A recording cost the company 70 billion yuan

In recent days, a piece of news has sent shockwaves through the global tech industry.

On July 2nd, an internal Meta meeting took place. Reuters obtained a recording of the session.

After the recording was leaked, Meta's stock price fell by roughly 5% immediately. Overnight, over $700 billion in market value vanished.

Why did this recording have such a massive impact? The fundamental reason is that Mark Zuckerberg openly admitted to a decision-making error during the meeting.

He stated: "Over the past four months, the development trajectory of AI agents has not accelerated in the way we expected." He also mentioned that the company's large-scale organizational restructuring centered on AI "could have been executed more cleanlyly," and the executive team had misjudged the timing of the transformation.

Pay attention to these phrases he used: "Time's wrong" and "Judgement's wrong." It was not external changes or technical challenges that were at fault, but rather "the timing was wrong" and "the judgment was wrong."

For any enterprise, a strategic decision error is fatal.

As a leading company in the AI sector, its leader admitting a strategic misstep undoubtedly sent two extremely clear signals to the market:

First, Meta indeed made a major strategic deviation;

Second, the real-world implementation of AI is falling short of expectations. This delivered a strong blow to the entire AI sector. When a leading figure in the AI field candidly acknowledges that technology is not a universal solution, market doubts about whether an AI bubble exists are sharply amplified.

So, what exactly was the "wrong decision" Zuckerberg referred to?

01

A Decision Driven by Fear

Let's rewind the clock back to January of this year.

In the recording, Zuckerberg laid out the context of the decision at that time: Back then, Anthropic's Claude Code saw explosive growth among programmers, with its growth rate far outpacing Meta's in-house coding tools.

Meta's core management team fell into a panic, concluding that without rapidly ramping up AI efforts, the company would be completely left behind by competitors.

This panic subsequently spawned a series of aggressive decisions.

In May, Meta laid off roughly 10% of its global workforce, nearly 8,000 people; at the same time, it froze 6,000 vacant positions and reassigned over 7,000 employees to AI-related departments, tasking them with full-time work on training AI tools that could potentially replace themselves in the future.

This move sparked huge controversy at the time. But Zuckerberg held a firm stance, arguing that companies in the AI track must streamline their staff, and failing to slim down would cause them to fall behind in the industry competition.

Theoretically, the original intentions of these two measures seemed impeccable: Layoffs can drastically cut labor costs, while redeploying manpower to support AI operations holds the potential to boost efficiency, making it a win-win strategy for reducing costs and increasing efficiency.

However, reality delivered a completely different outcome.

Just as Zuckerberg admitted in the recording, the AI agents failed to gain traction. Money was spent, employees were let go, but the product never materialized.

Even more devastating was the collapse of morale among Meta's employees.

8,000 people lost their jobs. 7,000 others were forcibly moved into roles they never imagined taking on. Many didn't even know their new responsibilities, who they reported to, what they were supposed to do, or how their KPIs would be evaluated.

A senior engineer with 11 years of tenure at Meta posted a message on the internal forum that was read and shared by nearly 20,000 people within two days: "I don't want to live in a world where humans are exploited as AI training data."

The newly formed AI department was bluntly described by internal engineers as a "soul-crushing concentration camp." When employees can no longer perceive the value and meaning in their work, the collapse of productivity becomes only a matter of time.

Even Meta's Chief Technology Officer Andrew Bosworth criticized without mincing words: Morale inside Meta is almost hitting an all-time low.

As you can see, Fear is not a strategy. Decisions driven by fear never lead to good outcomes.

02

Is Meta the Only One?

Meta's story is a microcosm of global tech companies. In the AI arms race, everyone is ramping up their investments, and no one dares to be the first to hit the brakes.

An article from Huxiu shows that: In 2026, the combined capital expenditure of the world's four major tech giants — Meta, Microsoft, Google, and Amazon — will reach $725 billion, a staggering 77% year-on-year surge from the $410 billion recorded in 2025.

What does $725 billion mean? This figure exceeds the annual GDP of the vast majority of countries worldwide. The giants are pouring massive amounts of capital into GPUs, computing power, large models, and infrastructure, but what is the return? The entire industry's incremental revenue directly generated by AI only amounts to tens of billions of dollars.

Investing thousands of times more capital while generating a fraction of the returns results in a severely imbalanced input-output ratio. This is by no means rational commercial investment; it's a typical case of involution and a prisoner's dilemma.

What is the prisoner's dilemma?

It essentially refers to a trap where individuals, unable to trust one another, each make the "most rational" self-protective choices, ultimately leading the collective to the "worst possible" outcome.

All four giants are fully aware that this endless burning of money cannot be sustained, but none dares to be the first to stop. Because everyone is betting that one extra GPU or one additional model iteration by a competitor could become the deciding factor in determining the final industry landscape.

As a result, everyone is forced into passive involution, knowing full well it's a war of attrition yet having no choice but to follow along.

This is the current reality of the AI industry: The technology is not yet mature, but capital has already gone wild; value has not been realized, yet involution has already begun.

03

The More Anxious You Are, the More You Need to Grasp the Essence

Meta's lesson teaches us that all AI strategic failures stem from failing to see AI's true capabilities clearly and misunderstanding the core relationship between humans and AI.

There are two fundamental truths about AI that you must understand.

First, AI cannot replace human judgment.

When massive amounts of information flood in, humans are ultimately the ones who need to make decisions. The performance of AI across various industries has been less than satisfactory to date. Once it enters the deep waters of business, facing complex, non-standardized core business scenarios, AI's performance falls far short of that of senior industry experts.

As Anthropic emphasizes, AI must be integrated into specific business scenarios, but this inevitably comes with profound transformations to organizational structures, management models, and management practices.

In this process, the judgment and decision-making capabilities of AI simply cannot replace the experience and insights of seasoned engineers.

Second, AI should serve humans, rather than humans serving AI.

Meta's choice to make humans work for AI was a directional mistake. The correct path will always be AI working for top experts.

We must recognize the inherent value of AI: It excels at replacing repetitive labor, accelerating standardized processes, and integrating massive volumes of basic information. But the prerequisite for all this is that humans set a precise, correct framework for it.

AI operates within this framework, and the outputs it generates still require human final judgment and verification. Without these steps, AI cannot deliver substantial value.

Let me give you an example. A friend of mine who runs a quantitative fund used AI to build a model for stock analysis and trading. On its first run, the AI successfully identified 10 stocks that hit their daily upside limit, generating considerable returns. However, by the second week, all 10 of those stocks hit their daily downside limit simultaneously.

After investigation, it was discovered that the model he designed had 12 steps in total, but the AI stopped and returned results at the eighth step, failing to execute the entire process. In other words, the AI "slacked off."

While the exact reason for the AI's slack behavior remains unknown, the consequence was that almost all of my friend's profits vanished, and he only managed to retain 10% of his gains in the end.

What does this illustrate? It shows that AI can slack off, generate hallucinations, provide large amounts of incorrect or weakly correlated information, and even exhibit somewhat "emotional" outputs.

This loss made my friend fully realize that he must supervise every single step of the AI's execution throughout the process to ensure its logic is not tampered with.

Today's AI is exactly like this: If it is completely separated from our design framework, our supervision, and our final checks, it cannot deliver reliable results.

04

My Three Judgments About AI

Based on these observations, I have the following three fundamental judgments about AI:

1. AI cannot take on the function of strategic decision-making

If you attempt to let AI formulate your company's strategy, the consequences will be disastrous.

Strategic decision-making requires far more than just data analysis; it demands an understanding of the essence of business, insight into human nature, and foresight into the future.

AI cannot do these things. AI can provide you with data and analysis, but the final decision must be made by humans.

We once conducted an experiment: We input the same set of materials into AI to let it draw conclusions, while a senior consultant performed independent analysis based on the exact same data. The results were shocking: The two conclusions were completely at odds with each other.

Many people prefer to trust AI's conclusions over the judgments of human experts. They believe AI is more objective, rational, and error-free.

But they overlook a key point: AI's judgment capabilities depend on the data fed into it and the framework set for it. Once the framework itself is flawed, the outputs AI generates will be a disaster.

2. AI cannot replace top experts in any industry

At least in the foreseeable future, this conclusion remains valid. The value of top experts lies not in repetitive labor, but in the things that cannot be standardized or algorithmized. For example, intuition, experience, and creativity.

AI can empower experts to amplify their efficiency, but it will never replace the experts themselves.

3. AI's role at the organizational level is extremely limited

When it comes to delving into business processes, transforming organizational structures, optimizing performance management, and personnel allocation, AI's efforts have all remained at the surface level and have not yet touched the core.

Peter Drucker said: "The essence of management is to unleash the goodwill and potential of people." To inspire a person, you need leadership, you have to start from the depths of human nature and strike a chord in their hearts. Only humans can create such interactions, and AI does not possess that kind of warmth.

So, what is AI's true value? The answer is: To help experienced individuals enhance their efficiency.

But please note that the prerequisite is "being experienced." For a complete beginner with no experience, the massive amount of information provided by AI will only leave them more confused and adrift, because they lack the necessary judgment capabilities.

But for an expert with rich experience, AI can help them automate repetitive tasks, freeing them up to focus more energy on truly valuable work.

Therefore, AI is indeed effective at empowering individuals to boost their efficiency. But when it comes to improving organizational efficiency and the overall performance of a company, its role remains limited.

Furthermore, there is another practical issue that cannot be ignored: Cost.

Many people think that using AI is cheaper than hiring humans. This is a huge misconception. You might think laying off 8,000 employees saves billions of dollars, but the token costs consumed by AI operations are equally staggering, and could even far exceed labor costs.

In fact, The cost of employing people is far lower than the cost of maintaining AI operations. Even super-large enterprises find the current return on AI investment unbearable, let alone small and medium-sized companies.

Finally, let's summarize: Meta spent $700 billion to learn a lesson. The core of this lesson is not that "AI doesn't work," but that companies must not be superstitious about AI. Don't let technological anxiety replace strategic thinking. Don't take fear as your guide for action. Let AI serve you, rather than you serving AI.

Because the ultimate essence of all technologies is to empower humans, not to replace them. The AI bubble will eventually deflate, and what will truly remain forever are the organizations and talents who understand business, know how to judge, and master the art of trade-offs.

This is the most precious insight that Meta's lesson has given us.

Sources: Reuters, TechCrunch, Huxiu, Li Shanglong's WeChat Channel.

This article is from the WeChat public account "Zhang Lijun", authored by Zhang Lijun, and published by 36Kr with authorization.