Meta's $70 Billion Capex Problem: When AI Spending Outgrows Free Cash Flow

Meta's capital allocation history has three distinct chapters before the current one. The first ran through 2021, when the business was structurally simple: an advertising platform with exceptional margins, minimal capex needs relative to revenue, and buybacks that quietly compressed the share count. Free cash flow in 2021 was $39.1 billion on $117.9 billion in revenue. The machine worked cleanly.

The second chapter was 2022, when Zuckerberg's metaverse ambition coincided with a digital advertising recession. Capex surged to $31.4 billion, revenue stalled at $116.6 billion, operating margins collapsed from 39.6% to 24.8%, and net income halved to $23.2 billion. The stock fell roughly 65% that year. Investors had legitimate questions about whether the company had lost its financial discipline.

The third chapter started in 2023 with what Meta called the Year of Efficiency. Headcount was cut, operating costs were restructured, and the business snapped back hard. Revenue grew to $134.9 billion in 2023, margins recovered to 34.7%, and free cash flow more than doubled to $43.8 billion. By 2024, operating margins had reached 42.2%, a level most software businesses never achieve.

Now a fourth chapter is underway. The efficiency discipline that restored investor confidence is being deployed in service of a new ambition: AI infrastructure at a scale that is reshaping Meta's cost structure from the ground up.

In 2025, Meta generated $201.0 billion in revenue, up 22.3% from $164.5 billion in 2024. Operating income reached $83.3 billion at a 41.4% margin, essentially flat with the 42.2% margin in 2024 despite the near-doubling of capital expenditure. That margin stability in the face of an $32 billion increase in annual capex is genuinely notable.

Operating cash flow was $115.8 billion, up 26.9% year-over-year. That number matters because it shows the core advertising business is generating cash at an accelerating rate. The structural problem, from a free cash flow perspective, is that $69.7 billion of those operating dollars went straight back into infrastructure.

To put the capex trajectory in context: Meta spent $18.6 billion on capital expenditures in 2021, $31.4 billion in 2022, $27.3 billion in 2023, and $37.3 billion in 2024. The 2025 figure of $69.7 billion represents an 87% increase in a single year. Management's guidance for 2026 capex ranges from $60 billion at the optimistic end to higher depending on how the AI buildout accelerates.

Gross margins held at 82.0% in 2025, up slightly from 81.7% the year prior. The AI spending is flowing through the balance sheet in a way that has not yet materially compressed the profitability of the advertising business itself. Whether that holds across the next two years depends on how the depreciation burden from 2025's spending begins to flow through.

Meta is not spending $70 billion on a product in the traditional sense. It is building infrastructure: data centers, custom silicon through its MTIA chip program, fiber connectivity, and the compute base that underlies both its recommendation algorithms and its generative AI products. The Corning fiber expansion announced in early April 2026 is one visible piece of the supply chain Meta is constructing to support this buildout. These are not software investments with short depreciation cycles. They are long-duration physical assets.

The direct monetization argument is that AI recommendation improvements drive feed engagement, which drives time-on-platform, which drives advertising CPMs. Meta has been testing this loop for years, and the correlation between increased AI investment and accelerating revenue is clear: $134.9 billion in 2023, $164.5 billion in 2024, $201.0 billion in 2025. Causation is hard to isolate cleanly, but the data does not argue against it.

The harder question is the AI assistant and standalone product layer. Meta AI, deployed across WhatsApp, Messenger, Instagram, and Facebook, reaches a user base that dwarfs any competitor's AI product by orders of magnitude. Whether that scale eventually translates into a revenue line independent of advertising is genuinely unresolved. It is where the long-duration value case lives, and where the uncertainty is highest.

The market is currently pricing this uncertainty conservatively, which is reflected in the word weights from recent coverage: "ai" and "platform" dominate the conversation about META, with "invest" and "billion" close behind. The market knows the bet is large. It is still deciding whether the return is there.

Meta reinstated shareholder returns in 2024 alongside the AI capex ramp, a deliberate choice that signaled management believed it could fund both the infrastructure bet and ongoing capital return simultaneously.

In 2024, the company returned $30.1 billion in buybacks and $5.1 billion in dividends for a total of $35.2 billion. In 2025, capital returns fell to $26.2 billion in buybacks and $5.3 billion in dividends for $31.5 billion total. The buyback reduction is directly attributable to the capex increase absorbing a larger share of operating cash flow. Cash on the balance sheet ended 2025 at $35.9 billion, with net debt of approximately $22.8 billion.

The share count compression over the past five years is material and often underweighted. From 2.89 billion shares outstanding at the end of 2020 to 2.61 billion by the end of 2024, Meta retired roughly 280 million shares, approximately 9.7% of the outstanding float. That reduction amplifies per-share earnings growth beyond what headline revenue numbers suggest.

Going forward, the capital return program will likely remain compressed relative to peak buyback years until the capex cycle moderates. Management has characterized 2025 and 2026 as peak infrastructure intensity. If that holds, free cash flow should recover toward $55-60 billion over the following two years, and the buyback program can reaccelerate. If capex continues to outpace cash generation, the compression becomes a permanent feature, not a cyclical one.

Meta operates four of the world's largest social platforms, with a combined daily active user base exceeding three billion people. That scale creates an AI training and deployment advantage that is structurally different from what any standalone AI company can replicate.

AI recommendation systems improve with data. Meta has more behavioral data on human attention, engagement, and social connection than any other private entity on earth. The more it invests in AI infrastructure, the better its recommendations become, and the harder it is for advertising-dependent competitors to match its targeting precision. This is a reinforcing loop, not a simple infrastructure bet. The advertising business itself sits atop this moat: at 82% gross margins and 41% operating margins on $201 billion in revenue, Meta's unit economics have no peer in traditional media.

The reputational headwinds noted in recent coverage are real but analytically distinct from the competitive position. Brand perception among consumers and brand perception among advertisers are different variables. Advertisers follow measurable return on investment. As long as Meta's ad platform delivers that, the spending will continue regardless of editorial coverage or public sentiment cycles.

The competitive pressure from AI commoditization is more relevant. If DeepSeek-style low-cost architectures close the intelligence gap that Meta's scale-driven spending is trying to widen, the capex justification weakens. This is the most credible long-run threat to the investment thesis, and it is one that cannot be resolved with financial analysis alone.

At 24.6x trailing earnings on $60.5 billion in net income, Meta is priced cheaper than the historical median multiple for the S&P 500 on a profits basis. An EV/EBITDA of 13.7x on $104.5 billion in EBITDA for a business growing revenue at 22% annually is, in any rational comparison, not an expensive number.

The price-to-sales multiple of 7.3x reinforces the same picture. This is not a speculative growth stock trading on optionality. It is one of the most profitable businesses in existence, trading at multiples that historically applied to slow-growing industrials. Sixty of 68 analysts rate it a buy or strong buy, with a consensus price target of $861.76 implying roughly 29% upside from current levels near $669.

The discount embedded in the current multiple is primarily the capex overhang. Investors are applying a haircut for the uncertainty around whether $70 billion in annual infrastructure investment generates sufficient return, and for the free cash flow compression that investment produces. The 30-day sentiment trend adds nuance here: normalized sentiment dropped from 0.77 in mid-March to 0.36 in the March 25-27 window during a period of heavy news volume, reflecting real investor anxiety about the spending trajectory. The partial recovery in early April suggests that anxiety is easing but not resolved.

If the AI infrastructure cycle peaks in 2025-2026 and FCF recovers toward $55-60 billion, the current multiple looks attractive in hindsight. That is a meaningful conditional, but it is not an unreasonable one given the business quality underneath.