AI & TechnologyMay 29, 2026·9 min read·Last updated: May 29, 2026

Meta AI Capex 2025: $65B Raised to $64–72B — What Zuckerberg Is Building and Why It's Different

Meta's 2025 AI infrastructure spend makes it the third-largest capex spender in tech — behind only Microsoft and Amazon. But unlike its peers, Meta isn't selling the compute. It's consuming every watt internally across 3.27 billion daily active users.

TC
Trace Cohen
3x founder, 65+ investments, building Value Add VC

Quick Answer

Meta's 2025 AI capex is $64–72B (raised guidance from an initial $60–65B), up from $38.4B in 2024 — a 67% increase. The spending targets NVIDIA GPU clusters, US hyperscale data centers, and compute for training Llama 4 and Llama 5. Unlike Microsoft and Google, Meta earns no cloud revenue from this infrastructure — the entire ROI case depends on ad targeting improvements, engagement gains, and Meta AI reaching 1B+ monthly active users.

Meta's 2025 AI capex started at $60–65B and got raised to $64–72B before Q1 was over. That's not a rounding error — it's Zuckerberg telling the market he's not done.

For context: Meta spent $38.4B on capex in all of 2024. The 2025 raise represents a ~67% jump in a single year. And unlike Microsoft or Google, Meta earns zero cloud revenue to offset it. Every dollar has to come back through ads, engagement, or consumer AI — which makes this one of the highest-stakes unilateral infrastructure bets in corporate history.

Meta AI Capex 2025 vs. Prior Years

YearTotal CapexYoY ChangePrimary Driver
2022$32.0B+20%Data centers, VR (Reality Labs)
2023$28.1B−12%Efficiency year — headcount cuts
2024$38.4B+37%AI infrastructure ramp begins
2025E$64–72B+67–88%GPU clusters, Llama training, Meta AI serving

Where the $65B+ Is Actually Going

Meta's capex has four primary buckets. The weighting has shifted dramatically toward AI compute since 2023:

~55%
GPU Clusters & AI Compute
NVIDIA H100, H200, and B200 procurement; custom MTIA inference chips; 350,000+ H100-equivalent GPU cluster target by end of 2025
~30%
Data Center Construction
New hyperscale campuses in Louisiana, Wyoming, and abroad; power infrastructure upgrades for AI workloads; $10B+ committed in US data center real estate
~10%
Networking & Storage
InfiniBand and Ethernet fabric for GPU-to-GPU communication; petabyte-scale storage for training datasets and model checkpoints
~5%
Reality Labs / AR Infrastructure
AI for Ray-Ban smart glasses, Quest headsets, and future AR hardware; down from prior years as AI compute dominates

Why Meta's AI Strategy Is Structurally Different

Microsoft and Google sell cloud compute. That creates a natural feedback loop: more infrastructure spending → more cloud revenue → justifies more spending. Meta has no equivalent. Its capex logic is entirely internal:

Distribution moat, not cloud revenue

3.27B daily active users across Facebook, Instagram, WhatsApp, and Messenger. Every AI improvement ships to more people than any cloud provider serves.

Ad ROI compounds quietly

Meta's 2024 revenue was $164.5B — 91% from ads. A 1% improvement in targeting efficiency adds ~$1.6B. AI-driven feed ranking and ad relevance already drove a 7% engagement increase in 2024.

Llama as competitive moat via commoditization

By open-sourcing Llama 3 and Llama 4, Meta forces every competitor to spend billions training models anyone can download. The real moat is Meta's proprietary social graph and behavioral data — which no open-source release transfers.

Self-funded at scale

Meta generated $62.4B in net income in 2024. It can fund $65–72B in capex from a single year of profits without diluting equity or raising debt — a position most AI companies can only dream of.

Meta AI: The Consumer Product That Has to Justify This

Meta AI — the assistant embedded across all Meta apps — is the single most important vehicle for converting GPU spend into consumer value. Zuckerberg set a target of 1 billion monthly active users for Meta AI in 2025. By Q1 2025, he claimed 700M+ MAUs, making it larger than ChatGPT by user count even if usage depth differs significantly.

The business model here isn't a subscription — it's engagement. Every interaction with Meta AI keeps users inside the Meta app ecosystem longer, generating more ad impressions. The AI assistant doesn't need to monetize directly if it reduces churn and increases session time.

Meta also deployed AI across its core ad products: generative ad creative, automated audience targeting, and AI-based bidding optimization. These are already measurable in revenue. The incremental gains from Llama 4 powering ranking models may not show up in a press release but they show up in the $164B top line. Track the ad revenue per user trend on the Big Tech Earnings Dashboard.

The Risks Nobody Is Pricing

Bear Case

  • ✕ AI assistants commoditize; Meta AI has no differentiated reason to use vs. ChatGPT or Gemini
  • ✕ Ad spend shifts to AI-native platforms that generate higher-intent clicks
  • ✕ Regulatory risk: EU DMA, FTC scrutiny on data use for AI training
  • ✕ Reality Labs has lost $60B+ over five years with no clear path to profitability

Bull Case

  • ✓ Distribution advantage is real: 3B+ daily users means even modest AI improvements have massive aggregate value
  • ✓ Llama open-source strategy establishes Meta as infrastructure-layer — not just an app
  • ✓ $62B/year in net income means Meta can sustain this spending for 5+ years without financial strain
  • ✓ Social graph + behavioral data is genuinely unique training signal no competitor can replicate

Meta's bet isn't that AI assistants win the market.

It's that AI makes 3 billion people more engaged — and Zuckerberg can afford to find out if he's right.

Track Meta and big tech AI spending on the Big Tech Earnings Dashboard and AI Valuations at Value Add VC. Originally published in the Trace Cohen newsletter.

Frequently Asked Questions

What is Meta's AI capex in 2025?

Meta initially guided $60–65B in 2025 capex in February 2025, then raised it to $64–72B after Q1 2025 earnings. This compares to $38.4B in 2024 — a roughly 67% year-over-year increase. The majority funds AI infrastructure: GPU clusters, hyperscale data centers, and networking.

What is Meta spending its AI capex on?

Meta's 2025 AI capex goes primarily to NVIDIA H100 and H200 GPU clusters for training and inference, new US data center campuses (Louisiana, Wyoming, and others), networking infrastructure, and the compute needed to serve Meta AI across 3.27 billion daily active users on Facebook, Instagram, WhatsApp, and Messenger.

How does Meta's AI capex strategy differ from Microsoft and Google?

Microsoft and Google monetize AI infrastructure externally through Azure and Google Cloud. Meta does not sell cloud compute — every dollar of infrastructure investment must generate ROI through improved ad targeting, user engagement, or consumer AI adoption. This makes Meta's capex math more concentrated and higher-risk, but also more direct.

Is Meta's AI capex generating a return?

Meta's 2024 revenue was $164.5B and net income was $62.4B — it can fund this spending from earnings. The measurable returns so far: AI-driven recommendation systems contributed to a 7% increase in time-on-platform in 2024, and Meta AI reached 700M+ monthly active users by mid-2025. The full ROI picture will take 2–3 years to validate.

How does Meta's Llama open-source strategy affect its capex bet?

Meta releases Llama models publicly, which seems counterintuitive for a company spending $65B+ on AI compute. The strategy is deliberate: commoditizing frontier AI models forces every competitor to spend at the same rate, while Meta's real moat is distribution across 3B+ users and proprietary social graph data that no open-source release can replicate.

Related Tools & Dashboards

Explore 45+ free VC tools, dashboards, and recommended startup software.