few years. And Tepper made quite a bit of money on those stocks.</p>
<div class="lb-trans"><p>尽管如此，他并不是那么固执于逆流而上，以至于不愿意购买像人工智能（AI）这样的明显趋势中的股票。像 <strong>英伟达</strong>（NVIDIA）和 <strong>超微半导体</strong>（AMD）这样的 AI 股票在过去几年中价值飙升。泰珀在这些股票上赚了不少钱。</p>
</div><p>But he's been selling them recently in favor of another AI chipmaker instead, possibly taking a bit of a contrarian stance against the two big GPU makers.</p>
<div class="lb-trans"><p>但他最近开始出售这些股票，转而投资另一家 AI 芯片制造商，可能是在对抗这两家大型 GPU 制造商。</p>
</div><p><img src="https://imageproxy.pbkrs.com/https://g.foolcdn.com/image//query-b3A9cmVzaXplJnVybD1odHRwczovL2cuZm9vbGNkbi5jb20vZWRpdG9yaWFsL2ltYWdlcy84MjU0ODkvZ2V0dHlpbWFnZXMtYWktY2hpcC1vbi1hLWNpcmN1aXQtYm9hcmQuanBnJnc9NzAw" alt="" original-src="https://g.foolcdn.com/image/?url=https%3A%2F%2Fg.foolcdn.com%2Feditorial%2Fimages%2F825489%2Fgettyimages-ai-chip-on-a-circuit-board.jpg&amp;op=resize&amp;w=700"/></p>
<p>Image source: Getty Images.</p>
<div class="lb-trans"><p>图片来源：Getty Images。</p>
</div><h2>The essential infrastructure behind the AI revolution</h2>
<div class="lb-trans"><h2>AI 革命背后的基础设施</h2>
</div><p>Graphics processing units (GPUs) are computer chips or systems that have proven exceptionally adept at crunching all the data that goes into training and running large language models. GPUs are designed to do the types of calculations needed for training AI algorithms, and they can run the processes in parallel, making them far more efficient than a standard CPU, which uses serial processing.</p>
<div class="lb-trans"><p>图形处理单元（GPU）是计算机芯片或系统，已被证明在处理用于训练和运行大型语言模型的数据方面极为高效。GPU 被设计用于进行训练 AI 算法所需的计算，并且可以并行运行这些过程，使其比使用串行处理的标准 CPU 更高效。</p>
</div><p>Nvidia has long been a leader in GPUs, dating back to the days when they were mostly just used for high-end visuals in gaming (hence why the G stands for &#34;graphics&#34;). As the processing needs of large language models grew exponentially larger, it's seen incredible demand for its leading GPU systems.</p>
<div class="lb-trans"><p>英伟达长期以来一直是 GPU 的领导者，追溯到它们主要用于高端游戏视觉的时代（因此 G 代表 “图形”）。随着大型语言模型的处理需求呈指数级增长，英伟达的领先 GPU 系统需求激增。</p>
</div><p>Even after the strong growth in 2023 and 2024, Nvidia's data center revenue climbed another 73% year over year last quarter. With strong operating leverage, the company has seen its earnings zoom higher, and investors have rewarded it. It's now the most valuable company in the world by a substantial margin, worth over $4 trillion.</p>
<div class="lb-trans"><p>即使在 2023 年和 2024 年强劲增长之后，英伟达的数据中心收入在上个季度同比增长了 73%。凭借强大的经营杠杆，公司盈利大幅上升，投资者也给予了回报。现在，它是全球最有价值的公司，市值超过 4 万亿美元。</p>
</div><p>But AMD is starting to make progress in catching up to Nvidia. The company's MI400 chips coming next year could offer better price performance than Nvidia's current Blackwell line of chips. While Nvidia will be on to its next-generation Vera Rubin platform by then, AMD is offering Nvidia's biggest customers a viable alternative, which could keep its pricing from climbing substantially higher.</p>
<div class="lb-trans"><p>但 AMD 开始在追赶英伟达方面取得进展。该公司的 MI400 芯片预计将在明年推出，可能提供比英伟达当前的 Blackwell 系列芯片更好的性价比。虽然到那时英伟达将进入其下一代 Vera Rubin 平台，但 AMD 正在为英伟达最大的客户提供一个可行的替代方案，这可能会阻止其价格大幅上涨。</p>
</div><p>AMD's stock hasn't performed nearly as well as Nvidia's. After peaking in early 2024, the stock crashed more than 60% to its low in April this year. The first quarter could have been a great opportunity to buy the stock, especially for a contrarian investor looking to take a stance against Nvidia's continued dominance.</p>
<div class="lb-trans"><p>AMD 的股票表现远不及英伟达。在 2024 年初达到峰值后，该股在今年 4 月暴跌超过 60%。第一季度本可以是购买该股票的绝佳机会，尤其是对于希望对抗英伟达持续主导地位的逆向投资者来说。</p>
</div><p>But Tepper sold his entire stake in AMD during the first quarter, a position first established in the second quarter of 2023. He also continued to cut his Nvidia stake, leaving him with just 3% of the shares he held for Appaloosa in mid-2023. Instead, he's betting on a different chipmaker that poses an increasing threat to the dominance of GPUs in AI data centers.</p>
<div class="lb-trans"><p>但泰珀在第一季度出售了他在 AMD 的全部股份，这一头寸最初是在 2023 年第二季度建立的。他还继续削减他在英伟达的持股，留下的仅为他在 2023 年中期为阿帕卢萨持有的 3% 的股份。相反，他正在押注另一家对 AI 数据中心 GPU 主导地位构成日益威胁的芯片制造商。</p>
</div><h2>The AI chipmaker Tepper's buying instead</h2>
<div class="lb-trans"><h2>泰珀正在购买的 AI 芯片制造商</h2>
</div><p>While GPUs are extremely flexible and capable of handling all sorts of tasks, many of the biggest companies developing leading-edge AI capabilities are working on custom-made silicon that can handle specific tasks far more efficiently than power-hungry GPUs. These application-specific integrated circuits, or ASICs, represent a significant threat to GPUs, as hyperscalers like <strong>Meta Platforms</strong> and <strong>Alphabet</strong>'s Google design more advanced chips capable of handling AI training and inference.</p>
<div class="lb-trans"><p>虽然 GPU 极为灵活，能够处理各种任务，但许多开发尖端 AI 能力的最大公司正在研发定制硅，以比耗电的 GPU 更高效地处理特定任务。这些应用特定集成电路（ASIC）对 GPU 构成了重大威胁，因为像 <strong>Meta Platforms</strong> 和 <strong>Alphabet</strong> 的谷歌这样的超大规模公司正在设计更先进的芯片，能够处理 AI 训练和推理。</p>
</div><p>The capabilities of ASICs are expanding. Meta says its custom chips, which have historically handled machine learning AI, are expanding to training large language models after starting with machine learning algorithms and moving on to AI inference. Google trained its large language model Gemini on its own chip designs, and it just released its first Tensor Processing Unit (TPU) designed for AI inference in April.</p>
<div class="lb-trans"><p>ASIC 的能力正在扩展。Meta 表示，其定制芯片历史上处理机器学习 AI，现在正在扩展到训练大型语言模型，最初是从机器学习算法开始，然后转向 AI 推理。谷歌在其自有芯片设计上训练了其大型语言模型 Gemini，并在 4 月发布了其首个用于 AI 推理的张量处理单元（TPU）。</p>
</div><p>The company helping Meta and Google design their ASICs is <strong>Broadcom</strong>(AVGO -1.12%). On top of that ASIC business, Broadcom is also the leading networking chipmaker. Networking is an essential piece of AI data centers, as solid network performance ensures all the data gets to the expensive GPUs or ASICs quickly and efficiently. These businesses are spending billions on those chips, so they don't want them sitting idly any longer than necessary.</p>
<div class="lb-trans"><p>帮助 Meta 和谷歌设计其 ASIC 的公司是 <strong>博通</strong>（Broadcom）。除了 ASIC 业务外，博通还是领先的网络芯片制造商。网络是 AI 数据中心的重要组成部分，因为良好的网络性能确保所有数据能够快速高效地传输到昂贵的 GPU 或 ASIC 上。这些企业在这些芯片上花费数十亿美元，因此他们不希望这些芯片闲置太久。</p>
</div><p>Broadcom also has an enterprise software business, led by virtual machine software VMWare.</p>
<div class="lb-trans"><p>博通还有一个企业软件业务，由虚拟机软件 VMWare 主导。</p>
</div><p>That is to say, Broadcom offers a more diversified chipmaker compared to Nvidia or even AMD (which also has a strong CPU business). That may be why Tepper took a small stake in the company during the first quarter, as it's a leading competitor in AI chips while offering some downside protection with its VMWare business.</p>
<div class="lb-trans"><p>也就是说，博通相比于英伟达甚至 AMD（后者也有强大的 CPU 业务）提供了更为多元化的芯片制造商。这可能是泰珀在第一季度对该公司进行小额投资的原因，因为它是 AI 芯片的领先竞争者，同时通过其 VMWare 业务提供了一些下行保护。</p>
</div><p>Still, Broadcom stock is expensive. It trades for a forward earnings multiple close to 40. That's right in line with Nvidia and slightly less expensive than AMD. The company arguably holds more upside if ASIC designs capture more real estate in data centers over time. Consider the potential efficiency gains of ASICs versus GPUs, which seems likely to happen in the long run.</p>
<div class="lb-trans"><p>不过，博通的股票价格昂贵。其前瞻性市盈率接近 40。这与英伟达持平，略低于 AMD。如果 ASIC 设计随着时间的推移在数据中心占据更多市场份额，该公司的潜在上行空间可能更大。考虑到 ASIC 与 GPU 相比的潜在效率提升，这在长期内似乎是可能的。</p>
</div><p>But investors may have to settle for more slow and steady growth compared to Nvidia or AMD. As such, it's worth keeping an eye on Broadcom's stock to see if it falls back down to a more attractive price before following Tepper into the stock.</p>
<div class="lb-trans"><p>但投资者可能不得不接受比英伟达或 AMD 更缓慢而稳定的增长。因此，值得关注博通的股票，看看它是否在价格回落到更具吸引力的水平之前，跟随泰珀进入该股票。</p>
</div>

Billionaire David Tepper Sold Nvidia and AMD and Is Piling Into This Specialized AI Chipmaker Instead

- David Tepper 的 Appaloosa Management 自 1993 年成立以来年化回报超过 28%。  
- 他近期出售 Nvidia 和 AMD 股票，转而投资 Broadcom。  
- Broadcom 在 AI 芯片领域具竞争力，提供多元化业务。  

亿万富翁大卫·泰珀卖掉了英伟达和 AMD 的股票，转而投资于这家专注于 AI 芯片制造的公司