Google has obtained the right to purchase up to 58.97 million shares of chipmaker Marvell Technology as part of an expanded agreement focused on the development of AI processors. This information was disclosed in a report filed with the U.S. Securities and Exchange Commission (SEC) under Form 8-K.
This agreement exemplifies the increasing complexity of financial relationships within the AI sector. Chip manufacturers, model developers, and hyperscalers are increasingly engaging in arrangements that go beyond mere technology sales, involving guarantees, stock options, and joint capital raising to build computing infrastructure.
Option Linked to Purchases
The collaboration between Google and Marvell encompasses several components of Google's TPU infrastructure, which are proprietary AI accelerators. This includes processors for inference, data storage controllers, networking solutions, memory interfaces, and compute-in-memory technologies.
Of the 58.97 million shares, only approximately 1.36 million will be unlocked in equal parts over the first year following the agreement's signing.
The remainder of the warrant is directly tied to Google’s order volume. According to the report, the shares are divided into 240 tranches, with each tranche becoming available after Google generates an additional $500 million in revenue for Marvell under the agreement.
Thus, to fully unlock the option, the total sales volume must reach $120 billion.
Google can purchase the shares at a price of $206.58 until August 18, 2033. If the warrant is fully exercised, the company could become the fifth largest shareholder in Marvell, as noted by Reuters.
Following the announcement of these terms, Marvell's stock saw an increase, while shares of its larger competitor Broadcom fell by over 5%.
Data Source: Yahoo Finance.“This is a significant win for Marvell,” stated Morningstar analyst William Kerwin.
However, the expert does not view this agreement as indicative of Broadcom being pushed out of Google's supply chain. He believes that the hyperscaler is simply expanding its supplier base in response to growing demands for AI chips.
Google Reducing Dependence on Nvidia
The demand for specialized processors is rising as major tech companies seek to lower the costs of AI computing and reduce their reliance on Nvidia’s accelerators.
For years, Google has been developing its own TPUs optimized for training and running AI models. Marvell's role will not only involve creating the accelerators themselves but also developing components necessary for integrating numerous chips into computing systems.
According to Reuters, the bolstered emphasis on proprietary infrastructure coincides with a restructuring of Google’s AI division, where leaders affiliated with Google Cloud have gained more influence. For Marvell, the agreement not only presents a potentially large customer but also ties the warrant to spending, making the hyperscaler financially invested in the supplier’s growth.
Such arrangements are becoming increasingly common in the AI industry. In October, AMD entered into a deal with OpenAI that also combined significant processor purchases with an opportunity for the ChatGPT developer to acquire a substantial stake in the chip manufacturer.
Manufacturers Financing Their Own Demand
The intertwining of interests is even more evident in the relationship between Nvidia and OpenAI. On August 17, Nvidia agreed to provide a guarantee of up to $105 billion for OpenAI’s data center in Ohio, a project being executed by SB Energy, a company owned by SoftBank.
If OpenAI fails to meet its obligations, Nvidia will have to compensate for the difference between the guaranteed amount and what the owner can recoup through re-leasing or selling the infrastructure.
Nvidia is also investing $1.5 billion directly in SB Energy and will serve as the exclusive supplier of computing infrastructure for the facility.
“We are securing long-term infrastructure for Nvidia computing,” explained CEO Jensen Huang.
The first 800 MW of approximately 8 GW are slated to come online in 2028, with OpenAI having signed a 20-year lease agreement.
This arrangement creates a unique situation for Nvidia, as it finances the construction of a data center that subsequently becomes a major buyer of its own accelerators. Huang dismissed claims that this represents circular financing.
However, Reuters noted that such agreements raise investor concerns about the interdependence of participants in the AI market.
Wall Street Engaging in AI Infrastructure
The scale of necessary investments has already far exceeded the capacity of individual AI startups. According to Reuters, total spending by major tech companies on artificial intelligence is projected to surpass $730 billion this year.
In this context, Nvidia, alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, is creating platforms to attract over $500 billion in external capital for computing infrastructure.
The chip manufacturer is willing to provide guarantees of up to $125 billion under certain conditions, which would account for a quarter of the potential financing. Huang referred to AI chips as "revenue-generating assets": they can be installed in data centers, transferred among operators, and used to secure debt financing.
Private equity and credit funds have begun to treat computing infrastructure as a distinct asset class. For instance, Apollo and Blackstone are participating in financing the expansion of Anthropic by approximately $35 billion using Broadcom equipment. Analysts at Bank of America suggested that the financing program related to the latter could grow to $370 billion in senior debt by mid-2029.
AI Has Transformed from Software to Capital-Intensive Industry
This shift was highlighted on August 20 by Amit Joshi, an AI and strategy professor at IMD Business School, in a column for Fortune. He argues that the two-decade-long model of tech business is reversing: investors have long valued software companies for their low capital requirements and high margins, while in AI, the main competitive factors have become data centers, chips, energy, and the ability to finance their construction.
Joshi connects this trend to the gradual convergence of leading models:
“Competitive advantage is increasingly less about the model itself and more about who can finance, build, and operate infrastructure more cost-effectively.”
However, new deals indicate that the financial aspect is indeed becoming more intricate, with blurred lines between suppliers, buyers, investors, and lenders. Such arrangements allow for faster infrastructure development than if each participant relied solely on their own balance sheet. At the same time, they increase the interdependence of companies on one another.
In late June, the Bank for International Settlements (BIS) warned of macro-financial risks associated with the AI boom. Among the potential issues highlighted were rising debt financing, private credit, and opaque agreements among hyperscalers, chip manufacturers, and model developers.
The BIS particularly emphasized the risks of circular financing:
“The terms of such deals are often poorly disclosed, creating the risk of re-pledging the same asset.”
It is worth noting that in August, Huawei's chief semiconductor scientist Liao Heng cautioned that large AI chips from Nvidia and other manufacturers are approaching the physical limits of scaling.