Meta spent roughly $65–72B on capital expenditure in 2025 — nearly double the ~$39B it spent in 2024 — and the overwhelming majority of it went to AI infrastructure: Nvidia GPUs, custom MTIA silicon, and gigawatt-scale data centers.
That's the short answer. The longer answer is more interesting — because Meta isn't a chip company, a cloud provider, or an AI lab that sells tokens. It's an ad business spending the GDP of a small country on compute it mostly uses internally. Here's where the $65B+ actually goes, why Zuckerberg can self-fund a build that would bankrupt anyone else, and how the bet stacks up against Amazon, Microsoft, and Google.
Meta's AI Capex in 2025: The Number and What It Buys
Meta's AI capex in 2025 reached roughly $65–72B, up from about $39B in 2024 — the steepest single-year jump in the company's history. The bulk funds AI infrastructure: hundreds of thousands of Nvidia GPUs, Meta's custom MTIA accelerators, and multi-gigawatt data centers in Louisiana and Ohio built to train Llama and power AI features across Facebook, Instagram, and WhatsApp.
The number moved twice during the year. In January 2025, on the Q4 2024 earnings call, management guided full-year capex to $60–65B. By summer, as demand for AI compute kept climbing and new data-center projects were announced, the range was lifted to roughly $66–72B. Either way, the trajectory is the story: Meta's capital intensity has gone vertical. Here is the multi-year arc.
| Year | Capex | YoY Change | Context |
|---|---|---|---|
| 2021 | ~$19B | — | Pre-AI build, mostly servers + offices |
| 2022 | ~$31B | +63% | Metaverse peak spending |
| 2023 | ~$28B | −10% | Efficiency year, capex pullback |
| 2024 | ~$39B | +39% | AI build begins in earnest |
| 2025 | $65–72B | +67–85% | Full AI infrastructure ramp |
| 2026E | $100B+ | +40%+ | Guidance signals further acceleration |
Figures are blended from Meta 10-K and 10-Q filings, quarterly earnings-call capex guidance, and consensus analyst estimates for 2026. The 2025 figure reflects the raised $66–72B range; 2026 is a forward estimate and not yet reported.
Where Meta's 2025 AI Capex Actually Goes
"AI infrastructure" is a vague phrase, so let's be specific. Roughly three buckets absorb the spend. First, AI servers and accelerators — the GPUs themselves. Meta has been one of Nvidia's largest customers, buying H100, H200, and GB200 systems by the hundreds of thousands; Zuckerberg has publicly targeted roughly 1.3 million GPU-equivalents of compute and the equivalent of ~600,000 H100s of training capacity. At list prices north of $25,000–40,000 per high-end GPU before networking, that single line item runs into the tens of billions.
Second, data centers and power. In 2025 Meta announced two flagship gigawatt-class campuses: Prometheus in Ohio, targeted to come online around 2026 at roughly 1GW, and Hyperion in Louisiana, a multi-year build designed to scale toward 5GW of capacity. For scale, 1GW is enough to power roughly 750,000 homes — Meta is building several of them just to train and serve models. Third, custom silicon: Meta's in-house MTIA (Meta Training and Inference Accelerator) chips, designed to handle ranking and recommendation inference more cheaply than buying Nvidia for every workload.
| Spend Bucket | Est. Share | What It Covers |
|---|---|---|
| AI servers & GPUs | ~50–55% | Nvidia H100/H200/GB200 + networking |
| Data center construction | ~25–30% | Prometheus, Hyperion, campus expansions |
| Power & energy | ~8–10% | Substations, grid deals, on-site generation |
| Custom silicon (MTIA) | ~3–5% | In-house inference accelerators |
| Network & storage | ~3–5% | Backbone, interconnect, data storage |
| Non-AI / facilities | ~3–5% | Offices, legacy infrastructure, Reality Labs hardware |
Allocation shares are directional 2025 estimates derived from Meta earnings commentary, supply-chain reporting on Nvidia GPU shipments, and announced data-center projects. Meta does not publish a precise line-item breakdown of capex, so these are informed approximations, not reported figures.
Meta AI Capex 2025 vs Big Tech: The Spending Race in Context
Meta's $65–72B looks enormous until you line it up against the other hyperscalers. In absolute dollars, Meta actually spends the least of the big four — but as a percentage of revenue it's among the most aggressive. Here's the 2025 capex picture across the companies building out AI at scale.
| Company | 2025 Capex | ~Capex / Revenue | Primary AI Focus |
|---|---|---|---|
| Amazon (AWS) | ~$100–105B | ~16% | Cloud capacity, Trainium, Anthropic |
| Microsoft | ~$80–85B | ~30% | Azure, OpenAI compute, Copilot |
| Alphabet (Google) | ~$80–85B | ~22% | TPUs, Gemini, Google Cloud |
| Meta | $65–72B | ~40% | Llama, ranking, Meta AI, MTIA |
| Oracle | ~$25–30B | ~45% | OCI data centers, AI hosting |
| Apple | ~$12–14B | ~3% | On-device AI, modest data centers |
Figures are 2025 estimates blended from company 10-K/10-Q filings, earnings-call guidance, and analyst consensus (Morgan Stanley, Bloomberg). Capex/revenue ratios use trailing-twelve-month revenue and are rounded; Oracle and Apple are included for range, not as direct hyperscaler peers.
The combined 2025 capex of the four largest US tech companies clears $330B — a number that would have sounded absurd three years ago. What makes Meta's position distinct is the ~40% capex-to-revenue ratio: unlike Amazon or Microsoft, Meta isn't reselling this compute to external customers. It's spending almost half its revenue on infrastructure it uses itself. You can track how these numbers move quarter to quarter on our Big Tech Earnings dashboard.
Llama, Superintelligence, and the Return on $65B
The obvious question for any LP or public-markets investor: what does Meta get for $65B+ a year? The honest answer is that most of it is defensive and internal. AI compute powers the ranking and recommendation systems behind Facebook and Instagram feeds, Reels, and ad targeting — and even single-digit-percent improvements in engagement and ad efficiency are worth billions against a ~$164B revenue base. Zuckerberg has repeatedly framed AI capex as the highest-ROI spend available because it compounds across every surface.
The offensive bets are Llama and the "superintelligence" push. Meta released Llama 4 (Scout and Maverick, with a larger Behemoth model in training) in April 2025, continuing its open-weight strategy that has driven hundreds of millions of model downloads and made Llama the default base model for much of the open-source AI ecosystem. In mid-2025, Meta also took a roughly $14.3B stake in Scale AI — about 49% — and stood up Meta Superintelligence Labs, hiring aggressively with reported nine-figure compensation packages for top researchers. That talent and data spend sits alongside the capex, not inside it.
The strategic read: Meta is buying optionality. If frontier AI becomes the platform layer of the next decade, Meta wants to own its own compute, its own models, and its own silicon rather than rent them from a rival. You can see how the market prices these AI bets on our AI Valuations dashboard, and where the broader infrastructure dollars are flowing on the AI Spending tracker.
Reality Labs: The Other Hole Behind the Capex Story
It's worth separating two things people conflate. The AI capex build is not the metaverse. Reality Labs — Meta's AR/VR and metaverse division — posted an operating loss of about $17.7B in 2024 and has cumulatively lost well over $60B since 2020. But that's an operating-expense story that flows through the income statement, distinct from the capital expenditure funding AI data centers.
The reason Meta can run both a $60B+ cumulative metaverse loss and a $65B+ annual AI build at the same time is the cash machine underneath: the Family of Apps segment generated tens of billions in operating income in 2024, and Meta posted roughly $62B in net income on ~$164B of revenue. That profitability is precisely why Meta can self-fund infrastructure that capital-constrained AI labs have to raise tens of billions in debt and equity to match — a structural advantage I'd weigh heavily if I were underwriting who actually wins the compute race.
The Bottom Line
Meta's $65–72B of 2025 capex is a self-funded bet that owning compute, models, and silicon beats renting them — paid for by an ad business throwing off ~$62B in net income.
Nearly doubling capex in a single year would terrify most boards. Meta can do it because its core business is wildly profitable and the spend is mostly defensive — protecting a ~$164B revenue engine that AI directly improves. The risk isn't bankruptcy; it's ROI discipline. If Llama, Meta AI, and superintelligence don't eventually justify gigawatts of GPUs, this becomes a second Reality Labs. But unlike the metaverse, AI capex pays for itself today through ranking and ads. That's why, of all the hyperscaler bets, Meta's is the one I'd least worry about — and the 2026 guidance north of $100B says Zuckerberg agrees.
Track big-tech capex, AI valuations, and quarterly earnings on the Big Tech Earnings, AI Valuations, and AI Spending dashboards at Value Add VC. Originally published in the Trace Cohen newsletter.