Google’s Marvell Deal Shows Custom AI Silicon Is Becoming a Full-System Supply Chain

Related daily report

August 20, 2026: Google deepens its custom-silicon supply chain through Marvell

Today’s Daily Report records the warrant mechanics and the broader AI-safety developments. This article focuses on the infrastructure story behind the transaction: why a “custom chip” relationship now extends into memory, networking, storage and supplier incentives.

Open the August 20 Daily Report

The immediate development

The $12.2 billion headline hides the more important contract

Marvell has given Google a warrant to buy as many as 58.97 million Marvell shares at an exercise price of $206.58. Multiplying those figures produces roughly $12.2 billion, which explains the headline attached to the deal. But Google is not writing Marvell a $12.2 billion cheque today, and Marvell has not received a guaranteed $120 billion order.

The commercial structure is more conditional. Reuters reports that most of the warrant vests as qualifying Google purchases accumulate through Marvell’s fiscal 2033. Filing-based reporting gives the mechanics more precisely: about 1.36 million shares vest over the first year, while the remaining shares are divided into 240 performance tranches, with one tranche vesting for every $500 million of qualifying custom-product revenue. Full performance vesting therefore corresponds to roughly $120 billion of qualifying revenue. That number is a contractual milestone ceiling, not a forecast of sales.

The product scope is the part that deserves more attention. The relationship covers custom technology around Google’s AI infrastructure, including AI inference accelerators, storage controllers, network-interface controllers, memory-interface controllers and near-memory compute.

That list changes the meaning of the transaction. If this were simply Google hiring a second design house to help produce one accelerator, the strategic story would be supplier competition. Instead, the agreement reaches across several layers that determine whether an accelerator can actually be fed with data, connected to its peers and kept useful at data-center scale.

The immediate market reaction—Marvell up, Broadcom down—was understandable. The more durable question is not which semiconductor vendor “won” one day of trading. It is how much of an AI system a hyperscaler now needs to co-design outside its own walls.

“Custom” does not mean one company builds the system

Google’s Tensor Processing Units are routinely described as in-house chips. That description is directionally true and operationally incomplete.

A hyperscaler can own the workload knowledge, architectural priorities, compiler strategy and many of the design decisions without manufacturing every piece of silicon or developing every interface block itself. Modern custom accelerators are assembled from an ecosystem: foundry processes, packaging, high-bandwidth memory, serializer/deserializer technology, chip-to-chip links, network interfaces, power delivery, optical components, storage controllers and specialized design services.

The practical meaning of custom silicon is therefore closer to “hardware designed around a specific operator’s workloads and system architecture” than “hardware created entirely by that operator.” Competitive advantage can come from choosing which functions to own and which to source, then integrating them tightly enough that the complete machine behaves as one system.

Marvell’s own custom-platform materials make that ecosystem visible. Its portfolio spans custom compute, electrical and optical SerDes, die-to-die interconnect, advanced packaging, silicon photonics, custom HBM-related technology and PCIe-class interfaces. In March, Marvell and Nvidia expanded their partnership around NVLink Fusion, with Marvell supplying custom XPUs and compatible scale-up networking while Nvidia supplies other rack-scale components. Supposedly competing architectures can therefore be assembled from overlapping supplier relationships.

For buyers and policymakers, this makes supply-chain maps more complicated. A company may reduce its dependency on one accelerator vendor while increasing its dependency on a different collection of memory, networking, packaging or optical suppliers. “Vertical integration” is no longer a binary condition. It is a question of which layers are controlled, which are co-designed and which remain external.

Ironwood shows why the accelerator is only one part of the machine

Google’s current Ironwood TPU makes the system-level problem concrete.

According to Google Cloud documentation, one TPU7x Ironwood chip carries 192 GiB of HBM with 7,380 GB/s of HBM bandwidth. A full pod contains 9,216 chips. Each chip also has 1,200 GB/s of bidirectional inter-chip interconnect bandwidth, while Google’s optical circuit switching and data-center network connect progressively larger groups of accelerators.

Those numbers are not included as a benchmark comparison. They show why compute cannot be evaluated in isolation. A processor that can execute trillions of arithmetic operations per second is useful only if model weights, activations and intermediate state can reach the arithmetic units quickly enough. At pod scale, thousands of chips must exchange data without letting communication overhead consume the benefit of adding more compute.

Google describes Ironwood as a co-designed hardware-and-software stack and treats the TPU pod as a single supercomputer rather than a bag of accelerators. Its system combines compute, high-bandwidth memory, inter-chip networking, optical switching, the data-center network, compiler infrastructure and workload software. The architecture is particularly revealing because Google already controls the TPU architecture. Even then, performance depends on a long chain of surrounding technologies.

This is closely related to the trend AIUpdateWatch examined in AI chips are becoming packaging systems. There, the issue is bringing compute dies and HBM physically close together with adequate yield, power and cooling. The Marvell deal extends the same logic outward: once the package is built, the rest of the rack and pod still has to move data fast enough to keep that expensive silicon busy.

Marvell is selling the connective tissue around AI compute

Marvell’s recent product announcements help explain why Google might want a relationship wider than one processor family.

On August 4, Marvell introduced an AI-memory portfolio spanning server storage, CXL memory expansion and pooling, and what it calls pod-level optical shared memory. The technical premise is straightforward: inference increasingly stores large KV caches, the working memory that keeps previously processed context available while a language model generates later tokens. Long conversations, reasoning traces and agent sessions can make those caches large enough that keeping everything in expensive accelerator-attached HBM becomes difficult.

Marvell’s proposed answer is a hierarchy. Fast HBM remains close to the accelerator. CXL can expand or pool memory across servers. SSDs provide cheaper capacity farther down the stack. Its Photonic Fabric concept would add another shared tier across multiple accelerators and racks, connected optically. Marvell says that design can support up to 32 TB of warm KV-cache offload and, in its own projections, improve token throughput by two to three times in some configurations. Those throughput figures are vendor claims, not independent production benchmarks, and the relevant products are at different stages of availability.

The important point is architectural rather than promotional. Memory capacity, bandwidth and connectivity are becoming separate resources that operators try to pool and schedule more independently from raw compute. That is exactly the kind of environment in which memory controllers, network interfaces, storage controllers and optical links become strategic rather than peripheral.

Marvell has been making the same argument about networking. Its May COMPUTEX materials describe connectivity across accelerators, servers, racks and data centers as a limiting factor for the next stage of AI scaling. In the company’s telling, the opportunity is the data movement between compute elements as much as the compute element itself.

That positioning obviously serves Marvell’s commercial interests. Google’s agreement nevertheless provides separate evidence that a major hyperscaler is willing to place a broad set of those functions inside one long-term custom-silicon relationship.

The warrant turns procurement into part of the capital structure

The equity component is unusual enough to deserve separate treatment because it changes the supplier relationship without being a conventional acquisition or strategic investment.

Google receives the right to buy Marvell shares at the specified exercise price as the agreement’s vesting conditions are satisfied. That can give the customer equity upside if Marvell’s market value rises above the exercise price, while Marvell has a strong commercial incentive to win and retain large volumes of Google business. The purchasing relationship and the capital relationship become partially linked.

There is precedent. In December 2025, Marvell disclosed a warrant allowing Amazon to acquire up to roughly 1.05 million shares, with vesting tied to Amazon purchases of Marvell photonic-fabric products through 2030. Marvell’s May 2026 quarterly filing also describes customer warrants whose shares vest primarily as qualifying product revenue is generated.

The Google agreement is much larger in potential scale, but the mechanism points to a broader procurement strategy: a hyperscaler can use equity-linked milestones to align a critical supplier with a long deployment horizon.

That can be rational for both sides. Custom silicon involves substantial engineering investment before production revenue arrives. A supplier has to assign scarce design teams, validate interfaces, coordinate manufacturing and support a long roadmap. A customer wants assurance that the supplier keeps investing in that roadmap. Under this structure, the buyer’s potential equity participation expands as qualifying procurement revenue accumulates, while Marvell receives the product revenue that triggers the performance vesting.

It also creates new governance questions. Procurement concentration can become financially entangled with supplier valuation. Switching costs may rise once several generations of custom IP and operational tooling are shared. And a customer trying to diversify away from one dependency can still become deeply coupled to the additional supplier.

That does not make the structure inherently problematic. It means the economics of AI infrastructure are moving beyond spot purchases of chips toward long-duration relationships that combine architecture, capacity planning and capital incentives.

Broadcom’s existing Google agreement argues for diversification, not replacement

The sharp fall in Broadcom’s share price after the Marvell news encouraged a simple interpretation: Google is replacing Broadcom.

The public evidence does not establish that.

Broadcom disclosed in April that it had signed a long-term agreement with Google to develop and supply custom TPUs for future TPU generations, plus a separate supply-assurance agreement for networking and other components used in Google’s next-generation AI racks. That relationship can extend through 2031.

Unless one of the companies later says otherwise, the safer interpretation is that Google is widening its supplier base while preserving major existing relationships. Reuters likewise quoted Morningstar analyst William Kerwin describing the situation as a “growing pie” rather than necessarily direct displacement.

That distinction matters because demand for AI infrastructure is expanding fast enough that second sourcing can be about capacity and schedule risk as much as price negotiation. One supplier may handle one TPU generation or subsystem while another handles inference accelerators, networking, memory interfaces or future designs. The exact division of work is not public.

Diversification can still change bargaining power. Adding Marvell gives Google another design partner and potentially another route to scale custom infrastructure. But the Broadcom filing is a useful factual brake on claims that one agreement has already redrawn Google’s entire TPU supply chain.

The bottleneck is moving from arithmetic toward data movement

The larger technical trend is not that compute stops mattering. It is that adding compute exposes whichever surrounding resource is least able to scale with it.

For large inference workloads, memory can become that resource. Model weights have to be fetched. KV caches grow with context. Mixture-of-Experts models may need to move activations among specialists. Disaggregated serving can separate prompt processing from token generation, which shifts more state across the network. AIUpdateWatch’s analysis of prefill/decode disaggregation describes how a performance optimization can turn the KV cache into a networking problem.

At larger scales, networking becomes the constraint. Thousands of accelerators participating in one training or inference job have to synchronize, exchange tensors and recover from failures. Optical switching and faster electrical links are not support equipment in that environment; they influence how much of the theoretical accelerator performance can be converted into useful work.

Power and physical infrastructure impose another boundary. The most sophisticated custom stack is still limited by where it can be deployed, which is why AI data centers are increasingly following available power and why financing, grid access and hardware refresh cycles are becoming part of compute strategy.

The Google–Marvell agreement sits at the intersection of these constraints. It treats processor design, memory movement, storage and networking as parts of one purchasing relationship. That is a more informative signal about AI infrastructure competition than a single accelerator benchmark.

What the deal still does not prove

Several conclusions would go beyond the available evidence.

It does not prove that Marvell will replace Broadcom as Google’s main TPU partner. Broadcom’s April filing documents a continuing long-term TPU agreement. The allocation of future Google programs among suppliers is not public.

It does not prove $120 billion of Marvell revenue. That figure is associated with full performance vesting of the warrant. Actual purchases may be lower, spread differently across products or never reach every milestone.

It does not establish the performance of any new Google–Marvell accelerator. No independent benchmark, production efficiency figure or public system specification for the products covered by the new agreement has been released.

It does not mean Google is abandoning Nvidia. Google Cloud continues to offer Nvidia GPU infrastructure alongside TPUs, and Marvell itself is working with Nvidia through NVLink Fusion. The modern AI supply chain is overlapping rather than cleanly divided into rival camps.

It does not make Marvell’s memory or networking claims automatically transferable to Google’s systems. The company’s 2–3x token-throughput projections for Photonic Fabric are vendor figures for its own design assumptions. They illustrate the type of bottleneck Marvell is targeting; they are not evidence that Google has adopted that exact architecture or will see that performance.

Those boundaries matter because the transaction is already strategically significant without exaggerating what has been demonstrated.

What would confirm that this is a durable supply-chain shift?

The next useful evidence will come from implementation rather than more partnership language.

Watch for which Google products actually qualify under the Marvell agreement and when they enter production. A public tape-out, Cloud TPU generation, networking platform or memory subsystem tied to Marvell would clarify whether the company is becoming a second accelerator-design partner, a broader infrastructure supplier, or both.

Watch the warrant vesting. Because most of the potential equity is tied to qualifying revenue, future filings can turn a vague strategic relationship into measurable procurement evidence. That is more informative than the maximum headline value.

Watch Broadcom’s role rather than assuming its decline. Its agreement through 2031 means Google can pursue a multi-supplier architecture in which different generations and subsystems are split among partners. The competitive question is allocation, not merely presence.

Finally, watch how Google exposes its custom hardware to outside customers. The economics of proprietary accelerators change when they move from internal infrastructure into a cloud product used by other large model developers. Broader external demand can justify a deeper supplier ecosystem and make capacity assurance more valuable.

The phrase “custom AI chip” once sounded like a story about escaping the commodity processor market. It now describes almost the opposite: a hyperscaler can differentiate more aggressively at the system level while depending on an increasingly specialized network of semiconductor partners.

Google’s Marvell deal is important because it makes that network unusually visible.

Sources

Primary documentation and independent reporting

Evidence boundary: Product-performance figures attributed to Marvell are vendor claims unless stated otherwise. Google–Marvell product allocation, future purchase volumes and the exact division of TPU work among suppliers have not been publicly established.