Issue #4 – December 15, 2025

The Week AI Infrastructure Showed Its Cracks

This week delivered a wake-up call for communities negotiating data center deals: the companies making billion-dollar promises are struggling with trillion-dollar debt. While AI deployment remains real and measurable, the infrastructure boom financing it looks increasingly precarious.

LEAD STORY: Oracle’s $10 Billion Problem

Oracle reported earnings Wednesday that should alarm any community in data center negotiations. The company posted negative $10 billion free cash flow in the quarter ending November 30—double Wall Street’s expected deficit—while announcing $248 billion in lease commitments for data centers it hasn’t built yet.

The stock dropped 11% Thursday, erasing $80 billion in market value. By Friday, Bloomberg reported Oracle pushing OpenAI data center completions from 2027 to 2028 due to “labor and material shortages.” The stock fell another 4.5%. Oracle disputes this.

The Michigan Connection:

Oracle promises Michigan $7 billion for the Saline Township Stargate campus. But Oracle’s financial position reveals critical vulnerabilities:

Total debt: $124 billion (up from $89B a year ago)
Capex raised to $50 billion for fiscal year (from $35B in September)
Credit default swaps at 5-year highs (financial crisis levels)
Morgan Stanley warning: Oracle’s debt could triple in next 3 years

Fortune magazine’s analysis hit the core issue: “Oracle’s rapid descent from market darling to market warning sign is revealing something deeper about the AI boom—no matter how euphoric investors became, the industry can’t outrun the laws of physics or the realities of debt financing.”

The timing couldn’t be worse for Michigan. Dana Nessel is challenging Whitmer’s fast-track approval while Oracle’s finances crater. Communities are approving 20-year tax breaks for a company Wall Street questions can meet its commitments.

Before you question our earlier analysis that “AI is not a Bubble“, read on…

The Debt Crisis Extends Beyond Oracle

Oracle isn’t alone in financial trouble. Research this week reveals systemic overleveraging across the data center industry:

CoreWeave: The AI Infrastructure Cautionary Tale

Fortune’s deep-dive on CoreWeave (November 2025) shows the bleeding edge of data center debt:

$11 billion total debt on only $5.15B projected 2025 revenue
$7.6 billion in current liabilities (due within 12 months)
Lost $600 million on $2.2 billion revenue in first half of 2025
Operating margins: 1.6% (turns sharply negative after interest expenses)
71% of revenue from Microsoft alone (dangerous customer concentration)
$34 billion in lease payments kicking in 2026-2028

Kerrisdale Capital calls CoreWeave “a debt-fueled GPU rental business with no moat, dressed up as innovation” and projects 90% downside.

The financing strain hit closer to home Friday when Fermi, a data center REIT, saw its stock plummet 40% after a potential anchor customer backed out of funding commitments. The collapse demonstrates that financing instability isn’t limited to hyperscalers—even specialized real estate investment trusts built specifically for data center development are struggling to secure committed capital.

The Industry-Wide Pattern

Bank of America analysis reveals the five largest AI infrastructure companies—Google, Meta, Amazon, Microsoft, and Oracle—collectively issued $121 billion in bonds in 2025 to fund data center buildouts, a level “far above historical averages.”

Key differences matter:

The Big Four (Google, Meta, Amazon, Microsoft) maintain AA/A credit ratings and generate large positive free cash flow
Oracle sits in BBB territory with negative cash flow and $100B+ debt

Bloomberg’s December 12 investigation uncovered what skeptics call “financial engineering”:

Borrowers seeking loans for 150% of construction cost
Two data center billionaires minted before anything built
Companies using “special purpose vehicles” to keep liabilities off balance sheets
JPMorgan estimates AI needs $650 billion in annual revenue by 2030 just to generate 10% return on capex

The last time “special purpose vehicles” dominated headlines? Enron in 2001.

OpenAI’s Strategy: Leverage Other People’s Balance Sheets

Perhaps most revealing: OpenAI’s data center partners are on track to amass $100 billion in borrowing tied to the loss-making startup, while OpenAI itself avoids direct financial risk.

A senior OpenAI executive admitted to reporters: “That’s been kind of the strategy. How does [OpenAI] leverage other people’s balance sheets?”

This $100 billion in bonds, bank loans, and private credit equals the net debt of the six largest corporate borrowers worldwide—including Volkswagen, Toyota, AT&T, and Comcast.

Texas Warning: The Math Simply Doesn’t Work

While Michigan debates individual projects, Texas provides the national reality check. The Electric Reliability Council of Texas (ERCOT) reported 226 gigawatts in data center interconnection requests for 2030.

The problem: Texas has 103 GW total generation capacity and 85 GW peak summer demand.

Requested capacity exceeds the entire Texas grid by more than 2x.

“The top line numbers are almost laughable,” said Joshua Rhodes, University of Texas researcher and energy consultant. “It definitely looks, smells, feels—is acting like a bubble.”

The Capacity Crunch Details

225 applications in 2025 alone—more than 2022-2024 combined
73% are data centers (AI has replaced Bitcoin mining)
Generation queue: 432 GW total requests, but 77% solar/battery (intermittent)
Firm power (gas): Only 48 GW—nowhere near enough to support reliable data center operation

Translation: Load growing fast, firm generation NOT keeping up.

Michigan Faces the Same Math

Michigan’s situation mirrors Texas at smaller scale:

Current capacity: 18.6 GW from DTE alone
Data center demand: At least 22 GW already announced
Saline Township alone: 1.4 GW (roughly 8% of DTE’s total capacity)

DTE promises Saline won’t require new generation, using “existing resources augmented by battery storage.” But where does that power come from when 20+ other Michigan projects make similar demands?

Michigan Update: Lyon Township Joins the Pattern

Oakland County discovered its data center surprise this week. Residents in Lyon Township learned via Facebook that Project Flex—a 172-acre, six-building AI data center—received conditional approval from the planning commission in September.

The playbook is identical to Saline and Washington Township:

No public hearings (industrial zoning permits data centers as-of-right)
Planning commission approval (appointed officials, not elected)
Residents discover via social media after approval
Just 300 feet from residential zones

Developer Verrus (California-based, backed by Sidewalk Infrastructure Partners) promises “smart energy use” through micro-grids and battery storage. But like other Michigan projects, critical details remain undisclosed:

Investment amount unknown
Job numbers not specified
Water and power requirements unclear
Still needs sound study and final engineering review

WXYZ reporter Christiana Ford captured resident Kimberly Killian’s reaction: “I’m extremely upset, I’m heartbroken, nobody wants that in their backyard.”

Hours after Lyon Township’s discovery, MLive reported another data center plan emerging for 280 acres off I-94 in Southeast Michigan. Details remain scarce, but the rapid-fire announcements—Lyon Township Wednesday, I-94 site Thursday—reinforce the pattern of projects appearing faster than communities can track them.

The Michigan Landscape

Planet Detroit’s weekly roundup now tracks multiple simultaneous battles:

Active Projects:

Saline Township (Oracle/OpenAI Stargate) – $7B
Lyon Township (Verrus) – 172 acres
Lowell Township – Public hearing rescheduled to high school due to overflow crowd
Augusta Township (Thor Equities) – Facing zoning rejection
Ypsilanti (U-M/Los Alamos) – $100M grant, $6M per promised job

Policy Battles:

Dana Nessel challenging MPSC rules on Consumers Energy data centers
Activists planning Lansing rally for statewide moratorium
Howell Township enacted 6-month moratorium (trustees say projects “inevitable”)

The Pattern: Data center industry sources estimate 59 existing facilities in Michigan operated by 29 providers. University of Virginia researcher João Ferreira notes that “much more are being proposed than will be built”—developers pitch multiple counties simultaneously.

His analysis of Great Lakes region data shows about 525 operating data centers as of December 2024, with 220 more planned through 2030. Michigan currently trails Illinois and Ohio, which account for 50% of operational facilities and 60% of new development.

Critical insight: “These aren’t new; they just happen to be new to Michigan,” says Phil Santer of Ann Arbor SPARK. “But this is a pretty mature industry in most of the United States. Michigan is not a hot spot. We’re not a hot spot yet, at all.”

AI Deployment vs. Infrastructure Speculation: Critical Distinction

This week also demonstrated why the “AI bubble” narrative requires nuance. AI deployment is real and measurable. Data center financing might be speculative.

But here’s the critical part: debt problems don’t mean the infrastructure isn’t needed.

Why Overleveraged Winners Survive (and Losers Get Absorbed)

The debt-to-revenue ratios alarming Wall Street don’t necessarily predict infrastructure collapse. They predict consolidation.

History shows that transformative infrastructure often gets overbuilt on shaky financing—then the winners absorb the losers as demand catches up to capacity.

The Facebook/Meta Precedent:

Meta faced similar scrutiny during its infrastructure buildout:

Massive capex spending before monetization proved out
Wall Street questioned whether social media justified the investment
Short-term metrics looked terrible
Long-term demand validated the infrastructure spend

Meta now generates tens of billions in annual free cash flow and leads AI infrastructure deployment. The demand was real; the timing just required patience and deep pockets.

Current AI Infrastructure: Similar Pattern

The $121 billion in bonds issued in 2025 looks different when you separate winners from losers:

Winners (will survive debt):

Google, Meta, Amazon, Microsoft: AA/A credit ratings, massive positive cash flow
Can weather 2026-2027 debt maturities
Will likely acquire distressed competitors’ assets

Losers (perhaps facing absorption):

Oracle: BBB rating, negative $10B cash flow, $124B debt
CoreWeave: $11B debt on $5B revenue, 1.6% margins
Smaller operators with concentrated customer bases

The Absorption Mechanism:

When CoreWeave’s $7.6 billion in current liabilities come due in 2026-2027, one of three potential things happen:

Microsoft (71% of their revenue) acquires them outright
They restructure debt at punishing terms
Asset sales to better-capitalized players

Either way, the physical infrastructure doesn’t disappear. The data centers, power contracts, and GPU clusters get absorbed by stronger operators. Bondholders take losses. But the infrastructure survives because demand is real.

Why This Time Really Is Different: “Too Big to Fail” Infrastructure

The federal government and Fortune 500 corporations have made AI infrastructure systemically important:

Government Commitment:

Trump’s July + December executive orders fast-tracking development
$500 billion Stargate project with presidential backing
National security positioning (beating China narrative)
Federal agencies deploying AI at scale

Corporate Integration:

Microsoft India: 200,000 Copilot licenses deployed (not testing—deployed)
Salesforce: 19,000 Agentforce customers
Snowflake: 5,600 customers
Combined: 24,600+ companies with AI agents in production

The Numbers That Matter:

When enterprise deployment reaches this scale—hundreds of thousands of workers using AI daily, tens of thousands of companies in production—the infrastructure supporting it becomes critical infrastructure.

The federal government won’t allow catastrophic infrastructure failure when:

Major defense contractors depend on it (Lockheed Martin, Raytheon, etc.)
Financial system runs on it (JP Morgan, Goldman Sachs deploying AI)
Healthcare requires it (AI diagnostics at scale)
Government itself uses it (federal AI deployment across agencies)

What Debt Crisis Actually Means for Communities

This “too big to fail” dynamic creates a counterintuitive situation for communities:

The Infrastructure Will Get Built (demand is real, government backing is real, corporate deployment is real)

But WHO Builds It Matters:

Scenario A: Oracle completes Saline before debt crisis hits. Community gets infrastructure, Oracle potentially gets acquired/restructured later. Infrastructure remains.
Scenario B: Oracle fails before completion. Stronger player (Microsoft, Google, Amazon) acquires assets. Community still gets infrastructure, but negotiated terms might change.
Scenario C: Project never breaks ground. Community gave tax breaks for nothing. Oracle’s debt problems become community’s lost revenue.

The Risk Isn’t “No Data Centers”—It’s “Wrong Deal Terms”

Given that AI infrastructure is likely “too big to fail” in addition to the demand, communities should negotiate assuming:

Infrastructure will eventually get built (demand is real)
Current developer might not be final owner (debt problems)
Tax breaks and incentives should protect community if ownership changes
Grid upgrades and public investment should benefit community regardless of which company operates the facility

Why Oracle Specifically Concerns Us

The “absorption will happen” argument works for the industry overall but creates specific risks for Oracle deals:

If Oracle is acquired/restructured in 2026-2027:

Will new owners honor community commitments?
Are tax agreements transferable?
Who pays for DTE grid upgrades if project stalls?
Does community have recourse?

The Smart Play:

Communities should negotiate terms assuming ownership might change. Include provisions for:

Tax break clawbacks if construction doesn’t start
Performance guarantees that survive ownership changes
Grid upgrade costs that don’t fall on residents
Job creation minimums with financial penalties

The infrastructure IS coming. But communities can still get burned by weak deal terms with unstable developers.

Real AI Deployment: Microsoft India

On December 11, Microsoft announced partnerships with four Indian IT giants—Cognizant, Infosys, TCS, and Wipro—to deploy 200,000+ Microsoft Copilot licenses collectively. Each company commits to 50,000+ licenses.

This represents:

Immediate workforce transformation (not future speculation)
Enterprise-scale deployment (not pilot programs)
Measurable worker impact (200,000 employees using AI daily)

TCS alone ran a hackathon with 281,000 participants. All TCS employees now have personalized AI coaches. Wipro deployed 50,000+ licenses and upskilled 25,000+ employees in Microsoft Cloud and GitHub technologies.

This is corporate AI integration at scale—happening now, with clear workforce implications.

Speculative Infrastructure: The Debt Pyramid

Contrast Microsoft’s operational AI deployment with the data center debt structure:

Operational AI (Healthy):

Microsoft generates massive positive free cash flow
Maintains AA/A credit rating
Deploys AI to improve existing operations
Revenue and investment aligned

Infrastructure Speculation (Concerning):

Oracle: Negative $10B cash flow, BBB rating, promises $248B in leases
CoreWeave: $11B debt on $5B revenue, 1.6% operating margins
OpenAI partners: $100B debt for loss-making startup
Industry: $121B bonds issued in 2025, 300% above historical norms

The AI isn’t the bubble. The financing structure underneath it might be.

What Communities Should Watch

The Oracle earnings caution and Texas capacity warnings provide a roadmap for community protection:

1. Financial Verification Before Approval

CIO.com published analysis Thursday advising enterprise customers on Oracle negotiations. Their guidance applies equally to communities:

Oracle customers face “clear and escalating risk of price increases” due to capital cycle where spending outpaces monetization
“The combination of heavy capex, negative free cash flow, increasing financing cost and long-dated revenue commitments forms a structural pressure” that shows up in commercial terms
Enterprise CIOs should develop “architectural optionality”—essentially, escape plans

Translation for communities: Don’t approve until you verify the developer can actually fund construction. Oracle’s promises while posting negative $10B cash flow should raise red flags.

2. Grid Capacity Reality Checks

Texas shows promised capacity often exceeds physical possibility. Michigan communities should demand:

Independent verification of DTE/Consumers capacity (not utility promises)
Written guarantees that residential rates won’t increase
Clarity on who pays for grid upgrades if capacity insufficient
Details on “battery storage” claims (how much, from where, financed how?)

3. Job Calculation Honesty

Tech Policy Press analysis reveals Michigan’s job creation math:

U-M/Los Alamos data center: $6 million per promised job
Detroit Diesel initiative: $63,000 per job
Switch Grand Rapids: Promised 1,000 jobs, delivered 26 by 2022, kept tax exemptions

The construction jobs are temporary. The operational jobs are often automated within 3-7 years. Communities should calculate total public investment per permanent job before approval.

4. Federal Policy Monitoring

Trump’s two-pronged executive order strategy (July data center fast-tracking + December state law preemption) shows federal pressure to override local control. Communities should:

Monitor Federal Register for regulatory changes
Watch state Attorney General responses (Dana Nessel model)
Track industry celebration as warning sign (when developers cheer, communities should worry)
Follow permitting timeline compression (60-day approvals = red flag)

5. The Negotiating Window

CIO.com’s most important insight for enterprise customers applies to communities:

“Oracle’s need to demonstrate utilization and revenue conversion over the next several quarters create windows of disproportionate buyer leverage. CIOs that come to the table now can secure far more favorable economic outcomes than those that wait until Oracle’s cash flow stabilizes and its bargaining power returns.”

Translation: Communities have leverage RIGHT NOW while Oracle is desperate. But only if they haven’t already approved.

Lyon Township approved in September before Oracle’s financial crisis became visible. Saline is still negotiating during Oracle’s weakest position. The difference in leverage is enormous.

Looking Ahead: Infrastructure Beyond Data Centers

Next week’s analysis will explore the infrastructure landscape beyond data centers:

Space ground stations (Starlink and competitors)
Subsea cables (the actual AI infrastructure bottleneck)
Small Modular Reactors (data centers driving nuclear renaissance)
Carbon capture facilities (climate infrastructure boom)

The data center frenzy obscures a broader infrastructure transformation. Understanding which projects represent genuine economic opportunity versus speculative bubbles requires tracking all major infrastructure categories.

Preview question: If data centers prove to be overbuilt speculation, what infrastructure actually supports long-term job creation and community benefit?

Intelligence Infrastructure Update

PivotIntel is expanding from national data center tracking to comprehensive Michigan project monitoring. Our database will track:

All known Michigan data center projects (currently 20+ active proposals)
Developer financial positions (Oracle situation shows why this matters)
Grid capacity vs. announced demand
Actual job creation vs. promises
Community opposition and success stories

This real-time intelligence helps workers and communities see patterns before making irreversible commitments.

Why This Matters: By the time Oracle’s stock crashed 11%, Lyon Township had already approved their project. Real-time intelligence creates leverage before approval, not regret after.

The Bottom Line

This week revealed a critical gap between AI reality and infrastructure promises:

Real: Microsoft deploying 200,000 Copilot licenses to Indian IT firms. Workers experiencing transformation now.

Questionable: Oracle promising Michigan $7 billion while posting negative $10 billion cash flow. Texas requesting 226 GW from 103 GW grid. Industry issuing $121 billion in bonds at 3x historical rates.

The Disconnect: AI deployment is measurably real. Data center financing increasingly resembles the dot-com fiber buildout bubble—real technology, speculative infrastructure, inflated promises.

Communities approving projects today based on developer promises should consider: If Oracle can’t generate cash flow to service its debt, who actually pays for the infrastructure upgrades, tax breaks, and grid expansions?

The answer, as always, is the community.

Next Week: Deep-dive analysis of infrastructure alternatives—the projects that might actually deliver on job creation and economic benefit when the data center bubble deflates.

PivotIntel tracks AI infrastructure development with focus on worker impact and community consequences. Sign up for weekly intelligence at theopenrecord.org/pivotintel

Previous Coverage:

Under the Radar: AI Isn’t a Bubble series
Under the Radar: What the 1980s-90s IT Revolution Teaches Us About AI
PivotIntel #3: Transparency & Data Center Approval Battles

Sources archived via Wayback Machine for permanent reference. Full source list available at theopenrecord.org/pivotintel

Sources

All sources archived via Wayback Machine for permanent reference:

Lyon Township & Michigan Projects:

Oracle Financial Crisis:

Industry Debt & Financial Analysis:

Texas Grid Analysis:

Microsoft Enterprise Deployment:

Microsoft News: India Copilot partnerships

Worker Retraining Research:

Additional sources pending Wayback Machine archiving. Full methodology at pivotintel.org/about/methodology