Meta released its quarterly earnings report, showing revenue of $60.8 billion, a 28% year-over-year growth, but free cash flow amounted to only $784 million due to significant investments in AI infrastructure.

Author and source: Letter AI

In the second quarter, Meta generated $31.862 billion in operating cash flow, spent $31.078 billion on building data centers and purchasing servers, and ended up with $784 million in free cash flow.

Keep in mind that Meta's advertising engine is still generating substantial revenue.

In the second quarter, Meta's revenue reached $60.801 billion, a 28% year-over-year increase; advertising revenue amounted to $59.363 billion, up 27% year-over-year. Both ad impressions and average prices rose, driving revenue above Wall Street expectations.

The profits from advertising machines are being completely consumed by AI.

The company also raised its lower bound for 2026 capital expenditures from $125 billion to $130 billion, with the upper limit still potentially reaching $145 billion.

A earnings report that exceeded expectations was met with an nearly 8% decline in stock price. Meta's stock closed at $585.61 on the day, and after hours, the drop approached 9% following the earnings release.

During the call, Bank of America Securities analyst Justin Post directly asked Zuckerberg: A year ago, Meta replaced the senior leadership of its AI lab—when will Wall Street see a noticeable acceleration in the release speed of models, chips, and other products?

Zuckerberg replied that he is "quite satisfied" with the current progress of the lab. Muse Spark and Muse Spark 1.1 are merely early achievements in Meta's climb up the ladder of model capabilities, as the company is training larger, more powerful models.

He did not announce the release date for this flagship model or present any new benchmark results, only saying that more information will be coming soon.

A year ago, Llama 4 underperformed, so Zuckerberg invested $14.3 billion in Scale AI, brought in 28-year-old Wang Tao at Meta, and entrusted him with rebuilding the AI system.

There are many similarities between Wang Tao and Zuckerberg—both dropped out around age 19 to start businesses, both became billionaires in their early twenties, and both believe in founder-led centralized power and rapid execution. At 40, Zuckerberg has effectively entrusted Meta’s future to his younger self.

A year later, Wang Tao had restructured Meta’s AI team and launched the first models and products. However, the flagship model Watermelon, which was expected to lead Meta’s counterattack, was still under training.

Competitors are still accelerating their iterations, while the cost for Meta to wait for this answer continues to rise.

Unless Wang Tao delivers the flagship model, Zuckerberg will struggle to demonstrate what these funds, this restructuring, and another year of waiting have achieved. Changing personnel or scaling back at this point would amount to admitting that the previously incurred costs are unlikely to be recovered.

Wang Tao has become Zuckerberg's largest sunk cost.

Wrong savior?

Wang Tao was certainly not idle.

Zuckerberg’s choice of Wang Tao is itself a somewhat risky decision. Scale AI has long provided data and evaluation services to companies such as OpenAI, Google, and Meta, so Wang Tao knows what each lab is training and where they commonly encounter bottlenecks—but he has no prior experience leading cutting-edge model development.

Wang Tao's most outstanding ability has always been management and resource allocation.

Even after scaling to thousands of employees, he insisted on personally approving new hires and randomly checking data before delivering it to clients. He called this persistence “quality is fractal”: if managers don’t care about details, their teams will soon learn not to care either.

What Zuckerberg values is the focus, decisiveness, and speed of execution that are hard to find in a company with tens of thousands of employees. The Financial Times therefore referred to Wang Tao as the "wartime CEO" that Meta needs.

After joining Meta, Wang Tao quickly introduced this founder-style management approach to the AI department. The establishment of Meta directly impacted the frontier model-focused TBD Lab, poaching top talent from companies like OpenAI and Google, while simultaneously streamlining existing management layers in an effort to transform the bloated AI department into a more agile, “wartime team.”

The cost is obvious.

Meta laid off approximately 600 positions from its original AI team, and Yann LeCun left Meta after 12 years to start fresh with research on world models. A rift has emerged between the high-paid new team and the old research structure.

The cost of this restructuring has now begun to appear in Meta's financial statements.

In the second quarter, Meta's research and development expenses reached $21.656 billion, a 67% year-over-year increase. CFO Susan Li stated on the earnings call that the growth in compensation expenses was primarily due to the increased number of technical staff hired over the past year, particularly high-salary AI talent.

Even after deducting $2.4 billion in legal expenses and $1.18 billion in severance costs, Meta's total costs for the second quarter increased by approximately 42% year-over-year, far outpacing the 28% revenue growth rate.

Wang Tao brought not only a new organizational structure but also an increasingly expensive AI team.

The logic behind Zuckerberg’s bold bet becomes clear here: he believes Wang Tao’s proven ability to rebuild organizations and mobilize resources will ultimately translate into the capacity to train cutting-edge models. The former has already been demonstrated; the latter remains to be proven by the models themselves.

In April 2026, Meta launched the first model from Wang Tao's team, Muse Spark. From the outset, Meta emphasized that it is smaller and faster, serving as the introductory model in the Muse series, with larger models still to come.

By July, Muse Spark 1.1 raised its Artificial Analysis composite score from 43 to 51, with the main improvements concentrated in programming, scientific reasoning, and knowledge capabilities. The pricing is $1.25 per million input tokens and $4.25 per million output tokens.

This model has improved rapidly and is priced low enough, but its overall capabilities still lag significantly behind the state-of-the-art models, making it insufficient to justify Meta’s $14.3 billion investment and the high-cost team it assembled.

The one truly bearing this expectation is Watermelon, who is still in training.

Wang Tao stated at an internal Meta meeting that Watermelon, which is still being trained, uses an order of magnitude more computing power than Muse Spark and has already matched GPT-5.5 on some benchmark evaluations. However, he did not specify the exact evaluation tasks, and Meta has not publicly released related results.

The latest earnings call also provided no additional evidence for this claim. Zuckerberg only confirmed that Meta is training larger, more capable models, without mentioning the name "Watermelon" or disclosing any release timeline or evaluation results.

More realistically, Watermelon hasn't been released yet, but OpenAI, Anthropic, and Google won't wait for it.

Wang Tao has been at Meta for a year and has not yet become a "savior."

Two years late

The most ironic thing is that Meta was never a latecomer to the AI field.

In 2013, Facebook established FAIR, with Yann LeCun as its head. Over the next decade, Meta accumulated deep expertise in computer vision, self-supervised learning, and other areas, and developed PyTorch; Facebook and Instagram’s recommendation and advertising systems have long relied on AI to process billions of data points generated by users daily. The release of Llama in 2023 further rapidly established Meta’s influence in the open-model space.

After ChatGPT's release, OpenAI built its models, APIs, and developer ecosystem around a consumer-facing entry point, while Meta's AI resources remain scattered across FAIR, the generative AI team, product departments, and infrastructure teams.

By 2025, Zuckerberg had elevated AI to a corporate-level priority, planning to invest $60 billion to $65 billion in infrastructure, only to be met with Llama 4’s underwhelming performance.

The centralized reforms subsequently driven by Wang Tao were actually reforms that Meta should have completed in 2023. By the time it began restructuring, OpenAI, Anthropic, and Google had already established advantages in models, talent, and products—leaving Meta two years behind.

Wang Tao is more like Zuckerberg receiving a bill for two years of hesitation.

In early 2025, Zuckerberg plans to invest $60 billion to $65 billion in infrastructure. By 2026, capital expenditures are expected to reach $130 billion to $145 billion, exceeding the previous plan by more than double.

However, Meta did not pour all of this money into a single unreleased flagship model. In the second quarter, Instagram user engagement grew by double digits, and Facebook video viewing time increased by 9%. Meta stated that its new ad model boosted Facebook ad clicks by 8.3% and conversion rates by 15.7%; more than 9 million small businesses are now using AI-generated ad creatives.

These results demonstrate that AI has helped Meta's advertising engine earn more money, but they cannot be directly counted as Wang Tao's achievement.

Meta's disclosed improvements primarily stem from its recommendation system, ad ranking, and content understanding, without specifying the extent to which advanced models from the Watermelon or Wang Tao teams contributed. Advertising and recommendations have long been Meta’s core strengths, honed over more than a decade.

In other words, Meta can continue improving its advertising using existing AI technologies while waiting for Wang Tao’s flagship model. The strong advertising business gives Zuckerberg the confidence to keep waiting, and this high-stakes gamble currently has no set deadline for termination.

For the additional massive computing power, Meta has prepared more uses. Zuckerberg outlined several directions during the call, including personal agents, enterprise agents, model APIs, and direct sale of computing power.

Over one million businesses already use Meta Business Agents each week, and the company has received numerous offers to purchase computing power, with reportedly significantly higher prices than Meta’s cost of acquiring this computing power.

Morgan Stanley analyst Brian Nowak asked which of these businesses is most likely to scale first and deliver measurable returns to investors. Zuckerberg did not select a specific answer, but expressed optimism about all these directions and said more details would be announced soon.

CFO Susan Li put it more directly: Even if Meta’s models don’t reach the cutting edge, the company is confident it can use this computing power to improve its existing products. If there’s excess capacity, it can also sell it directly to other companies.

This provides several exit strategies for a massive investment, but it also indicates that Meta is still uncertain about which new business will ultimately bear the cost.

Can you still afford to wait?

In the second quarter, Meta issued approximately $24.91 billion in long-term debt. As of the end of June, the company’s long-term debt reached $83.664 billion, an increase of $24.92 billion from $58.744 billion at the end of last year. Meanwhile, Meta did not repurchase any shares this quarter.

The reason is not complicated.

In the second quarter, Meta generated $31.862 billion in operating cash flow, but capital expenditures consumed $31.078 billion, leaving only $784 million in free cash flow—insufficient to cover the $1.35 billion in dividends paid during the same period.

Cash generated from advertising business can no longer simultaneously fund AI expansion, dividends, and share repurchases.

Meta still has $90.26 billion in cash and marketable securities on its balance sheet and is not short on funds in the short term. However, it is noteworthy that it is no longer relying solely on cash generated from its advertising business to build AI; instead, it is increasing debt and bringing in partners to raise additional funding for long-term projects.

During the call, Goldman Sachs analyst Eric Sheridan pressed Susan Li on how Meta plans to balance its aggressive investments with its funding needs.

Susan Li responded that the company has been increasing its debt ratio in recent years to secure lower-cost, longer-term funding that aligns with the extended construction timeline of its AI infrastructure.

Meta’s partnership with BlackRock is one example. The day before its earnings report, the two companies announced the construction of a 1-gigawatt data center in El Paso, Texas. Meta shares the construction costs with external partners while retaining the ability to use the computing power.

Money needs to be spent now, but returns won’t come until the data centers are built. Zuckerberg acknowledged on the earnings call that the company’s data centers will not generate value until they go live. Susan Li also stated that Meta’s immediate priority is to scale up its computing power for 2026 and 2027 as much as possible, and decide later, based on actual demand, how many chips to purchase after 2028.

This means that while the capabilities of Watermelon have not yet been publicly validated, the data centers, servers, and networks needed to train larger models must be prepared years in advance. Even if the final results fall short of expectations, land, power, and data center infrastructure are already in place.

Yet it is particularly difficult for Meta to define a clear stop-loss line for this high-stakes gamble.

And don’t forget that Meta has previously demonstrated remarkable patience in sticking with long-term bets. In the second quarter, Reality Labs generated only $431 million in revenue but incurred an operating loss of $4.619 billion. AI glasses drove a 16% increase in revenue for the division, and Zuckerberg noted that sales of the new glasses exceeded expectations—but they are still far from covering Reality Labs’ massive losses.

The external skepticism directed at Wang Tao also lands on Zuckerberg’s shoulders—Wang Tao’s success or failure is no longer just about Meta’s fate, but has become a litmus test for whether Zuckerberg is acting irrationally again after the metaverse.

Wall Street is concerned that AI could become the next high-stakes gamble with no clear deadline.

The longer the wait, the more money and computing power Meta invests. The more they invest, the harder it becomes for Zuckerberg to admit that this path may not work.

Watermelon still has a chance to prove everything for Wang Tao. But before that day arrives, every additional quarter Meta waits will only increase the size of Wang Tao’s “sunk cost.”