Recently, various investment banks in the U.S. have been vigorously promoting TPU chips. Since Google was targeted by Buffett, its stock price has been rising steadily 📈, nearing the $4 trillion mark. Today, let's briefly analyze the potential of TPU and the companies that may benefit or be negatively impacted.

I have recently followed Buffett in betting on Google's TPU. I believe this is not just about chips, but a "conspiracy" in cloud computing.

You may have heard of GPUs—NVIDIA (#NVDA) has become the "money printer" of the AI era thanks to them, with its market value once soaring to $5 trillion. But TPU is Google's secret weapon that has been in development for 12 years. In the past, it was only used internally, for example, to train the Gemini large model, optimize YouTube recommendations, and support AI features in Google Search. But now, Google suddenly says, "Hey, Meta, Anthropic, why don't you try my chips too?"

AI chip competition may unfold among three major players: NVIDIA, AMD, and Google. Google's TPU could be a game-changer in the entire AI chip landscape.

What is TPU? Simply put, it is "Google's self-developed AI accelerator card."

You can think of it as a "competitor" to NVIDIA's GPU, but deeply optimized for Google's AI tasks. TPU is not a general-purpose chip; it is an ASIC (Application-Specific Integrated Circuit), like custom running shoes designed for a marathon—other paths may not work well, but within Google's AI ecosystem, it is fast and energy-efficient.

In the past, TPU only ran in Google's own data centers and was never offered externally. But now, the winds have changed. A report from Morgan Stanley shows that Google has signed a large order for 1 million TPUs with #Anthropic and is negotiating with #Meta for next year's rental and the following year's direct purchase for training the Llama model—note that this is for training, not just inference. Training requires much higher chip performance, indicating that Google's Gemini has already gained public recognition for its practical use.

What financial impacts could this have on Google? According to Morgan Stanley's estimates: for every 500,000 TPUs sold externally, Google's cloud business revenue could increase by $13 billion (+11%) by 2027, and earnings per share (EPS) could rise by $0.37, equivalent to a 3% overall EPS increase. Don't underestimate 3%—for a company the size of #GOOGL, this represents a real profit increment and could push the valuation multiple higher.

Currently, Google Cloud is still catching up to AWS and Azure, but TPU could be its "differentiation ace": while others use expensive and hard-to-get NVIDIA GPUs, Google says, "I not only have cloud services but also my own, more cost-effective AI chips that can seamlessly integrate with PyTorch." This is a huge temptation for most AI customers.

📝 Which companies are affected positively or negatively?

Google (#GOOGL): Clearly the biggest beneficiary. Not only will cloud revenue grow, but it can also dilute TPU R&D costs, forming a "chip-cloud-large model" closed loop. This is a classic positive flywheel effect.

NVIDIA (#NVDA): Short-term impact is minimal, but long-term investor sentiment may be affected. After all, NVDA is expected to sell 8 million GPUs by 2027, and even if Google sells 1 million TPUs, it would only be a small portion of the market. However, in the long run, if the TPU ecosystem takes off, NVDA's "monopoly" position will be weakened, which could have a negative impact in the long run.

AMD: A bit awkward. It originally hoped to gain market share with MI450, but now Meta has turned to try TPU. AMD has always emphasized "versatility + cloud compatibility," but now TPU has also become a "quasi-commercial" product, diluting this advantage, and Su may be anxious again.

Broadcom (#AVGO): It may be the invisible biggest beneficiary. Because TPU is designed by Google and manufactured by Broadcom. Morgan Stanley estimates that Broadcom will produce 1.8 million TPUs for Google by 2025 and ramp up to 3 million by 2027. The hotter TPU gets, the steadier Broadcom's orders will be. Although the profit margin for this business may not be as high as NVDA's, the large volume, stability, and binding with top clients make it a quality asset.

⚠️ My three concerns, as mentioned in the Morgan Stanley report, highlight some risk points that need attention:

1️⃣ Business Model: Is Google selling chips directly? Or just renting them? If selling, it will be accounted as hardware revenue (low margin); if in the form of cloud services, it will be counted as high-margin cloud revenue. This has a huge impact on profit structure.

2️⃣ Developer Ecosystem: TPU has been criticized in the past for being "difficult to use." Now Google has created a "TPU Command Center" that can also be called using PyTorch, which is a significant improvement. But can it really eliminate developers' reliance on NVDA's CUDA? That remains to be seen based on actual experience.

3️⃣ How much will Meta actually buy? If it's just a "trial," the impact will be limited; but if Meta truly hands over large-scale training of Llama 4 or 5 to TPU, it would be equivalent to giving TPU an "industry certification," prompting other AI companies to follow suit in large numbers.

I have recently increased my bets on Google and Broadcom. I am not betting on "whether NVIDIA will be affected," but rather on "whether Google can run out a second curve." The external sale of TPU is a key step for Google to transform from an "AI user" to an "AI infrastructure provider." If Google can indeed sell 2 to 3 million TPUs by 2027, not only will the cloud business valuation need to be reassessed, but the entire #GOOGL AI narrative will shift from "follower" to one of the "rule-makers." Therefore, I am now more firmly holding #GOOGL and even preparing to add a bit more during a pullback. This is not about speculating on concepts, but about betting on the rise of an undervalued integrated AI platform of hardware + software + cloud. After all, in this marathon of AI, having just the engine (GPU) is not enough; you also need the whole car—and Google is quietly building its own. 🧐

If you are still limited by the inconvenience of opening a U.S. stock account domestically, you can try using U to trade U.S. stocks for a smooth experience. I am personally using the #RWA tokenized platform #MSX to participate in the U.S. stock market: http://msx.com/?code=Vu2v44

Currently, you can join our U.S. stock community for free, read first-hand reports from overseas investment banks for free, limited to 10 people per week. You can message me to fill out a form to enter the U.S. stock discussion and exchange community (the number of new members has been increasing recently, and the assistant's review takes time, thank you for your understanding) 🙏

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。