Meta spent $75B in 3 months on AI infrastructure (CoreWeave, Oracle, Blue Owl)

10 hours ago 2

I watched Meta’s capital expenditure ratio hit 37% of revenue in Q3 2025.

That’s not a typo. Meta is now spending more than a third of every dollar it makes on AI infrastructure. For context, that’s nearly double what they spent last year (20%), and it’s the highest capex-to-revenue ratio in the company’s history.

But here’s what really caught my attention: in just three months (September to October 2025), Meta announced $75.5 billion in infrastructure deals. That’s more than most countries spend on their entire tech sector in a decade.

The $75 Billion Question Nobody’s Asking

When Mark Zuckerberg talks about building “superintelligent” AI systems, most people focus on the models. But the real story is in the infrastructure-and the unprecedented way Meta is financing it.

Between September 30 and October 27, 2025, Meta signed four massive deals:

CoreWeave: $14.2 billion (6+ years)
Oracle: ~$20 billion (multi-year)
Scale AI: $14.3 billion (49% stake)
Blue Owl/Hyperion: $27 billion (joint venture)

Total: $75.5 billion

For perspective, that’s:

More than Netflix’s entire market cap ($280B) × 0.27
Equivalent to building 15 nuclear power plants
Enough to buy every data center in Ireland twice

But what’s really interesting isn’t the size-it’s the structure.

Here’s where it gets fascinating.

Traditional tech infrastructure spending works like this: Company makes money → Company spends cash on servers → Company owns infrastructure.

Meta just rewrote that playbook.

The Hyperion Deal: A New Financial Model

On October 21, 2025, Meta announced a joint venture with Blue Owl Capital for the Hyperion data center in Louisiana. The structure is unlike anything I’ve seen in tech:

Ownership Split:

Blue Owl: 80%
Meta: 20%

Financing Structure:

Morgan Stanley arranged $27B+ debt
$2.5B equity into Special Purpose Vehicle (SPV)
PIMCO as anchor lender (144A bonds, maturing 2049)
Meta received $3B cash distribution upfront

The Kicker: Meta doesn’t own most of it, but they’re on the hook for 16 years through a residual value guarantee.

Think about that: Meta is essentially leasing infrastructure they’re building, financed by private credit, with a 16-year financial commitment.

Why This Matters

This isn’t just creative accounting-it’s a fundamental shift in how tech infrastructure gets built.

Traditional Model:

Revenue → CapEx → Owned Assets → Depreciation

New Model:

Revenue → JV Partnership → Leased Assets → Operating Expense

The implications:

Balance sheet stays cleaner (assets off-balance-sheet)
Faster deployment (external capital accelerates build)
Shared risk (partners absorb some downside)
Higher leverage (debt financing amplifies scale)

But there’s a catch: if AI doesn’t deliver ROI, Meta is still paying rent for 16 years.

The Technical Specs: What $75B Actually Buys

Let me break down what Meta is actually getting for this money.

Hyperion Data Center (Louisiana)

Scale:

Size: 2,250 acres (1,700 football fields)
Power: 2 gigawatts (2,000 megawatts)
Completion: 2030
Location: Richland Parish, Louisiana (between Rayville and Delhi)

For Context:

2GW is enough to power 1.5 million homes
That’s more power than some small countries use
Requires dedicated power infrastructure (Entergy building $1.2B transmission line)

What It Can Do: Train models the size of GPT-4 multiple times simultaneously. We’re talking about infrastructure that can handle:

Hundreds of thousands of GPUs
Petabytes of training data
Months-long training runs without interruption

CoreWeave Infrastructure

Hardware:

Nvidia GB300 server racks
72 Blackwell GPUs per rack
Access through December 2031 (optional extension to 2032)

Why GB300 Matters: The Blackwell architecture (GB300) is Nvidia’s latest generation. Each GPU delivers:

2.5× performance vs previous gen (H100)
Better power efficiency (critical for multi-GW facilities)
Native support for FP4 precision (faster inference)

Real-World Impact: According to CoreWeave’s SEC filing (September 30, 2025), this infrastructure can reduce training time for large models from months to weeks.

Oracle Cloud Computing

Confirmed Details:

$20 billion multi-year contract
Part of $65 billion in Oracle Cloud Infrastructure (OCI) bookings in single 30-day period
Announced October 16, 2025

Oracle’s Projections:

FY2030: $166 billion cloud infrastructure revenue
Cloud gross margins: 35% target
AI database revenue: $20B by FY2030

What Meta Gets: Flexible cloud capacity for:

Inference workloads (serving AI models to users)
Distributed training (across multiple data centers)
Backup and redundancy (if owned infrastructure fails)

Energy Infrastructure: The Hidden Cost

On October 27, 2025, Meta signed a deal with ENGIE for 1.3 GW of solar power across four Texas projects.

Key Project: Swenson Ranch Solar

Capacity: 600 MW
Location: Stonewall County, Texas
Meta’s Commitment: 100% of output
Operational: 2027
Significance: ENGIE’s largest solar project (11 GW total portfolio)

Why This Matters: AI training is energy-intensive. A single training run for a large language model can consume:

1,287 MWh (equivalent to 120 US homes for a year)
Cost: $100,000+ in electricity alone

Source: Stanford HAI Report, 2024

With 1.3 GW of dedicated solar, Meta can:

Reduce energy costs by ~30% vs grid power
Meet sustainability commitments
Insulate against energy price volatility

The Numbers That Made Me Skeptical

I’m a numbers person, so let’s talk about the economics.

Meta’s Spending Trajectory

YearTotal CapEx% of RevenueYoY Growth

2024	~$50B	20%	-
2025	$66-72B	37%	+44%
2026 (est)	~$97B	TBD	+35%

Source: Meta Q3 2025 Earnings, Wall Street consensus estimates

Through 2028: Meta has outlined plans to invest $600 billion in US data centers and infrastructure.

That’s $200 billion per year for three years.

The ROI Question

Here’s where it gets uncomfortable.

Current AI Revenue: Meta doesn’t break out AI-specific revenue, but analysts estimate:

AI-enhanced ads: ~$5-8B incremental revenue (2025)
Direct AI products: Minimal (mostly free)

Infrastructure Costs:

2025 CapEx: $66-72B
Operating costs: Additional $10-15B/year (estimated)

Simple Math:

Revenue from AI: ~$7B Cost of AI infrastructure: ~$80B Net: -$73B

That’s a 10:1 cost-to-revenue ratio.

When Does This Pay Off?

Meta’s bet is that AI will:

Improve ad targeting (higher CPMs, better conversion)
Enable new products (AI assistants, business tools)
Create infrastructure business (sell excess capacity)

Bull Case Timeline:

2026-2027: AI products launch, revenue ramps
2028-2029: Infrastructure-as-a-service business scales
2030+: Positive ROI on cumulative investment

Bear Case: AI doesn’t deliver proportional revenue growth, and Meta is stuck with:

$600B in sunk costs
16-year lease commitments
Massive depreciation expenses

I had to ask: is this sustainable?

The Bull Case: “It’s Different This Time”

Argument 1: Cash Flow Unlike dot-com startups, today’s AI giants generate massive cash:

Meta Q3 2025 operating cash flow: $24.7B
Microsoft: $30B+ per quarter
Google: $25B+ per quarter

Source: Company earnings reports, Q3 2025

Argument 2: Real Demand AI infrastructure is being used:

ChatGPT: 200M+ weekly active users
GitHub Copilot: 1.8M+ paid subscribers
Enterprise AI: $50B+ market (2025)

Source: OpenAI Blog, GitHub Stats

Argument 3: Defensive Necessity Companies aren’t spending because they want to-they’re spending because they have to. If Meta doesn’t build this infrastructure, Google or Microsoft will, and Meta loses competitive position.

The Bear Case: “Circular Financing Red Flags”

Concern 1: Vendor Financing Nvidia is investing in its customers (OpenAI, CoreWeave), who then buy Nvidia chips. That’s circular.

Example:

Nvidia invests $5B in Intel (October 2025)
Nvidia commits $6.3B to buy CoreWeave capacity through 2032
CoreWeave uses that to buy more Nvidia chips

DA Davidson analyst Gil Luria told Yahoo Finance (October 14, 2025):

“They’re using that capital to raise debt. It’s the levering up that’s the truly unhealthy behavior.”

Concern 2: Valuation Disconnect CoreWeave market cap: $67 billion CoreWeave contracted revenue: $43 billion (OpenAI + Meta + Nvidia)

That’s a 1.6× revenue multiple for a company that:

Isn’t profitable yet
Has 71% revenue concentration (Microsoft)
Operates in capital-intensive business

Concern 3: Macro Dependency Deutsche Bank analysis (September 2025) found:

“Without AI-related investment, the US economy might already be in a recession.”

AI spending accounted for 1.1% of US GDP growth in H1 2025.

Source: Deutsche Bank Research, September 2025

If AI spending slows, it could trigger broader economic weakness.

My Take: Bubble with Substance

Here’s what I think after digging through the data:

Yes, there are bubble characteristics:

Circular financing
Aggressive valuations
Hype-driven investment

But there’s real substance:

Actual products with millions of users
Cash-generative businesses (not dot-com burn rates)
Infrastructure that will be useful regardless (compute demand is real)

The Risk: Not that AI is fake, but that the timing and scale of returns don’t match the timing and scale of investment.

Meta might be right about AI’s importance but wrong about how quickly it generates revenue.

The Competitive Landscape: Who’s Spending What

Meta isn’t alone in this infrastructure arms race.

Big Tech AI CapEx (2025)

Company2025 CapExAI FocusKey Deals

Meta	$66-72B	37% of revenue	CoreWeave, Oracle, Blue Owl
Microsoft	$125B (FY26 est)	Azure AI, Copilot	OpenAI partnership
Google	$82.4B	TPUs, Anthropic	1M TPU chips to Anthropic
Amazon	~$75B	AWS AI services	Trainium chips
Apple	~$30B	On-device AI	Apple Silicon

Source: Company earnings, Bank of America estimates, FactSet

Total: ~$400 billion in AI infrastructure spending across Big Tech in 2025.

The CoreWeave Factor

CoreWeave has become the infrastructure kingmaker:

Major Contracts:

OpenAI: $22.4 billion
Meta: $14.2 billion
Nvidia: $6.3 billion (backstop)

Total: $42.9 billion in contracted revenue through 2032.

Stock Performance:

IPO: March 2025
Current: +235% YTD
Market cap: $67 billion

But here’s the concern: 71% of revenue comes from Microsoft (Q2 2025). If Microsoft builds its own infrastructure or switches providers, CoreWeave’s business model breaks.

What I Learned: Three Key Insights

After spending weeks researching this, three things stand out:

1. Infrastructure Is the New Moat

In the AI era, the competitive advantage isn’t just algorithms-it’s infrastructure.

Meta can’t compete with OpenAI on model quality alone. But if Meta has:

2GW of dedicated compute
Exclusive access to latest Nvidia chips
Vertically integrated stack (data centers → models → products)

Then Meta can iterate faster, train bigger models, and serve more users.

The Insight: AI is becoming an infrastructure business, not just a software business.

2. Private Credit Is Reshaping Tech Finance

The Hyperion deal represents a new model: infrastructure-as-a-service financed by private credit.

Traditional tech companies avoided debt (except Apple). But AI infrastructure is so capital-intensive that equity financing alone can’t scale fast enough.

The Shift:

Old: Equity → CapEx → Owned assets
New: Debt → JV → Leased assets

The Risk: If AI doesn’t deliver, tech companies are stuck with debt obligations they can’t service.

3. The Winner Isn’t Clear Yet

Everyone’s spending billions, but nobody knows what the winning AI product looks like.

Current Revenue Leaders:

ChatGPT Plus: ~$2B ARR (estimated)
GitHub Copilot: ~$1B ARR
Midjourney: ~$500M ARR

Infrastructure Spending:

Meta: $75B in deals (Q3-Q4 2025)
Microsoft: $125B (FY26)
Google: $82B (2025)

The Math: $400B+ in infrastructure spending to support ~$10B in AI product revenue.

That’s a 40:1 investment-to-revenue ratio.

The Question: Will AI revenue scale 40× in the next 3-5 years? Or will infrastructure spending collapse when companies realize the ROI isn’t there?

Let me be honest about what could go wrong.

The Good

1. Scale Advantage With 2GW+ of dedicated infrastructure, Meta can:

Train models faster than competitors
Serve billions of users without cloud costs
Experiment with new architectures cheaply

2. Vertical Integration Owning the full stack (data centers → chips → models → products) means:

No vendor lock-in
Better margins long-term
Faster iteration cycles

3. Optionality Even if Meta’s AI products fail, the infrastructure has value:

Sell excess capacity (like AWS)
Lease to other companies
Repurpose for other workloads

The Not-So-Good

1. Massive Capital Commitment $600B through 2028 is irreversible. If AI doesn’t deliver:

Sunk costs can’t be recovered
Depreciation hits earnings for years
Shareholder pressure mounts

2. Execution Risk Building 2GW data centers is hard:

Construction delays (Hyperion target: 2030)
Power infrastructure challenges
Cooling and energy efficiency

3. Competitive Pressure If Google or Microsoft’s AI products win, Meta’s infrastructure advantage doesn’t matter. Users will use the best product, regardless of who has the biggest data center.

The Cost Calculation That Surprised Me

Let me show you the math that made me realize how big this bet really is.

Meta’s AI Infrastructure Costs (2025-2028):

CapEx (2025-2028): $600B Operating costs (4 years @ $15B/year): $60B Total: $660B

Break-even Scenarios:

Scenario 1: Ad Revenue Improvement

Assumption: AI improves ad targeting, increasing CPM by 10%
Meta’s 2025 ad revenue: ~$150B
10% improvement: $15B/year
Years to break even: 44 years

Scenario 2: New AI Products

Assumption: Meta launches AI assistant, business tools
Target: $50B/year in new revenue by 2030
Years to break even: 13 years

Scenario 3: Infrastructure-as-a-Service

Assumption: Meta sells 30% of excess capacity
Pricing: $0.50/GPU-hour (market rate)
Potential revenue: $20B/year
Years to break even: 33 years

Reality Check: Meta needs a combination of all three scenarios to hit reasonable ROI timelines (5-7 years).

Resources:

Further Reading:

Connect

GitHub: @0xReLogic
LinkedIn: Allen Elzayn

Read Entire Article