Anthropic $35bn chip deal puts private credit in AI

Apollo Global Management and Blackstone have finalized a $35 billion financing package for Anthropic’s chip capacity, turning one of the AI boom’s largest compute purchases into a private-credit test case rather than a conventional Big Tech capex line.

In practical terms, the move pushes Wall Street deeper into the physical plumbing of artificial intelligence. Anthropic gets access to custom chips and data-centre capacity linked to Broadcom and Google. Apollo and Blackstone get a new asset class built around leases, SPVs and hardware collateral. Public investors, meanwhile, get an early look at how AI companies may fund infrastructure before they reach the IPO market.

Viewed through credit markets, the package described by the Financial Times as a $35 billion chip financing deal is not just debt attached to another fast-growing software company. It is a claim on future compute demand. That makes it closer to aircraft leasing or project finance than to the buyout loans private credit became famous for after the financial crisis.

For insiders, the structure has a clean logic. Anthropic needs more compute without forcing every dollar of capacity onto its own balance sheet. Broadcom wants its custom AI silicon and networking gear to become a financeable platform, not merely a component sale. Apollo and Blackstone want long-duration assets with contractual cash flows. If the leases perform, all three can argue that the AI infrastructure market has found a repeatable funding model.

Analysts are likely to read the same evidence less comfortably. The structure that makes compute easier to buy also moves risk into places investors cannot see as clearly: note tranches, residual-value assumptions and the resale value of expensive chips whose economic life may be shorter than their depreciation schedule.

The trade Wall Street wants

At launch, the $35 billion transaction funds more than 1 gigawatt of capacity and starts a platform that Blackstone said is designed to accelerate more than 20 gigawatts of global AI deployments through 2028. That is the point. The first deal is large enough to matter on its own, but the bigger prize is a financing template for the next wave of frontier-lab demand.

In Apollo’s telling, it led the $35 billion capital solution for Broadcom’s AI XPV platform with Blackstone and a group of global banks. Reuters reported that the package backs Anthropic’s capacity expansion under a new Broadcom tie-up, including access to 3.5 gigawatts of Google compute capacity starting next year.

From there, the transaction looks different from a normal corporate borrowing. The borrower story is Anthropic. The asset story is compute. The finance story is whether private credit can turn chips, networking equipment and data-centre leases into collateral that pension funds and insurers are willing to hold at scale.

Rao cast the arrangement as disciplined growth. In Anthropic’s announcement of the Google and Broadcom partnership, CFO Krishna Rao described the deal as part of the company’s disciplined approach to scaling infrastructure.

His wording matters because every AI lab now needs to sound disciplined as investors ask whether model revenue can catch up with the cost of training and inference. Compute is no longer just a technical bottleneck. It is a balance-sheet problem.

Chips become collateral

At the chip layer, Broadcom sits at the centre of that shift because custom AI chips are becoming easier to finance when they are embedded in a broader platform. Anthropic said its expanded partnership with Google and Broadcom gives it access to next-generation compute. Reuters said the deal funds Google chips Broadcom helped develop. The result is a financing package that treats non-Nvidia capacity as bankable, provided enough demand is locked in.

Broadcom benefits from more than a hardware order. Nvidia still dominates the market for AI accelerators, but the biggest cloud and AI customers are trying to diversify supply. If Broadcom-linked XPUs and networking equipment can be funded through outside capital pools, the company is helping define the financing layer around that hardware.

According to Bloomberg, the debt stack includes $6 billion of A1 notes, $24 billion of A2 notes and $4.5 billion of B notes, with coupons of 5.75 per cent on the A2 notes and 8.5 per cent on the B notes. Bloomberg also reported that Broadcom is providing support tied to the senior notes and residual values. Those details turn the story from a simple AI-growth headline into a credit-underwriting question.

Tan described the transaction as a way to combine technology and investor capital.

“Our strategic vision is to bring together Broadcom’s leading technology and investor partners…”
— Hock Tan, CEO of Broadcom, according to Bloomberg

A useful structure is not the same as a riskless one. If demand remains tight, leases can look infrastructure-like. If model economics disappoint, newer chips arrive faster than expected, or a frontier lab slows capacity growth, the collateral could age more like technology hardware than a toll road.

In its Quarterly Review, the Bank for International Settlements has already warned that AI infrastructure investment is being supported by both on- and off-balance-sheet borrowing. Its analysis of AI infrastructure finance places the Anthropic-style structure inside a larger pattern: data-centre capex has become big enough to affect credit markets, not just tech valuations.

The private-credit pivot

Debt is expanding in the AI buildout because the spending curve is too steep for equity alone. Software companies used to offer lenders high margins, recurring revenue and relatively light fixed assets. AI changes that bargain. Training clusters, inference capacity, power contracts and data centres require cash upfront, long before the final revenue mix is obvious.

Oracle, Amazon and Nvidia show the same pressure from different angles. MarketWatch reported that Nvidia is pursuing a historic bond deal to fund AI ambitions. CNBC reported that Oracle’s shares fell after the company beat earnings but planned to raise more money for its AI buildout. TechCrunch noted that Amazon recently borrowed $17.5 billion from banks shortly after a bond sale. The common thread is simple: even the strongest technology franchises are treating debt as part of the AI supply chain.

Managers such as Apollo and Blackstone see opportunity in that mismatch. Their pitch is that they can underwrite complexity faster than public bond buyers, especially where contracts, equipment and specialist collateral sit between the borrower and the end user. Apollo has built a business around that kind of origination. Blackstone has the real-assets and credit machinery to scale it.

Sceptics will focus on whether the market begins to price AI infrastructure as if demand is certain. It is not. Claude, ChatGPT and Gemini are growing products, but frontier AI revenue still has to justify enormous capital intensity. A lease-backed structure can reduce the burden on Anthropic’s reported balance sheet, yet it cannot erase the system-wide exposure if too much compute is financed against aggressive utilisation assumptions.

For regulators, SPVs can make risk more distributable. They can also make it harder to see where the risk finally sits. If insurance portfolios, private funds and banks all hold slices of the same compute boom, a repricing in chips or leases would not remain a Silicon Valley problem for long.

IPOs need a funding story

Timing adds another layer. Bloomberg has linked the financing to Anthropic’s confidential IPO preparations, while the same capital-markets window is opening around OpenAI. MarketWatch and CNBC have both framed 2026 as a year of large technology listings, with AI companies preparing to test public-market appetite after years of private funding rounds.

Any IPO filing from a frontier lab will have to answer a question that user growth alone cannot settle: how much capital is needed to keep models competitive, and where does that capital come from? The menu is already visible. Equity dilution, corporate debt, vendor financing, cloud partnerships and outside infrastructure vehicles are all in play. The Apollo-Blackstone structure points toward the last of those options.

A successful version may become the template. A weaker version may become the stress point. If Anthropic can grow revenue into the capacity, private credit will look like an accelerant for the AI economy. If the economics fall short, the same deal will look like an early example of Wall Street manufacturing yield from a technology cycle it did not fully understand.

For now, the important signal is not that Anthropic raised $35 billion. It is that two of the world’s largest alternative-asset managers were willing to fund the machinery behind a frontier AI lab before public investors get their turn. The AI arms race is no longer just a battle for chips, models and users. It is becoming a race to decide whose balance sheet can carry the infrastructure first.