Meta's free cash flow turned negative at $2.12 billion in Q3 2026, capital expenditure guidance raised to $135–145 billion

In the AI era, Meta is facing a mandatory question that is hard to answer.

On October 21 local time in the United States, Meta released its financial results for the third quarter of 2026. The company's quarterly revenue grew 26% year over year to $64.601 billion; however, amid large-scale infrastructure investment, free cash flow fell to negative $2.12 billion, compared with positive $10.625 billion in the same period last year.

Meta's main financial data for the third quarter

Company founder and CEO Mark Zuckerberg said the investments are paying off, and Chief Financial Officer Susan Li made it even clearer that the company will continue to increase investment in 2027. Meta also announced in its financial results a new Class C non-voting share plan, seeking to reserve capital and governance space for multi-year investment.

For Meta, aggressive investment has already become an inevitable trend. The key issue now is not only whether the advertising business can continue to grow, but also whether the new investment can generate sufficient revenue and efficiency gains to cover expanding infrastructure costs.

In the third quarter, Meta achieved total revenue of $64.601 billion, up 26% year over year from $51.242 billion in the same period last year. On a non-GAAP basis, excluding the $113 million foreign exchange headwind, revenue growth at constant currency was also 26%.

This growth was mainly driven by the core advertising business, with advertising revenue reaching $63.061 billion in the quarter, up 26% from $50.082 billion in the same period last year.

Other revenue in the Family of Apps was $1.097 billion, bringing total Family of Apps revenue to $64.158 billion. Revenue in the Reality Labs segment, which focuses on the metaverse and hardware, was $443 million.

On the traffic side, daily active people across the Family of Apps reached 3.64 billion, up 3% year over year. Ad impressions increased 10% year over year, and the average price per ad rose 14%, as the platform's advertising monetization capability continued to strengthen.

Compared with revenue growth, profit fell sharply, dragged down by large legal expenses.

In the third quarter, Meta recorded $9.89 billion in settlement expenses related to litigation over adolescent mental health. As a result, consolidated operating income fell to $12.988 billion, down 37% year over year; reported net income was $9.860 billion, and diluted earnings per share were $3.83.

On October 21, the same day the financial results were announced, Judge Francis Matthew of the U.S. District Court for the District of New Mexico ruled in the state's lawsuit accusing Meta of misleading consumers about data privacy, ordering Meta to pay the state as much as $25 billion in damages.

Although the advertising business continues to provide funding, infrastructure investment is changing Meta's cash flow position.

According to officially disclosed data, Meta's operating cash flow in the third quarter was $32.786 billion, higher than $29.999 billion in the same period last year.

On the spending side, cash paid for purchases of property and equipment rose to $33.942 billion in the quarter, compared with $18.829 billion last year. Adding $964 million in principal payments on finance leases, total related cash spending reached $34.906 billion.

Based on Meta's official data, free cash flow in the period fell to negative $2.120 billion. By comparison, free cash flow in the third quarter of 2025 was positive $10.625 billion, and cumulative free cash flow for the first nine months narrowed to $11.049 billion from $29.508 billion in the same period last year.

Negative free cash flow does not mean the company is facing a liquidity crisis. Meta's operating cash flow in the quarter remained at a high level of $32.786 billion, and the negative free cash flow mainly reflects that capital expenditure grew faster than operating cash flow.

As of the end of the third quarter, the company's net property and equipment had increased to $253.222 billion from $176.40 billion at the end of 2025, and long-term debt had increased to $80.922 billion from $58.744 billion.

The key going forward is whether Meta's operating cash flow can continue to grow and whether the new assets can gradually translate into revenue and profit.

For fourth-quarter performance, the company expects revenue of between $73 billion and $76 billion, with a midpoint of $74.5 billion, up about 15.3% from the third quarter.

As resources are reallocated, losses in the Reality Labs segment are expected to be somewhat contained, with the full-year 2026 loss expected to narrow to $18 billion to $19 billion, and further narrowing to around $15 billion in 2027.

On capital expenditure, Meta raised the lower end of its full-year 2026 capital expenditure guidance by $5 billion to $135 billion to $145 billion. Chief Financial Officer Susan Li clearly stated that although the 2027 budget is still being formulated, the company expects to maintain an aggressive investment pace.

As for the sources of return on investment, based on management statements and product plans, they can be divided into three levels.

In the short term, AI's improvement of recommendation algorithms and advertising delivery efficiency is bringing incremental revenue to the core advertising business. In the medium term, the personal AI assistant Muse and the frontier model Muse Wonder launched in the third quarter are gradually testing subscription models and developer interfaces, exploring commercialization directions in enterprise services and transaction scenarios. In the long term, Meta hopes to improve internal operating efficiency and open up new commercialization opportunities through its superintelligence lab and computing infrastructure construction.

To support infrastructure construction and long-term strategic planning, Meta disclosed in its financial report a plan to adjust its corporate governance and capital structure; the board has approved the creation of a new class of non-voting Class C common stock.

Under the disclosed plan, the company intends to distribute a stock dividend to existing Class A and Class B shareholders in the form of one Class C share for each share held.

The company said the distribution of Class C shares will not change the existing shareholder voting structure. The company's special committee has signed a related agreement with Zuckerberg aimed at continuing the existing governance arrangements.

In 2014, Google parent Alphabet completed a stock split by proportionally distributing non-voting Class C shares, enabling founders Larry Page and Sergey Brin to successfully lock in control of the company while continuing to issue shares for employee equity incentives and acquisitions.

The non-voting Class C shares will mainly be used for Meta's future employee equity incentives, potential strategic acquisitions or other capital operations. The stock dividend itself will not directly bring new cash flow to the company. This arrangement reflects Meta's desire to emulate the mature capital tools of other tech giants while maintaining its existing control structure, reserving room for future equity incentives and capital operations.

Meta's third-quarter financial report presents two curves advancing simultaneously: the advertising business continues to generate revenue, while AI infrastructure continues to expand capital requirements. Advertising revenue growth provides the funding base, but it does not mean AI investment will necessarily yield corresponding returns.

Meta has clearly stated that it will continue to increase investment in 2027. For Zuckerberg, adhering to the long-term strategy is only the first step; as investment continues to expand, the company needs to prove that the new revenue and efficiency gains brought by AI are sufficient to support this multi-year infrastructure expansion.

The following is a condensed transcript of Meta's third-quarter earnings analyst call:

Zuckerberg first commented:

In the third quarter, 3.64 billion people used our family of apps every day, and engagement on both Facebook and Instagram reached all-time highs. Threads monthly active users have surpassed X, and WhatsApp message volume also hit a new high.

We are rapidly advancing our goal of "providing personal superintelligence to everyone," and the launch of the Muse personal agent has been very successful. In just a month and a half since launching in the US and Canada, it has surpassed 10 million setup users, with about 25% of them using it every day. The most common use case is shopping (over 25% of usage, with an average spend of over $200).

Because first-week transaction volume far exceeded expectations, we had to work with Stripe to temporarily raise credit limits. Next, Muse will launch on Meta smart glasses starting next month.

On model development, we released our first closed-source frontier model, Muse Wonder, which is at the industry's most advanced level on multiple key metrics; we also released Muse Spark 1.2/1.3, Muse Code, and two open-source open-weight models, Muse Glimmer and Muse Spark Open. Going forward, we will continue to push the frontier of intelligence and plan to distill frontier models into smaller, more efficient product models.

On compute and efficiency, we are building compute capacity capable of serving more than 1 billion people in the coming years. Although Muse is currently a compute-intensive application, we expect cost efficiency to improve more than 10-fold next year, which will enable it to be profitable at scale. We expect that a year from now we will be able to deliver equivalent intelligence in smaller models at a fraction of the current cost of Muse Wonder.

On organizational adjustments, we will change the release cadence of VR devices and instead prioritize delivering personal superintelligence. We expect Reality Labs' operating loss to narrow to about $15 billion next year and to move toward profitability from there.

Chief Financial Officer Susan Li then added information on capital expenditure and capital allocation. She noted that Meta's capital expenditure will continue to increase in the coming years, mainly to support the core business and emerging AI initiatives. Although AI investments take longer to generate returns than the core business, they are already bringing substantial incremental revenue to core advertising by improving recommendation and monetization efficiency.

The following is the analyst Q&A session:

Q: What are the user growth constraints for Muse as it moves forward? What is Meta's vision for its enterprise business?

Zuckerberg: Muse's biggest constraint right now is mainly compute intensity, so it has not yet been broadly rolled out outside the US and Canada; we are working to improve compute efficiency and build enough compute capacity to support a scale of more than 1 billion users in the future.

Q: What friction currently exists in agentic commerce? How is the company addressing it? Also, what are the considerations behind the new share class announced tonight in terms of capital structure and future AI investment?

Zuckerberg: The vast majority of businesses are very excited to support this, because they see that it can not only increase sales and improve the existing customer experience, but may also bring in new customers. Adoption speed will vary across businesses, and we will work on terms that are reasonable for all parties. In the long run, making this work well is valuable for everyone, and we are currently on a good trajectory.

Susan Li: Agents will become an important part of how reality evolves, and the market will eventually shift toward more permissioned, structured interactions—including authenticated APIs, approved agent partnerships, or direct communication between consumers' personal agents and enterprise agents. We are working with partners such as Sierra, Walmart, Stripe, and Shopify on the Personal Agent Protocol to define secure and reliable ways to interact. Many experiences in the near term will still be AI-assisted rather than fully autonomous, but as trusted integrations develop, more businesses will actively participate.

Regarding the new share class, we created this new class of non-voting capital stock (Class C shares) for the purpose of distributing it to all existing shareholders through a stock dividend, creating publicly tradable non-voting stock that can be used for general corporate purposes such as financing, employee compensation, and acquisitions, while preserving the voting rights of existing Class A and Class B shareholders.

Over the long term, as LLMs scale, we expect to be able to accomplish more tasks with fewer tokens and to rely on the AI models themselves to help improve the efficiency of the software used to run the models. There is still a great deal of work on the efficiency roadmap that can help us bend the cost curve, and the arc of AI technological progress will also bring greater savings.

Q: Does Muse's commercialization pace need to be faster than for past new products, because the associated costs are clearly higher than in the past? Also, does the data on 25% of users engaging in shopping and an average of over $200 include travel purchases?

Zuckerberg: We will start monetization earlier than the normal playbook, mainly for two reasons: first, both the costs and the scale of the opportunity are large, and we must make progress in order to scale; second, the fact that people use it heavily is itself a business. That data does indeed include travel. We will establish a model to capture a certain percentage of the value we help people create. Making the math work is very important so that we can continue to build computing capacity and reach the scale of serving everyone.

What we are seeing is a trend that, year over year, a year from now we will be able to deliver the same intelligence in smaller models at a fraction of the cost of the current Muse Wonder (or any current frontier model) level of intelligence. If this represents a baseline level sufficient to support product-market fit, continued model development will drive efficiency gains.

Of course, a year from now most people will not be satisfied with the current level of intelligence (because intelligence itself is also advancing), so we need to advance revenue, scaling, model development, and efficiency simultaneously in order to fund the buildout of computing and reach billions of people as quickly as possible. We are not charging everyone from the start because we want to first build a large-scale Muse network.

Zuckerberg: We released our first closed-source frontier model, Muse Wonder, which is state of the art across multiple key dimensions of personal superintelligence, validating the importance of training our own models to deliver the best experience. During the same period, we also released Muse Spark 1.2 and 1.3, Muse Code, and two open-source open-weight models, Muse Glimmer and Muse Spark Open, as well as Muse Voice.

Looking ahead, as the field advances rapidly, we plan to both push the intelligence frontier and distill these models into smaller, more efficient product models optimized for personal agents. We are already working on larger models after Muse Wonder, which will be very impressive and capable; we are also developing new smaller models to launch soon that will deliver Muse Wonder-level intelligence in a more efficient way.

My view is that we will be able to build superintelligence through scaling, so we are methodically climbing this ladder.

Q: Is compute currently the main constraint on scaling Muse? How is the company improving computational efficiency? What is the basis for the expectation of more than a 10x improvement in cost efficiency next year?

Zuckerberg: Yes, one of the biggest constraints right now is compute intensity, which is why we have not yet launched globally (currently only in the US and Canada). Part of the reason is that we want to optimize and refine connectors for each location, and more importantly, ensure efficient progress. We are building computing capacity that can serve more than 1 billion people in the coming years.

In the near term, Muse is very compute-intensive to run, but we expect its cost efficiency to improve by more than 10x next year, which will make it profitable once revenue reaches scale. If computational efficiency continues to improve rapidly over the next few years (which we consider highly likely), then Muse will be highly profitable for many years.

Q: On capital expenditures, the 2026 guidance has been raised to $135–145 billion. How will capital expenditures evolve in the coming years? How do these investments support the core business and AI?

Susan Li: Third-quarter capital expenditures (including principal payments on finance leases) were $34.9 billion, driven by investments in servers, data centers, and network infrastructure.

We now expect full-year 2026 capital expenditures in the range of $135 billion to $145 billion (previously $130–145 billion).

Looking beyond 2026, we are still finalizing next year's budget and capacity plans, but we expect to continue investing aggressively to support the core business and emerging AI initiatives. Our current capacity and planned capacity give us the flexibility to allocate resources among model training, core business growth, and new product development and support.

Our AI investments have already generated substantial incremental revenue for the core advertising business through stronger engagement, better recommendations, and more efficient monetization. Although these investments take longer to generate returns than the core business, the breadth of the monetization opportunity and internal efficiency give us confidence in the long-term potential. We expect to maintain strong overall revenue growth throughout 2027.

This article is from the WeChat public account "Tencent Technology," author: Su Yang, editor: Xu Qingyang, published by 36Kr with authorization.