In May 2018, Pan Luan penned an article titled "Tencent Has No Dreams," arguing that Tencent was losing its edge in product innovation and entrepreneurial spirit, morphing instead into an investment firm. The article sparked widespread discussion across China's internet, even prompting a response from Pony Ma, Tencent's founder.

Eight years later, a similar narrative is unfolding in Silicon Valley—but this time, the spotlight is on Google.

Google appears to be losing faith in pioneering AI research, transitioning instead into an AI infrastructure provider. It developed the Transformer architecture, acquired DeepMind, launched the TPU project, and built a chatbot a year before ChatGPT emerged. Yet now, it sells computing power to its fiercest competitors, watches top talent depart in droves, and has seen its flagship model, Gemini, fall out of the global top ten.

However, Google's situation fundamentally differs from Tencent's. Tencent's investments are defensive, aimed at protecting its core traffic moat. Google's investments, on the other hand, are a hedge against its own inability to execute in cutting-edge fields. Google hasn't abandoned its dreams; it has simply outsourced them.

1. DeepMind: From Pioneer to Follower

On August 5, Google announced a sweeping reshuffle of DeepMind's leadership. While the news itself isn't complex, the signal it sends is crystal clear.

• Demis Hassabis, DeepMind's co-founder and CEO, will step back from daily operations.
• Jeff Dean, Google's chief scientist for 27 years, co-founder of Google Brain, and architect of the TPU project, left to launch Discovery Loop.
• Joining Dean in departure were Sanjay Ghemawat, Quoc Le, and Oriol Vinyals—three core researchers at the Google Fellow level.
• DeepMind CTO Koray Kavukcuoglu will take over daily operations, reporting directly to Sundar Pichai.

SemiAnalysis bluntly stated in its report: "In every respect, we believe DeepMind is no longer a cutting-edge lab. Google will continue to release models, but the probability of them reaching industry-leading levels (SOTA) again has dropped to zero."

This decline didn't happen overnight. A timeline of Google AI's brain drain over the past two years reveals a troubling pattern:

• 2021: Noam Shazeer, a co-author of the Transformer paper, left Google to found Character.AI after the company refused to release his chatbot, Meena.
• August 2024: Google "bought back" Shazeer for $2.7 billion, appointing him as co-lead of Gemini.
• June 18, 2026: Shazeer left Google again, this time to join OpenAI. By now, all eight authors of the Transformer paper had departed.
• June 19, 2026: John Jumper, the 2024 Nobel laureate in chemistry and a core leader of AlphaFold, left DeepMind to join Anthropic.
• August 5, 2026: Jeff Dean, the symbol of Google's engineering culture and employee #30, announced his departure. Demis Hassabis stepped down from daily management.

A Google employee shared with reporters that jokes about Jeff Dean were ingrained in the company's engineering culture. Some claimed he wrote code faster than others could read it, never needed to check syntax before compiling, and if the compiler reported an error, it must be the compiler's fault.

Behind these jokes lies Jeff Dean's monumental contribution to Google's technical ecosystem over the past 20+ years—MapReduce, Bigtable, Spanner, TensorFlow. Any one of these projects alone would be industry-changing. His departure marks the loss of half of Google AI's kingdom.

But the greatest irony isn't who left, but what Google once had.

• In 2017, eight Google researchers published "Attention Is All You Need," proposing the Transformer architecture. This paper has since been cited over 260,000 times, becoming the technical foundation for almost all modern large language models—GPT, Gemini, Claude, DeepSeek.
• In 2021, DeepMind researcher Tibo's team built a chatbot called LMChat, a full year earlier than ChatGPT. But Google, fearing disruption, refused to release it.
• Tibo later wrote on X: "Google was too nervous to release it, and DeepMind was blocked from shipping products that could disrupt Google." He left for OpenAI in July 2024 and is now head of Codex.

The company that invented AI now fears being disrupted by it—a mirror image of Tencent's hesitation in the short-video sector. Weishi, Tencent's short-video platform, launched in 2013, disbanded in 2015, and shut down in 2017, when Douyin's daily active users were still below 1 million.

2. When the CFO Wins the AI War

If DeepMind's decline is "front-end bleeding," then Google Cloud Platform's (GCP) rise is "back-end blood transfusion." The two events are two sides of the same coin.

SemiAnalysis's report reveals a critical internal power shift. Gemini and GCP once competed fiercely for computing power allocation, but now the conclusion is clear—Thomas Kurian, CEO of GCP, has won.

With $200 billion in external sales and an EBIT margin exceeding 30%, compared to Gemini's first-party business at just $12 billion, Google's management has made its choice: prioritizing GCP's short-term financial gains over long-term competitiveness in cutting-edge fields.

Even more shocking is whom Google sells its computing power to. From Q3 2026 to Q4 2027, over 20% of total TPU shipments will go directly to Anthropic. This doesn't include the hundreds of thousands of TPUs GCP currently rents to Anthropic, nor the hundreds of thousands more committed to Anthropic and Meta over the next six quarters.

Google invented the Transformer; Anthropic uses it for Claude. Google spent $2.7 billion to bring Shazeer back for Gemini, only to watch him defect to OpenAI. Now Google sells TPUs to Anthropic to train Claude faster.

If you've heard Thomas Kurian's interviews, you know he's no believer in artificial general intelligence (AGI). In a podcast, he argued that TPUs becoming "general infrastructure" supporting Citadel, the Department of Energy, and general high-performance computing is a good thing. When asked why Google sells computing power to Anthropic despite competing with Gemini, he said it's the inevitable result of Google being a "platform company."

This isn't unreasonable. But the question remains: When a company's CFO wins the AI war, is it still an AI company?

【Chart】chart1

3. First-Ever Negative Cash Flow: Spending Like a Startup, Valued Like Old Money

In Q2 2026, Alphabet delivered a "contradictory" earnings report:

The most noteworthy data points are the last two rows. For the first time in Alphabet's history, quarterly free cash flow turned negative. The market's most reliable cash generator now spends more than it earns.

Of the $9.11 earnings per share (EPS), $6.26 came from unrealized gains in its investment portfolio—primarily from rising stakes in SpaceX and Anthropic. Excluding this, Google's "real" business profitability is far less impressive than it appears.

This mirrors Tencent's situation in 2018. Back then, Tencent's profits increasingly came from investment gains rather than its core business, prompting Pan Luan to question whether it had "become an investment company." The difference is that Tencent's investments were a logic of traffic monetization, while Google's are a logic of technological hedging.

【Chart】chart2

4. But Google Hasn't Given Up on Its Dreams—It Just Outsourced Them

This is the fundamental difference between Google and Tencent, and the point this article must clarify.

Tencent's investment logic is "social traffic monetization"—outsourcing non-core, non-specialized projects through investment, essentially a "startup tax." Investing in JD.com, Meituan, and Pinduoduo was because Tencent couldn't do these things itself but could empower them with traffic. These investments defend Tencent's social moat.

Google's investment logic is entirely different. It doesn't invest in "things it can't do," but in "things it believes in but can't execute." For cutting-edge technologies, Google invests aggressively:

Let's expand this list:

• Anthropic: Alphabet committed $40 billion in investment ($10 billion initial), binding exclusive TPU supply. The stake is valued at ~$124 billion. Ironically, Anthropic's Claude is using Google's TPUs to outperform Google's Gemini.
• SpaceX: Google invested $900 million in 2015, now worth $94.1 billion—a 100x return. SpaceX's xAI launched Grok 4.6 today—another AI sector Google indirectly bets on.
• Waymo: Raised $16 billion externally in early 2026, valued at $126 billion. Pichai said it will materially contribute to Alphabet's financials starting in 2027.
• Isomorphic Labs: After stepping down from DeepMind's daily operations, Hassabis will focus on this AI-driven drug discovery company.
• Discovery Loop: Jeff Dean's new venture, funded and supported with computing power by Alphabet. It uses AI to automate the entire scientific research process.
• GV + CapitalG + Gradient Ventures: These three funds manage over $8 billion, investing in hundreds of companies ranging from AI programming tools to quantum computing.

See the difference?

Tencent invested in JD.com because it couldn't do e-commerce well; Google invested in Anthropic because it couldn't do AGI well. The former is a capability issue; the latter is a belief issue. Tencent lacks execution; Google lacks that "religious conviction" to "keep going until hitting a wall."

SemiAnalysis put it well: "Google's problem isn't Jeff Dean or Noam Shazeer, but its extremely bureaucratic, slow-moving, and strategically timid corporate culture."

So Google made a rational choice: If it can't run at the cutting edge, it will invest in those who can. Sell TPUs to them, rent computing power to them, earn money from them via GCP, and then invest in their equity via GV. Google has become the arms dealer of the AI era—while holding stakes in both sides of the war.

【Chart】chart3

5. Gemini Falls Out of Top Ten, but Alphabet Ranks First

This is the complete picture of "Google without dreams":

• Google LLC (operating company): Ads + Search are cash cows; GCP is the growth engine; Gemini is... a product that still releases updates but no one expects much anymore. Gemini 3.5 Pro has been repeatedly delayed and never launched; Gemini 3.6 Flash ranks ~21st on Artificial Analysis's intelligence index, behind Muse Spark 1.2, Grok 4.5, and even several Chinese open-source models. DeepMind is no longer a cutting-edge lab.
• Alphabet Inc. (holding company): Holds $94.1 billion in SpaceX, $124 billion in Anthropic, $126 billion in Waymo. Q2 2026 investment gains were $98 billion, exceeding same-period operating profit of $40.8 billion. External equity investments exceed $50 billion. It has essentially become a top-tier venture fund with a search engine.

Breaking down that $9.11 EPS in Q2 2026: $2.85 came from the "real" business, and $6.26 came from investment gains. The market is valuing Google at just 17.75x earnings—no longer a growth stock valuation, but a signal it’s being priced as a value stock.

Is this reasonable?

From one angle, extremely so. If Alphabet’s AI strategy isn’t “build the best model” but rather “build the best AI infrastructure + own the best AI companies,” its moat runs deeper than any single lab’s:

• No matter who wins the model race, they’ll need TPUs—Google sells them.
• No matter who wins the model race, they’ll need cloud infrastructure—GCP provides it.
• No matter who wins the model race, Google owns its stock—directly sharing the returns.

This is a “bet on everyone” strategy. The upside: nearly impossible to lose. The downside: nearly impossible to win the biggest.

But from another angle, this strategy carries a fatal hidden cost: talent. When the smartest minds realize Google isn’t pursuing AGI but GCP’s profit margins, they leave. Shazeer left. Jumper left. Jeff Dean left. Who’s next?

As that TiMedia article put it: "In AI, the real moat isn’t data, compute, or even model architecture. It’s the people willing to stay and push boundaries day after day. And Google is losing exactly those people."

Conclusion

Google hasn’t lost its dream. It’s just moved it—from DeepMind’s labs to Alphabet’s portfolio.