Inside the $35bn deal: Apollo and Blackstone's chip-backed SPV for Anthropic signals a new financing era

Inside the $35bn deal reshaping AI infrastructure finance

Apollo Global Management and Blackstone finalised a $35 billion financing package for Anthropic last week. The record-breaking private credit deal, chip-backed special purpose vehicle is one of the largest transactions of its kind in history.

The initial commitment will increase Anthropic’s AI computing capacity by one gigawatt (enough to power approximately 750,000 homes) with that capacity scheduled for deployment at sites operated by Fluidstack starting mid-2026. And the broader partnership intends to enable more than 20 gigawatts of computing capacity for leading AI labs, including OpenAI, through 2028.

“Private equity firms have become a vital source of funding for AI companies that face a shortage of costly and supply-constrained AI infrastructure needed to meet rising demand,” said one infrastructure analyst tracking the deal closely.

“The question is no longer whether capital is available. It is whether the physical infrastructure can absorb it fast enough.”

As Capacity reported in November 2025, 83% of industry experts surveyed by Turner & Townsend do not believe supply chains are equipped to deliver the advanced cooling systems required for AI-ready facilities, a finding that sits awkwardly alongside commitments of this scale.

Why Fluidstack, and what it tells the market

The selection of Fluidstack as the infrastructure partner for Anthropic’s initial gigawatt deployment is the detail that has received the least attention and deserves the most. Anthropic cited Fluidstack’s “exceptional agility” in delivering gigawatt-scale power without long lead times.

The facilities being developed are designed from the ground up around Anthropic’s most advanced systems, giving the company dedicated compute lanes instead of waiting for traditional cloud providers to catch up.

This is a significant signal for the data centre sector. A company with access to Google Cloud TPUs at scale, an $11 billion AWS campus already operational in Indiana, and a multi-cloud strategy stretching across Microsoft Azure, nonetheless went to a UK-based neocloud provider for its most time-sensitive build. The reason was not cost. It was delivery certainty and workload customisation, two attributes that hyperscalers, despite their balance sheets, are increasingly struggling to guarantee at the pace frontier AI development demands.

Gary Wu, co-founder and CEO of Fluidstack, said: “We’re proud to partner with frontier AI leaders like Anthropic to accelerate and deploy the infrastructure necessary to realise their vision.” Anthropic’s distributed strategy contrasts sharply with the single joint-venture model pursued by OpenAI through Stargate, hedging compute access across multiple providers and ownership models simultaneously.

The SPV structure changes the tenant relationship

The financing architecture of this deal deserves close reading by anyone negotiating leases with AI-era customers, because it introduces a counterparty dynamic that is genuinely new.

The deal relies on a complex structure in which a special purpose vehicle raises debt and equity, with lease agreements for chips ultimately supporting the value of the transaction. The $35 billion facility was structured across three tranches, with Broadcom backing the senior layers to achieve lower borrowing costs aligned with its strong credit profile. Crucially, Broadcom agreed to support the residual value of the chips, which means that if Anthropic defaults on lease payments and the chips are sold for less than the amount owed, Broadcom covers any shortfall for investors holding the senior notes.

What this means in practice is that the risk architecture sitting behind Anthropic’s infrastructure commitments is now considerably more complex than a conventional lease agreement. The tenant is not simply an AI company drawing on its balance sheet. It is a company whose compute obligations are intermediated through a private equity-structured vehicle, backstopped by a semiconductor manufacturer’s credit, across a chip-leasing arrangement that has no clear precedent in the data centre leasing market. If Apollo and Blackstone can syndicate $35 billion for Anthropic, similar structures will emerge for other AI labs.

Hyperscale investment has surged past $93 billion in annual financing, with Google, Amazon, Microsoft and Meta collectively expected to spend approximately $725 billion on capital expenditure in 2026 alone, up 77% from the prior year’s record. JLL’s 2026 Global Data Centre Outlook forecasts global installed capacity almost doubling from 103GW today to around 200GW by 2030, pointing to a $3 trillion investment supercycle driven by AI. Against that backdrop, the Apollo and Blackstone deal is not an outlier – it is a template.

When your chip vendor becomes your lender

Perhaps the most structurally novel element of this transaction is the role Broadcom has assumed and what it signals for how vendor relationships in the AI infrastructure stack are evolving.

Broadcom is not simply supplying chips to this deal. It is carrying credit risk on them. That is a meaningful departure from the conventional vendor-operator relationship, and it has implications that extend well beyond this specific transaction.

In April, Broadcom signed a long-term agreement with Alphabet’s Google to develop and supply future generations of custom AI chips through 2031. The pattern is consistent: Broadcom is positioning itself not just as a component supplier but as an infrastructure finance partner, a role that changes the procurement and risk calculus for every operator deploying its hardware at scale.

This is a new kind of vendor concentration risk. When a chip manufacturer is also, in effect, a guarantor in your tenant’s financing structure, the traditional separation between technology procurement and financial counterparty exposure begins to blur.

The broader context is one of mounting demand-side pressure with constrained supply-side capacity. As Capacity previously reported, data centre construction in North America jumped 69% year-on-year in 2024, while early-stage projects in the UK increased from 2GW to 8GW between 2025 and 2026. Hyperscaler-operated data centres now account for 48% of worldwide capacity and are on course to represent 67% by 2031. The contractors capable of delivering at that pace are already stretched, a concentration risk Capacity examined in depth earlier this year.

Into that constrained environment, a $35 billion commitment targeting 20 gigawatts of additional AI compute capacity by 2028 is not just a financing story. It is a forward claim on power, sites, contractors, cooling supply chains, and grid connections that the market will have to work out how to honour.

DigitalBridge CEO Marc Ganzi, speaking to Capacity earlier this year, framed the structural challenge plainly: “The defining trend for 2026 will be power-informed deployment of digital infrastructure. Another misconception is that AI infrastructure can scale without rethinking power and sustainability. The reality requires disciplined, long-term coordination with utilities, regulators, and communities.”

The Apollo and Blackstone deal adds $35 billion worth of urgency to that coordination challenge. The capital is committed. The gigawatts are promised. Now someone has to build it.

Why Amazon is investing $25bn into Anthropic in a major AI deal

OpenAI’s IPO is the largest compute procurement vehicle in corporate history

Edged US closes $2bn in financing to power AI data centre push

Why is contractor concentration becoming a major risk for AI data centre growth?

Data centres at a crossroads: Can capital keep up with demand?

Marc Ganzi: Why power and AI will redefine the data centre industry