Learning NVIDIA: Google and Broadcom are both starting to open the "AI chip closed loop"

Author: Dong Jing; Source: Wall Street Insights

NVIDIA’s AI chip business empire, built with financial guarantees and revolving financing, is being replicated one by one by its strongest competitors. Google and Broadcom are using their respective balance sheets as weapons, borrowing NVIDIA’s “playbook” to break through the AI computing power market.

Google is launching the most direct challenge yet to NVIDIA’s dominance in AI chips, using NVIDIA’s playbook. According to a June 18 report by The Wall Street Journal, Google is copying NVIDIA’s familiar customer lock-in strategy in full: by providing financial guarantees for data center projects, and using revolving financing to leverage chip procurement, it is locking in customers—then backing the push with an $85 billion equity financing plan to aggressively compete for external computing power clients.

Meanwhile, Wall Street Insights also wrote that Broadcom is taking a similar route: teaming up with Apollo and Blackstone to set up a $35 billion AI computing power financing platform. Using its own credit to provide gap-top-up guarantees for senior bonds, it bundles chip manufacturers, private credit, and AI computing power demand into a new financing closed loop, targeting NVIDIA’s more than 90% share in the AI chip market.

The core logic behind this challenge is: when computing power scarcity becomes a decisive variable in the AI race, whoever can help clients solve their financing problem will win chip orders. Analysts point out that the market significance of the above moves is that NVIDIA’s business model—long helping data centers lower financing costs through financial guarantees, and leveraging chip procurement through revolving investments—is being systematically transplanted by Google and Broadcom.

This trend not only means the competitive landscape in the AI chip market is being reshaped, but also suggests that a deep linkage between private credit and AI infrastructure financing will become the industry’s new normal, posing a real challenge to NVIDIA’s more than 90% market share.

Google Copies NVIDIA’s Playbook: Financial Guarantees in Exchange for Chip Orders

According to reports, Google is systematically replicating NVIDIA’s core business strategy: helping data centers secure lower-cost debt financing through financial guarantees, while using “revolving financing” arrangements that allow part of the capital it invests to flow back in the form of chip procurement.

The most representative case is the Lake Mariner project on the south shore of Lake Ontario in New York State. Google provided $3.2 billion in financial guarantees for this AI data center cluster. The project was co-developed by TeraWulf and FluidStack, a cloud service provider supported by Google, and its computing power will be leased for use by AI giant Anthropic. TeraWulf co-founder and CTO Nazar Khan said: “These well-capitalized companies firmly believe that the market around computing power will create enormous value, and they don’t want to be left behind.”

Google’s financial guarantee rollout goes far beyond this. The report said insiders revealed that Google has also underwritten another Anthropic project—a $7 billion River Bend project near Baton Rouge, Louisiana—and provided an additional $1.4 billion in financial guarantees for an AI computing power leasing project in Colorado City, Texas.

At the broader strategic level, Google recently reached a $5 billion agreement with Blackstone to set up a new cloud services company, directly targeting CoreWeave and Nebius—two cloud services providers backed by NVIDIA that exclusively use NVIDIA hardware. Bernstein tech analyst Stacy Rasgon said:

“They are obviously more opportunistic than a few years ago, and more actively monetizing their own assets. But a few years ago, this opportunity simply didn’t exist. Now, all we hear is that there isn’t enough computing power.”

Direct Selling of TPUs: From Internal Tools to an External Competitive Weapon

Google’s commercialization path for its self-developed AI chip TPU (Tensor Processing Unit) involves a three-stage leap: from being used internally only, to opening it up to external customers, and then to direct sales.

According to The Wall Street Journal, everything can be traced back to 2013. Back then, Jeff Dean—an AI researcher at Google (now Chief Scientist at DeepMind)—conducted a “thought experiment” while researching speech recognition: if Google were to launch a speech model for 100 million users, the required computing power would be equivalent to twice the total number of servers Google had at the time. His conclusion was: “We need to build dedicated hardware.”

At first, TPUs were used only internally by Google to support AI functionality development for its search engine and other products. As external demand for computing power surged, Google began opening up TPUs to external customers through its Cloud platform, driving rapid growth in its cloud business. In May this year, Google further announced plans to sell TPUs directly to customers and rolled out its first TPU product specifically customized for inference scenarios, which is expected to directly compete with NVIDIA’s new Groq 3 LPU.

Mark Lohmeyer, Vice President of Google Cloud AI and Computing Infrastructure, said that inference-optimized chips, together with Google’s improvements in cross-system chip collaboration, have attracted new customer groups that previously had not considered TPUs. These include Citadel Securities, which has long used Google Cloud—recently, the company has started using TPUs for some research software workloads. CTO Josh Woods said that the operating cost for key workloads fell by 30%, with speeds up to four times faster.

Broadcom Bets on “Gap Top-Up”: Trading Credit for Market Share

Meanwhile, Wall Street Insights previously also wrote that Broadcom is using its own credit at a cost to gain market share in the AI chip space—creating a new financing model that bundles chip manufacturers, private credit, and AI computing power demand.

Broadcom, Apollo, and Blackstone jointly announced last week the establishment of the “AI XPV Platform,” with an initial deal size of $35 billion. The financing is intended to expand Anthropic’s AI computing power infrastructure of over 1 gigawatt, and it is one of the largest private credit special purpose vehicle (SPV) transactions to date. The core vehicle is an SPV set up with Apollo’s Atlas SP Partners taking the lead: it buys the chips and leases them to Anthropic, with rental payments serving as the source of debt repayment.

The debt structure is divided into three layers: $600 million A1 notes sold to banks at a 100-basis-point premium over Treasury rates; $24 billion A2 notes sold to institutional investors at a 5.75% yield; and $4.5 billion subordinated notes not backed by Broadcom, with a yield as high as 8.5%. In addition, Atlas SP Partners also provides an $800 million equity layer. The key to enabling low-cost financing for senior notes lies in the “gap top-up” guarantee provided by Broadcom: if Anthropic fails to perform and the proceeds from disposing of the chips are insufficient to cover principal and interest, Broadcom will make up the losses for A1 and A2 investors.

As early as this March, Broadcom CEO Hock E. Tan was cautious about using Broadcom’s balance sheet to provide such guarantees, but later changed his stance. The real pressure driving the change was that NVIDIA has accelerated chip sales using similar supplier financing methods; if Broadcom did not follow suit, it risked falling behind in the AI chip competition. Tan positioned this cooperation as “the first of many future deals,” and plans to provide more than 20 gigawatts of AI compute financing for leading AI labs via this platform by 2028—potentially up to $700 billion in chip procurement.

NVIDIA’s Moat: The CUDA Ecosystem and the “Jensen Prison”

Despite the pressure from Google and Broadcom, NVIDIA’s market position remains fairly resilient. Behind its more than 90% share in the AI chip market is a strong ecosystem built from plug-and-play interconnected hardware and easy-to-use CUDA programming libraries.

Reports say some emerging cloud service providers are worried that once they deviate from NVIDIA’s full hardware stack, they may face the risk of losing chip allocations—a dilemma insiders in the industry call the “Jensen Prison.” Adam Fisher, a partner at Bessemer Venture Partners, said:

“Not all NVIDIA cloud providers will say this. Some will say NVIDIA gave them everything they need, but there are also some that are urgently looking for other choices yet can’t get them from other suppliers.”

On the challenge from competitors, NVIDIA CEO Jensen Huang has publicly expressed calm confidence. In an April podcast episode, he said NVIDIA has a significant lead over Google and other ASIC chip manufacturers, and questioned the cost advantage of TPUs: “I’d love to hear them prove the cost advantage of TPUs—in my view, it makes no sense.” He also emphasized that Anthropic is the only important external customer for Google’s TPU.

However, Amin Vahdat, Google’s Chief Technology Officer for AI infrastructure, takes a different view. He said he is not focused on confronting NVIDIA or any other competitor—because NVIDIA is both a competitor and an important partner, given that Google data centers also use NVIDIA GPUs. “For me and for us, this isn’t a zero-sum game—the market demand is big enough.”

Trillion-Dollar Capital Expenditure Breeds a New Financing Landscape

The above moves by Google and Broadcom reflect the broader industry backdrop of a sharp surge in AI infrastructure financing demand.

According to a report, Morgan Stanley expects that capital market financing in the U.S. AI sector will reach $400 billion, and could exceed $1 trillion by 2028, to match estimated capital expenditure needs of about $1.8 trillion over the next two years. Traditional banks have already shown clear strain in absorbing large volumes of AI-related debt, which is why private credit has become an important alternative channel.

Google announced this month a plan to raise $85 billion in equity financing, mainly to support AI infrastructure construction. Recent comparable cases include:

Meta completed a $27.3 billion SPV transaction for the Louisiana Hyperion data center, with private credit provided by Blue Owl and arranged by Morgan Stanley, while Meta provided quasi-guarantee support; Amazon issued about CAD 14 billion (roughly $10 billion) in the Canadian bond market, setting a record for the largest single issuance in the CAD bond market.

For Anthropic, this arrangement is the clearest signal that it is shifting toward building its own computing power supply and reducing reliance on cloud service providers such as Google or Amazon. Wall Street Insights noted that, according to The Information, Anthropic has arranged for Google to provide backing for lease agreements across five of its data center facilities, ensuring physical space support for chip installation.

Together, these arrangements show that in the race for AI computing power, the bundling of interests among chip manufacturers, tech giants, and private capital is advancing with unprecedented depth and speed—and the “financial guarantees for market share” model pioneered by NVIDIA has become a new paradigm that the entire industry is rushing to emulate.