FT披露Google为Anthropic组建约2000亿美元TPU和数据中心融资网络

Inside Google’s $200bn Wall Street finance machine for Anthropic
Google has assembled one of the largest infrastructure financing programmes in history to sell more than $150bn of AI chips destined for Anthropic.
The effort brings together Google, Broadcom, Apollo, Blackstone, Morgan Stanley and a slew of crypto miners in a web of transactions that stretches from chip manufacturing to data centre development, according to people involved in the project and corporate filings reviewed by the FT.
The arrangements, which tie together contracts worth about $200bn, have not previously been reported in full. They reveal how Big Tech is creating funding models for AI that reach far beyond traditional corporate spending and tie more of the financial system to the industry’s growth.
At the centre of the project are Google’s tensor processing units, or TPUs — AI chips it has co-developed with Broadcom since 2016. Once used largely inside Google’s own data centres, the chips have begun to be sold externally, challenging Nvidia’s dominance of the AI processor market. Google sells them in “pods” — server racks taller than a person that connect thousands of chips into a single computing system.
“AI chips are some of the most valuable goods ever produced,” said a senior banker close to the deals. “It’s a scale like we’ve never seen before because it’s a product we’ve never seen before.”
Surging demand from Anthropic has required a sprawling financing operation to support sales to a start-up without a credit rating. To support that expansion, Google, Broadcom and Wall Street investors have each taken on different pieces of the financial risk.
Google, which is also an investor in Anthropic, guarantees the data centres. Broadcom commits to buying the chips and helps finance them. Apollo and Blackstone provide much of the private-credit capital that purchases the hardware before leasing it to Anthropic.
“This is each of us putting our balance sheet to work,” said a Google executive involved in the effort. “We’re doing it on the data centre side, [Broadcom’s] doing it on the chip side.”
The web of contracts underpinning these arrangements adds up to about $200bn, with roughly four-fifths tied to the chips themselves, making it one of the largest infrastructure financings ever assembled.
A programme of that size posed a problem: none of the companies involved wanted to carry tens of billions of dollars of AI chips on their balance sheets.
Google, which is already funding record capital spending, would prefer not to add them. So Broadcom agreed to buy the AI hardware from it. But Broadcom, too, would rather deploy its capital elsewhere.
“Broadcom is the financing provider, but they don’t want to be in the financing business,” said one person familiar with the effort.
That challenge produced an unusual solution. Morgan Stanley helped arrange a private-credit vehicle, funded by outside investors, that buys the chips and leases them to Anthropic in an adaptation of the vendor-financing model Boeing and GE built to sell aircraft and engines.
In June, the first tranche of TPU hardware passed from Google through Broadcom into this financing blender. A special-purpose vehicle known as Compute SPV paid $35bn for roughly 1GW of the AI hardware, representing around 1mn TPUs, according to people familiar with the matter.
The SPV’s cash came from three tranches of debt anchored by Apollo and Blackstone. Broadcom, in effect, guaranteed the two senior tranches by agreeing to cover any shortfall if Anthropic stopped paying and the SPV could not sell the hardware for enough to make the senior investors whole.
The arrangement, known as residual value support, covers about $30bn of the $35bn financing, with Broadcom’s exposure declining as Anthropic makes its lease payments.
That structure is the template that is expected to finance TPUs worth hundreds of billions more. The largest deal yet came together in April when Google agreed to sell another 3.5GW of TPU hardware to Broadcom for deployment by Anthropic.
Broadcom’s filings disclose $128bn of purchase commitments, with $55.2bn of deliveries scheduled for its 2027 fiscal year and $72.9bn for 2028. People familiar with the arrangements said those commitments relate almost entirely to its agreed purchase of 3.5GW of TPU hardware from Google.
Financing the chips solved only half of Google’s problem. The company also needed enough powered data centres to house them. “We have a schedule and we’re looking for capacity that will fit the schedule,” the Google executive said. “Crypto miners with excess capacity were helpful.”
It has helped transform several crypto miners with secured power into a new breed of AI infrastructure developers, with a small outfit called TeraWulf the first to land a Google backstop to add a 360MW data centre on its campus in upstate New York.
Google guaranteed the lease payments on the Anthropic-bound site, which Morgan Stanley packaged into a construction bond that in October raised $3.2bn to get it built. In return for the backstop, Google took penny warrants that gave it a stake in TeraWulf.
Google’s team rapidly replicated this approach in the following months, helping crypto miners like Cipher Digital and Hut 8 build data centres in Texas and Louisiana. The FT identified five projects with 1.4GW of power, which have raised $15bn of debt with the support of Google’s backstop.
People familiar with the matter said the Big Tech company had so far backstopped 10 developments with 2.4GW of power for TPUs. Google’s guarantees put it on the hook for as much as $44bn if all the leases go bad, though it marks the liability at $815mn on its balance sheet. It could also step into the leases itself.
The Google team is now racing to put together additional data centre projects with enough power to ultimately house all of the 4.5GW of TPU hardware they’ve agreed to sell. “We’re spending a lot of time on [power] right now — all of our time,” said the Google executive.
By this spring, Google’s financial heft had already begun to reshape the economics of AI infrastructure. Data centre projects backed by Google borrowed at a median rate of 7.1 per cent, compared with 9.3 per cent for the neocloud operators building around Nvidia’s chips. Jefferies analysts called this gap “a structural cost-of-capital disadvantage” for those in Nvidia’s orbit.
For Google’s project, the risk is big if concentrated: $200bn of contracts tied to Anthropic’s ability to pay its chip and data centre leases. It is one piece of a larger dilemma, with demand across the industry resting on a handful of large hyperscalers and frontier AI labs.
“There’s a whole world that’s been built underneath those companies, and if their appetite to invest decreases, all of it sees a slowdown,” said Jefferies analyst Jonathan Petersen. “That’s the big macro risk.”