AI Compute
The competition in computing infrastructure has entered deep waters, and traditional procurement contracts can no longer satisfy the sense of security required by tech giants.
Recently, Google presented chip giant Marvell with a seven-year capacity and R&D wager through a stock purchase warrant agreement valued at up to $12.2 billion. A tranche of shares can be unlocked only for every $500 million in custom product revenue generated for Marvell.
I. What Does Google See in Marvell?
Looking at the commercial cooperation agreement signed by both parties, Google is not targeting a single, breakout AI compute chip, but rather the peripheral infrastructure of the entire TPU (Tensor Processing Unit) ecosystem.
△ Image source: Google
The joint R&D scope defined by both parties is highly specific, covering five core areas: AI inference accelerators, storage controllers, network interface controllers (NICs), memory interface controllers, and near-memory computing products.
As large models shift from large-scale training to massive applications, the performance bottleneck of computing clusters is no longer purely the peak compute power of chips, but rather lies in data transmission and storage throughput efficiency. In a computing network composed of tens of thousands of accelerator cards, communication latency and the "memory wall" will directly drag down the overall system's operational efficiency.
To break away from over-reliance on general-purpose GPUs during the inference stage, Google must significantly improve the overall collaboration efficiency of its TPU clusters. Network interface controllers and storage controllers determine the speed at which data shuttles between racks, while near-memory computing technology is used to break through the storage bandwidth limitations in traditional architectures. Marvell's years of accumulation in high-speed interconnects and data transmission IP perfectly fills Google's shortcomings in data center rack-level hardware.
Behind this cooperation also lies Google's strategic considerations for balancing its supply chain.
In April 2026, Broadcom and Google renewed a long-term cooperation agreement, committing to jointly develop next-generation custom AI chips and next-generation AI rack components, extending the partnership all the way to 2031. In the current custom AI chip market, Broadcom holds an extremely large market share and is in an absolutely dominant position.
However, for a cloud provider of Google's caliber, betting entirely on a single supplier for underlying core hardware carries too much risk. Bringing in Marvell as a second core camp and diverting the development of peripheral high-speed interconnects and specific inference products not only makes Google's computing supply chain more resilient but also gives it more initiative in core R&D scheduling and resource allocation.
II. Equity Unlocked Only Upon Procurement Targets
Compared to traditional hardware procurement contracts, the most noteworthy aspect of this cooperation is the stringent terms designed by both parties in the transaction structure.
According to documents disclosed by Marvell, the company issued a stock purchase warrant to Google valued at up to approximately $12.18 billion. Google has the right to purchase up to 58.97 million shares of Marvell's common stock at a price of $206.58 per share. If this warrant is ultimately fully exercised, Google will hold approximately 7% of Marvell's outstanding shares, naturally becoming its fifth-largest shareholder.
But this is by no means a "free lunch" that can be collected step by step. The entire exercise mechanism is deeply tied to the commercial implementation performance of both parties in custom chips.
Apart from the first year, where 1.36 million shares serve as the base tranche to kick off the cooperation and are unlocked in quarterly batches, the vast majority of the remaining shares are divided into 240 equal tranches, with the unlocking window extending from the third quarter of Marvell's fiscal year 2027 all the way to the end of fiscal year 2033.
The rules for realization are highly stringent: Google must generate $500 million in actual recognized revenue for Marvell within the eligible custom product lines to trigger the release of one tranche of equity.
If Google wants to cash in this multi-billion-dollar warrant, it must generate hundreds of billions of dollars in actual business turnover for Marvell over the next seven years. For Marvell, this is equivalent to directly locking in a long-term revenue floor for the future; for Google, if the project falls short of expectations or the products cannot be mass-produced and deployed, most of the warrants will remain merely paper rights, and the company will not incur significant sunk costs.
III. Buying Companies, Trading for Compute, and Betting on Options: Three Strategies for Giants to "Lock In" Chips
Looking at the evolution trajectory of the entire industry, the alliance between Google and Marvell provides a very typical business sample: cloud providers use stock options as leverage to lock in the custom R&D capacity of chip companies; chip companies, in turn, give up future equity appreciation potential in exchange for deterministic revenue during the long development cycle.
The rise of this "options for orders" model is closely related to the massive capital consumption in current computing infrastructure construction. Designing a custom AI chip based on advanced process nodes requires hundreds of millions of dollars just for non-recurring engineering (NRE) and tape-out costs. Without highly certain procurement volumes, it is difficult for major chip companies to divert core teams to do deep customization for a single customer.
Looking back at the past decade of cloud computing giants taking control of underlying hardware, the way capital plays its role has undergone three clear iterations.
The first is the "wholly-owned acquisition" model represented by Amazon.
In 2015, AWS acquired the Israeli chip startup Annapurna Labs for approximately $350 million. At a time when custom chips had not yet become an industry consensus, Amazon internalized its hardware R&D capabilities at a relatively low cost. This full ownership model paved the way for the subsequent creation of the Nitro architecture, Graviton CPUs, and Trainium accelerator cards. The logic lies in absolute control: bringing all underlying architectures and teams in-house to exchange for ultimate hardware-software integration efficiency. But the price to pay is bearing years of R&D investment and management overhead.
The second is the "capital for compute and technology binding" model represented by Microsoft and OpenAI.
By injecting massive funds and cloud compute into OpenAI, Microsoft did not pursue absolute control over the latter, but instead exchanged it for priority access to the underlying large models and exclusive binding with its cloud service platform. The core logic of this model is to use capital leverage to lock in the output of upstream core algorithms, thereby driving the consumption of its own cloud infrastructure.
Now, with the current cooperation between Google and Marvell, a third form has evolved: "performance-bound stock purchase warrants."
Today, the competitive barriers and corporate valuations in the semiconductor industry have reached an entirely different level. For a mature semiconductor giant of Marvell's massive size and rich IP reserves, a wholly-owned acquisition by cloud providers is neither realistic and would also face lengthy and extremely stringent antitrust reviews. However, relying solely on traditional commercial procurement contracts makes it difficult to ensure that suppliers pin their most elite engineering teams on custom projects for the long term against the backdrop of high R&D risks in advanced process nodes.
The "options for orders" model perfectly hits the intersection of both parties' demands.
Google does not need to pay huge amounts of cash for acquisition premiums; simply with a stock purchase warrant tied to actual business performance, it has substantively "leased" Marvell's most core R&D capabilities. For Marvell, the company's independence remains completely unaffected, and giving up a portion of the equity premium is exchanged for extremely rare, guaranteed orders over the next few years.
Over the past two years, fluctuations in AI compute demand have repeatedly impacted the entire industry chain. Capacity planning by upstream wafer fabs, R&D investment by chip design companies, and infrastructure deployment by cloud providers have all been constantly navigating the uncertainty of demand.
This $12.2 billion agreement is, in essence, a long-term commitment signed and confirmed by both parties. Google has secured supply chain resilience and priority for customized resources, while Marvell has secured predictable revenue and a safety cushion for R&D investment. Ultimately, how much of this warrant can be realized depends on the deployment progress of TPUs over the next seven years, as well as the actual trajectory of AI compute demand.
The variables in the industry have not disappeared; rather, the upstream and downstream sectors are exploring more stable modes of collaboration.
#Google #Marvell #Compute #AIChips