Meta Can't Buy Its Way to a Top AI Model — How Will Zuckerberg Solve His AI Anxiety?

Hundreds of Billions Spent, Yet a Top-Tier Model Remains Elusive

In 2025, Meta's AI spending has once again set a new record. Zuckerberg confirmed during the latest earnings call that the full-year capital expenditure budget will reach $60 to $65 billion, with the vast majority allocated to AI data center construction and GPU procurement. Yet an increasingly unsettling fact for investors is this — despite pouring in such astronomical sums, Meta's Llama series of large language models has never managed to reach the pinnacle of the industry.

In the latest round of authoritative benchmarks, the Llama 4 series sparked widespread controversy. Llama 4 Maverick briefly topped the LMSys Arena leaderboard shortly after launch, but was quickly suspected of being "benchmark-optimized," with community feedback and real-world user experience severely at odds with its rankings. Meanwhile, OpenAI's GPT-4.1, Anthropic's Claude 4, and Google's Gemini 2.5 Pro continued to lead in reasoning capabilities and real-world application scenarios, making Meta's "open-source champion" narrative look increasingly thin.

Zuckerberg might have genuine reason to worry.

The "Halo" of the Open-Source Strategy Is Fading

The cornerstone of Meta's AI strategy is open source. From Llama 1 to Llama 4, Zuckerberg has repeatedly emphasized the strategic value of the open-source approach: by freely releasing model weights, building a developer ecosystem, and ultimately forming an AI application network centered around the Meta platform.

This strategy did work in the early stages. The release of Llama 2 triggered a wave of enthusiasm among developers worldwide and was once seen as a beacon of hope against OpenAI's closed-source monopoly. Llama 3 sparked an even bigger deployment frenzy in the open-source community, with countless startups building their own AI applications on top of Llama.

But the problem is that the "ecosystem prosperity" brought by open source has never truly translated into commercial revenue for Meta. Developers build products with Llama, and the beneficiaries are themselves — not Meta. Unlike Google, which can deeply embed Gemini into search, cloud services, and the Android ecosystem, or Microsoft, which can directly monetize OpenAI's capabilities through Azure and Copilot, Meta has yet to find a clear pathway to convert Llama's influence into real revenue.

More awkwardly, the technical moat of open-source models is inherently weak. When you release model weights for free, competitors can fine-tune superior versions on top of them. Players like Alibaba's Qwen and Mistral are steadily encroaching on Llama's open-source territory, and the models Meta spent fortunes training have effectively become a "starting line" for others.

The Core Problem: No "Second Curve" to Deliver on the AI Narrative

If you dig deep into Meta's AI predicament, all roads lead to the same fundamental issue — the lack of a second growth curve.

Meta's revenue structure is overwhelmingly dependent on advertising. In 2024, ad revenue still accounted for over 97% of Meta's total revenue. The role AI currently plays within Meta's ecosystem is primarily that of an "efficiency enhancer for the advertising business" — using AI to optimize ad recommendation algorithms and improve ad targeting precision and conversion rates.

This certainly has value. In fact, Meta's strong earnings rebound in 2023–2024 was largely attributable to AI-driven advertising efficiency gains. But this is fundamentally an optimization of existing business, not an incremental growth story. Wall Street's premise for giving Meta a premium valuation is the belief that AI can unlock entirely new revenue streams — whether through AI assistant subscriptions, enterprise AI services, or AI-powered new social products.

However, the realization of these "second curves" remains a distant prospect:

• Meta AI Assistant: Already embedded in WhatsApp, Instagram, and Facebook, with monthly active users reportedly exceeding 1 billion, yet there is still no clear monetization pathway — no subscription model, no standalone paid product.
• Enterprise AI Services: Compared to AWS, Azure, and GCP, Meta has virtually no cloud service infrastructure and cannot charge for API calls the way competitors do.
• AI Hardware: Ray-Ban Meta smart glasses have earned decent reviews, but annual sales of only a few million units are far from constituting a scalable revenue pillar.
• Metaverse and AI Integration: Reality Labs continues to hemorrhage money, with full-year losses exceeding $17 billion in 2024, and there is no prospect of breakeven in the near term.

In other words, Meta spends tens of billions of dollars each year training AI models and building AI infrastructure, but these investments can ultimately only "empower" the advertising business rather than open up independent revenue streams. It's like a company spending a fortune to build a highway, only to have just one of its own cars driving on it.

Capital Markets Are Losing Patience

Wall Street's attitude toward Meta's AI narrative is undergoing a subtle shift.

In early 2024, when Meta announced a dramatic increase in AI spending, its stock price plunged more than 15% at one point. Although it later recovered on the back of strong advertising performance, the core question from investors never went away: Where exactly is this money going? And where are the returns?

Compared with peers, the gap is even more glaring. OpenAI has achieved annualized revenue exceeding $4 billion, with staggering growth rates. Google's AI capabilities are comprehensively boosting the monetization power of its search and cloud businesses. Microsoft's Copilot product suite has delivered tangible ARPU increases for Office 365.

And Meta? Its most presentable AI commercialization achievement remains "more precise ad recommendations." This story is certainly real, but its ceiling is clearly visible — global digital advertising market growth is slowing, and efficiency optimization alone cannot sustain the market's return expectations for Meta's hundreds of billions of dollars in AI investment.

Even more concerning is that Meta's relative lag in model capabilities could eventually undermine its AI advantage in advertising. When competitors possess stronger foundation models, they are fully capable of challenging Meta in the intelligent ad delivery space. Google is already leveraging Gemini to reshape its advertising ecosystem, posing a very real threat to Meta.

Zuckerberg's Choices: What's Needed Beyond Open Source

Facing this predicament, Zuckerberg is not unaware of the problems. He has recently emphasized on multiple occasions that Meta AI will become "the world's most widely used AI assistant" and hinted at the possible introduction of monetization mechanisms. Internally, the company is also accelerating the development of AI Studio — a platform that allows businesses and creators to build custom AI characters — in an attempt to use it as a commercialization breakthrough.

But whether these efforts succeed depends on several key variables:

First, model capabilities must keep pace. Without a top-tier model as the foundation, all downstream applications will lose competitiveness. High hopes are pinned on Llama 5 — Meta is reportedly committing unprecedented computing resources to its training — but whether it can truly match or even surpass the next-generation products from GPT-5 and Claude remains unknown.

Second, finding a monetization path that doesn't rely on cloud services. Meta has no cloud business, which is its biggest structural disadvantage compared to Google, Microsoft, and Amazon. Zuckerberg needs to creatively leverage Meta's unique social graph and user relationship chains to carve out a differentiated AI commercialization path.

Third, controlling the pace of spending. Annual capital expenditure of over $60 billion cannot be sustained indefinitely, even for a company with cash flows as robust as Meta's. If Llama 5 fails to deliver a qualitative breakthrough, market confidence in Meta's AI narrative could suffer a systemic collapse.

Conclusion: There Are No Pay-to-Win Shortcuts in the AI Race

Meta's predicament reveals a harsh industry truth — in the AI race, money is a necessary condition but far from a sufficient one. Computing power can be bought and data can be accumulated, but top-level model design, engineering culture, talent density, and productization capabilities cannot be solved simply by throwing capital at the problem.

The deeper issue is that Meta is fundamentally a consumer social company, and its DNA dictates inherent shortcomings in enterprise AI services and foundational model R&D. Zuckerberg cleverly sidestepped some competition with his "open-source" strategy, but open source itself is not a business model — it is merely a means of gaining influence. When influence cannot be converted into revenue, even the grandest AI narrative will eventually face the market's reckoning.

Zuckerberg's AI mega-bet is entering its most critical validation phase. The window he has left to prove himself may be shorter than anyone imagines.