30+ sources. Zero spin.
Unbiased news you can read, scroll, or listen to.
Apollo and Blackstone Close $35 Billion Private Credit Deal to Finance Anthropic's AI Chip Expansion

The Deal in Plain English
Since our June 9 coverage of China's $297 billion data center buildout and the broader AI infrastructure arms race, a concrete piece of that puzzle has landed on this side of the Pacific.
Apollo Global Management and Blackstone finalized a $35 billion private credit arrangement to fund Anthropic's computing expansion, according to the Financial Times as reported by TradingKey on June 9, 2026. The structure: a Special Purpose Vehicle raises capital, buys the chips, then leases them to Anthropic. The SPV collects lease payments. The debt gets repaid from those payments — and from the chips' residual value if things go sideways.
Anthropic is not borrowing money directly. None of this touches Anthropic's balance sheet in any traditional sense.
How the $35 Billion Is Sliced
According to TradingKey, the debt breaks into three tranches:
- $6 billion A1 notes: Treasury rate + 100 basis points, sold to banks, backed by Broadcom's credit guarantee
- $24 billion A2 notes: 5.75% coupon, priced at par, also backed by Broadcom
- $4.5 billion Class B notes: 8.5% coupon, NO Broadcom backing — higher risk, higher yield
Apollo's Atlas SP Partners threw in $800 million in equity to become the SPV's owner.
Broadcom's "residual value support" agreement is central to the deal. If Anthropic defaults on its leases and the chips can't cover the debt, Broadcom compensates A1 and A2 investors 100% for the shortfall. This backstop allows those two senior tranches to price close to investment-grade — without formally appearing on Broadcom's balance sheet either, per TradingKey.
What Chips Are We Talking About?
The SPV is purchasing Google's custom Tensor Processing Units (TPUs) — not Nvidia GPUs, per TradingKey. Intellectia.AI reports the deal also involves Broadcom's custom chips and networking solutions, with a target of delivering over 20 gigawatts of computing capacity to leading AI labs by 2028. The physical infrastructure will be built by Fluidstack, with deployment expected to begin in mid-2026.
Anthropic's initial commitment is to expand capacity by one gigawatt — enough, per Intellectia.AI, to power approximately 750,000 homes. This isn't a server upgrade. It's a power-plant-scale buildout.
Similar Structures Are Multiplying
ZeroHedge noted in January 2026 that Meta's $27.3 billion SPV deal — "Project Beignet" with Blue Owl for the Hyperion data center in Louisiana — was the template, and that hundreds of billions in similar structures would follow in 2026. "Project Big Sky" has arrived, and it's bigger.
This structure's appeal is clear: massive off-balance-sheet debt, investment-grade pricing through creative guarantee structures, and private credit funds filling the gap that traditional banks — already choking on AI debt exposure, per ZeroHedge — can't or won't fill.
What the AI Safety Side Is Doing Meanwhile
Separately, Anthropic on Tuesday released two new models: Claude Fable 5 (public) and Claude Mythos 5 (restricted), according to Wired. Mythos 5 is the same underlying model as Fable 5 but without guardrails — and Anthropic is only releasing it to a limited set of cybersecurity partners under the "Project Glasswing" program, plus select biology researchers.
The concern: Mythos-level capabilities could help bad actors develop novel hacking tools or biological threats faster than defenders can respond. Anthropic's head of Product Management Diane Penn told Wired the company has been wrestling with how to handle these capabilities since before the April limited release, and this two-track rollout — a guardrailed public version and a restricted full version — was the most viable path they found.
Queries related to cybersecurity, biology, and chemistry on Fable 5 get rerouted to the older Claude Opus 4.8. If Anthropic detects someone trying to train a smaller model off Fable 5's responses — so-called "distillation" — that gets rerouted too, per Wired.
The Risk Built Into the Structure
The entire $35 billion structure rests on AI chips holding residual value over multi-year lease terms. GPUs and TPUs depreciate fast. If AI model architecture shifts — say, a major efficiency breakthrough makes current-generation chips obsolete faster than expected — the underlying collateral loses value rapidly. The Class B noteholders at 8.5% have zero Broadcom backstop. Even the A1 and A2 tranches depend on Broadcom's willingness and ability to honor guarantees that don't appear on its balance sheet. ZeroHedge noted that Broadcom's stock fell sharply last week before rebounding Monday, suggesting the market is uncertain about Broadcom's AI exposure right now.
Broadcom's guarantee is a legally binding contractual obligation. The senior tranches are priced accordingly — near investment-grade rates reflect real structural protection, not speculation.
What Retail Investors Should Know
Most financial coverage treats this as a triumphant milestone for private credit and AI infrastructure. What it isn't asking: what happens to retail investors in Apollo's and Blackstone's BDCs — Business Development Companies sold to everyday investors — if this bet goes wrong? ZeroHedge has documented rising redemption requests and gates on these retail-facing vehicles in recent months. The upside of "Project Big Sky" flows to partners and institutional allocators. The downside risk bleeds through to ordinary people who bought private credit products chasing yield.
The Ongoing Wager
The AI infrastructure financing machine is running at full speed, and the creative structures keep growing larger. Anthropic gets its chips. Apollo and Blackstone collect fees and spread. Broadcom cements its position as the non-Nvidia chip play. Fluidstack builds the physical facilities. Everyone wins — until the chips depreciate faster than the models, or the lease payments stop, or the next efficiency breakthrough makes last year's $35 billion in hardware obsolete.