The Anthropic Deal Changes the Scale of This Conversation

Apollo Global Management and Blackstone are actively shopping a $36 billion debt financing package to fund Anthropic's acquisition of Google's custom TPU chips, according to ZeroHedge citing Bloomberg reporting.

Thirty-six billion dollars. For chips. Through private credit.

This would be the largest private credit deal ever and the largest chip-financing debt transaction in history. The deal is partially backstopped by Broadcom's credit quality, creating a layered debt structure that essentially rents computing power to Anthropic through borrowed money.

Anthropic has seen its valuation surge rapidly in recent years. Now it's funding its hardware stack through what amounts to a leveraged buyout of compute capacity.

Moody's Just Made the Math Worse

Moody's Ratings issued a warning this week — not about one deal, but about the entire financing architecture underneath the AI infrastructure boom.

The five major U.S. hyperscalers — Amazon, Meta, Alphabet, Microsoft, and Oracle — have accumulated approximately $662 billion in future data center lease commitments that have not yet commenced, according to Moody's. Add in other commitments and the total undiscounted future lease exposure hits $969 billion.

These aren't speculative projections. These are signed contracts. Obligations. They just haven't shown up in the financial press because they haven't started billing yet.

Apollo Global Management's chief economist Torsten Slok put the broader spending picture in context: total capital expenditure on data centers is estimated at roughly $646 billion — approximately 2% of U.S. GDP. That's roughly equivalent to the combined GDP of Singapore, Sweden, and Argentina. For reference, the entire U.S. defense budget in 2025 ran about $917 billion, according to ZeroHedge citing Slok's analysis.

AI infrastructure spending is now approaching the scale of national defense. Funded increasingly by private debt. With the liabilities parked off-balance-sheet.

What Mainstream Coverage Is Getting Wrong

Most AI coverage focuses on record valuations, impressive demos, and paradigm shifts.

The debt structure underneath all of it is getting increasingly complex and increasingly opaque. The Anthropic deal isn't just big — it's a signal about HOW this buildout is being financed. Private equity firms are engineering layered instruments that route around traditional public market scrutiny. Moody's flagged this specifically: "aggressive financing structures... creating significant systemic risks that could ripple across global credit markets."

ZeroHedge covered this most directly. Bloomberg's article was inaccessible due to a paywall block. VentureBeat didn't touch the financing angle at all — their focus was on a different part of the AI stack entirely.

The Other Bottleneck Nobody's Fixing

While the capital markets debate plays out at the $36 billion level, there's a ground-level problem quietly killing enterprise AI adoption.

It's NOT model performance. According to VentureBeat, enterprise AI agents are stalling because of permissioning — the fundamental question of what an agent is allowed to touch, on whose behalf, and how the system verifies that.

Workday's president for product and technology, Gerrit Kazmaier, told VentureBeat that customers consistently hit this wall when they try to build their own agentic solutions. "The richness of the security model gets lost, and the results become overly broad," Kazmaier said.

Workday's solution is its Sana agent system of record, launched in March, which uses its existing HR and finance data infrastructure as the governance layer for agents. Workday expanded its Google partnership to bring Sana into the Gemini Enterprise ecosystem — meaning agents built on Sana are now discoverable there.

Accuracy and identity are the same problem. Does the system know enough about the agent, the authorizing human, and the current state of the record to act correctly? For HR and finance use cases, "almost right is not acceptable," Kazmaier said. A wrong paycheck processes. An interview gets scheduled incorrectly. The damage is already done before anyone catches it.

This isn't a philosophical AI safety debate. It's a plumbing problem. And it's why most enterprise agentic workflows are still stuck in pilot programs.

What This Means for Regular People

Tens of billions in private debt are funding the compute stack. Nearly a trillion dollars in lease obligations are sitting in the fine print of tech company balance sheets. Moody's — NOT a fringe outlet — is using the word "systemic" to describe the risk.

At the enterprise level, the AI tools companies are paying fortunes to deploy are stalling because nobody sorted out the permission structure.

The capital is flowing. The infrastructure is being built. The debt is piling up. Whether the productivity gains arrive fast enough to justify it remains an open question.