AI in the US stock market is approaching a new watershed moment.

If you only look at the index, you might think everything is still going smoothly, but if you dig deeper, you'll find two forces fiercely battling it out:

On one hand, the once invincible "OpenAI chain"—Nvidia, SoftBank, Oracle, and CoreWeave—has seen a continuous decline.

On the other hand, Google and Broadcom, representing "Google Chain," continued to rise.

It's worth noting that Google's stock price hit an all-time high of $328.83 per share during trading. Furthermore, Berkshire Hathaway initiated its first investment in Alphabet, Google's parent company, in the third quarter of 2025, with a holding value of approximately $4.34 billion.

In the new logic of "AI deals," Google seems to be replacing Nvidia and trying to be crowned the new king of AI.

1. Nvidia's moat has been breached by TPU.

Over the past three years, the biggest belief in the US stock market has been: "No matter who wins, Nvidia will make money." This "shovel-selling" logic is simple and crude, yet unbreakable.

But now, there's no longer just one company making shovels.

Google's series of moves in November 2025, especially the fact that the training of Gemini 3 Pro is based on the sixth and seventh generation TPUs (codenamed Trillium and Ironwood), are actually announcing to the market that training top-notch large models no longer requires Nvidia GPUs.

To understand this shift, we need to go back to the underlying logic of the technology.

Early large models (such as BERT and GPT-2) relied on flexibility, and their structures were not yet standardized, often requiring custom layers, dynamic shapes, and debugging iterations.

The versatility of GPUs and their mature toolchains (such as CUDA, cuDNN, and NVIDIA Nsight) make them more suitable for this exploratory phase.

，Llama、Gemini等基本都基于 Transformer架构。

The current mainstream large model architecture tends to be standardized , with Llama, Gemini, and others basically based on the Transformer architecture.

The core of the Transformer architecture is large-scale matrix multiplication plus the Attention mechanism, which is exactly the computation mode that the TPU systolic array is best at.

In short, GPUs are like a versatile construction team, capable of building any kind of structure; TPUs are like "automatic bricklaying machines designed specifically for Transformers," offering both speed and efficiency when building large-scale projects with uniform brick sizes.

Google's newly released Gemini 3 Pro relies entirely on its self-developed TPU cluster for training and inference. This proves that the ASIC chip solution with hardware and software co-design has been fully implemented for trillion-parameter MoE (Hybrid Expert Model) and ultra-long context (Infinite Context) scenarios.

In other words, large models can be trained on Google's self-developed TPUs, no longer relying on Nvidia's GPUs.

Previously, when invited to the "Bg2 Pod" talk show, Jensen Huang confidently stated, " Even if competitors' ASIC chips are free, customers will still choose Nvidia."

The core of Huang's view is that the operation of hyperscale data centers is limited by electricity, and the customer's core goal is to maximize the conversion of limited electricity into revenue.

Because Nvidia systems have an order-of-magnitude advantage in "tokens generated per unit of energy consumption," choosing an Nvidia solution can generate several times, or even tens of times, more revenue with the same power consumption. Therefore, the opportunity cost of forgoing such a huge revenue opportunity far exceeds the purchase cost of the chip itself.

This logic used to hold water, but it was shaken in November 2025.

According to Guoxin Securities' calculations, Google's TPU v6e is already close to Nvidia's Blackwell in terms of cost per unit of computing power (TOPS/USD).

Furthermore, based on currently available data and evaluations, in the specific scenario of "training large models such as standard Transformer/LLM", the TPU has an overall advantage in performance/power consumption (Perf/Watt) compared to the same generation of data center GPUs, and its energy efficiency is more outstanding.

Downstream customers also began to turn against them.

Google Cloud's officially launched TPU v7 Ironwood instance is no longer for Google's own use but is now open to the industry. Google and Anthropic have reached a cooperation agreement for millions of TPUs, which are expected to bring tens of billions of dollars in additional annual revenue to Google Cloud once fully operational.

Recently, media reports indicated that Meta is in talks with Google to use billions of dollars worth of TPU chips in its data centers in 2027, and plans to lease chips from Google Cloud next year, which could bring Google billions of dollars in additional revenue.

If both Meta and Anthropic begin to shift their focus, Nvidia's 90% monopoly in the AI ​​chip market will inevitably be diluted.

The ironclad rule that "if you're building a large-scale model, you must stockpile Nvidia GPUs" has been broken. This shift from "monopoly" to "optional" is the biggest blow to a hardware company that enjoys extremely high premiums.

II. AI Commercialization Begins to Close the Loop

Besides the "fundamental impact" of chip technology, Google's Gemini3 also surpasses mainstream large-scale models in terms of large-scale design.

Google released Gemini 3 Pro on November 19, 2025, and its flagship version, Gemini 3 Pro, ranked first on the LMARaena leaderboard. This model outperformed Grok-4.1, Claude-4.5, and GPT-5 in multiple fields, including text, vision, WebDev and coding, mathematics, and creative writing.

Benchmark scores and large model performance are only one aspect. At present, investors are more focused on the commercialization capabilities of large model companies like OpenAI.

There is currently a strong "AI bubble" theory circulating on Wall Street.

Many organizations question whether AI companies, led by OpenAI, can actually make money, and whether the future they depict is a mirage or a real vision.

Multiple studies and commentaries have pointed out that the capital expenditures of the "Seven Golden Flowers" (a group of leading AI companies) and a host of cloud vendors on AI+data centers are approaching $300 billion or even higher annually, and the market is increasingly worried that the "speed of realizing AI benefits" cannot keep up with the "speed of burning money on hardware".

In this respect, OpenAI faces a huge dilemma. Sam Altman's vision is grand, but the reality is harsh: although OpenAI has a large user base, its consumers' willingness to pay is relatively limited, its business-to-business (B2B) commercialization path is always constrained by Microsoft, and it lacks a sufficiently strong "traffic moat".

Google can then embed the capabilities of Gemini 3.0 into its ubiquitous ecosystem—search, YouTube ads, Workspace, Gemini Enterprise, Vertex AI—all of which can become vehicles for monetizing AI.

This not only enables a closed business loop but also shortens the time lag between "investment, validation, and revenue".

While other tech giants are still "betting on a business model that is not yet fully proven with increasingly expensive capital expenditures," Alphabet has finally made its story of "technological leadership + monetization path" complete with Gemini 3.0.

III. Endgame and Investment Opportunities

For investors, this is not just a debate over technological approaches, but also a real adjustment in return expectations.

Miao Tou believes that Google will experience a classic "Davis Double Play" in the short term.

On one hand, there's the recovery in valuation. For the past two years, Google's valuation has been suppressed due to fear of falling behind, as it was seen as a latecomer to the AI ​​field. Now, with the Gemini 3 reclaiming its throne and the TPU demonstrating hardware independence, Google has regained its voice in AI, and the capital market should rightfully award it a higher premium.

On the other hand, there is the realization of performance. The revenue from external cloud customers brought by TPU v7 (such as orders from Anthropic), as well as the increase in ARPU (average revenue per user) after Gemini is fully integrated into advertising and enterprise services, will be reflected in the profit statement.

Institutional investors have already reacted. In a recent research report, Shenwan Hongyuan directly raised its 2026 net profit forecast for Google from $123 billion to $137 billion, and gave it a target valuation of 34 times PE.

In contrast, companies like Nvidia and Oracle, which are part of the "OpenAI chain," will face significant valuation pressure in the short term. Once the market confirms that "accumulating computing power doesn't necessarily require Nvidia," then Nvidia's days of enjoying unlimited premiums will be over.

However, the AI ​​war is far from over.

Judging from the current technological evolution and business landscape, the "battle of the big models" is unlikely to end with one company completely eliminating its opponent. Instead, it will evolve into a two-giant landscape of OpenAI + Microsoft vs. Alphabet (Google).

In consumer and office scenarios, Alphabet leverages Gemini3 to deeply integrate with Search, YouTube, Workspace, and Android, while Microsoft uses Copilot to integrate OpenAI models into Office and Windows. Whoever has higher daily active users and retention rates will win this battle.

In the enterprise and developer ecosystem, Azure+OpenAI on one side and Google Cloud+Gemini on the other side form a binary structure similar to iOS/Android. Most enterprises will choose one as the primary and one as the backup, rather than betting on only one side.

However, it should be noted that Google's execution capabilities in cloud and enterprise software have historically been significantly weaker than Microsoft's. While Gemini3 is technically superior, whether this full-stack solution can be truly implemented across various industries will require two to three years of practical verification.