How Google's $150bn Anthropic chip deal reshapes AI finance

Google has quietly built one of the largest infrastructure financing programmes ever put together, a web of contracts worth more than $150bn designed to keep Anthropic supplied with the chips it needs to train and run Claude.

The arrangement pulls in Broadcom, Apollo, Blackstone, Morgan Stanley and a handful of crypto-mining firms turned compute providers, and stretches from chip fabrication all the way through to data centre construction.

For an industry used to headline gigawatt numbers and eye-watering capex forecasts, the detail that matters here is not the size of the cheque. It is the shape of the financing itself, and what that shape implies for everyone else trying to fund an AI build.

The plumbing behind the price tag

Anthropic does not have a credit rating. That single fact explains most of the deal’s architecture. Rather than lending directly to a start-up with no balance sheet history, the participants have split the risk between them. Google sells the chips, in the form of custom tensor processing units it co-designs with Broadcom, and separately guarantees the data centres that house them. Broadcom commits to buying back unsold capacity and stands behind a deficiency guarantee covering roughly $31bn of senior debt, meaning that if Anthropic ever stopped paying and the hardware resold for less than the loan, Broadcom would cover the shortfall.

Apollo and Blackstone supply much of the private credit used to buy the chips in the first place, through a special purpose vehicle that then leases the hardware back to Anthropic. Morgan Stanley advises, arranges and lends into the structure at several points at once.

The first tranche of this machinery, a $35bn SPV agreed in June and covered at the time by Capacity, was already the largest private credit transaction on record. It priced at around 5.75%, a rate only achievable because Broadcom’s guarantee pushed the paper into investment grade territory, opening it up to insurers and pension funds that would never touch an unrated start-up’s debt directly.

Vendor financing of this kind is not new; Lucent and Nortel ran similar schemes for telecoms kit in the late 1990s, with painful losses when customers folded. What is new is the scale, and the fact that the guarantees have been engineered so the exposure sits off almost everybody’s balance sheet at once, including Anthropic’s, which is useful for a company widely expected to eventually list.

Why the structure is a template, not a one-off

The reason this matters beyond one company’s chip order is that Alphabet owns roughly 14% of Anthropic, making Google simultaneously supplier, guarantor and shareholder. That triple role is unusual, but the underlying mechanism, hardware held off-balance-sheet in an SPV and repaid through lease income, is already being copied.

Broadcom has said its own financing platform could support more than 20 gigawatts of compute through 2028 for large language model developers including Anthropic and OpenAI, and rival structures are already forming elsewhere: OpenAI has been in talks with Nvidia over a reported $250bn backstop for a data centre campus in Ohio.

Capacity’s earlier reporting on the initial $35bn tranche flagged exactly this direction of travel. As one infrastructure analyst tracking the deal put it, “Private equity firms have become a vital source of funding for AI companies” facing chronic hardware shortages.

Fluidstack, the operator hosting the first gigawatt of capacity under the deal, was similarly direct about what the arrangement unlocks. “We’re proud to partner with frontier AI leaders like Anthropic,” said co-founder and chief executive Gary Wu, framing the financing as infrastructure delivery rather than a straightforward equipment sale.

Google’s earlier cloud partnership with Anthropic, covered by Capacity here, already gave the company access to up to a million TPUs. Thomas Kurian, chief executive of Google Cloud, described that expansion as reflecting “the strong price-performance and efficiency” Anthropic’s own teams had seen from the chips over several years. The $150bn programme is what happens when that kind of technical preference gets translated into balance sheet engineering at hyperscale.

What happens next

Two things stand out for anyone covering or operating in this market. The first is that the guarantee structure is not free. S&P downgraded Broadcom over the leverage the arrangement implies, a reminder that vendor backing carries a cost even for a company as large as Broadcom, and that cost will shape how aggressively other chip suppliers are willing to follow suit. Broadcom’s parallel long-term TPU supply agreement with Google, reported separately by Capacity, shows how tightly the commercial and financial relationships are now intertwined.

The second is geographic. No European laboratory can currently raise money on comparable terms, largely because none has the revenue base to support it or a domestic chip designer able to guarantee the debt. Europe’s difficulty is not a shortage of ambition so much as a shortage of this specific structural plumbing, and the EU Chips Act, focused on fabrication capacity rather than financing mechanics, does not fill that gap. That leaves an obvious opening for European private credit houses, sovereign funds and utilities willing to underwrite similar arrangements for regional AI infrastructure, and for data centre operators positioning themselves as the hosting layer for whoever moves first.

The practical takeaway is that procurement, credit and equity decisions are converging into a single negotiation. Anthropic’s parallel financing for its Texas campus, its Microsoft and Nvidia tie-up covered by Capacity, and this $150bn Google programme are variations on the same theme: hardware, credit and equity bundled together because none of the participants can afford to solve just one piece alone.

Expect more of these hybrid structures to surface through the rest of 2026, and expect the operators who understand the financing as well as the megawatts to be the ones winning the next round of deals.

Google Cloud’ Yasmeen Ahmad: The ‘dashboard fallacy’ is holding back enterprise AI agents

AI agents are starting to talk: Are the networks ready?