• Google expanded its Marvell AI chip deal, tied to up to $120 billion in revenue.
• Big Tech is building custom AI chips alongside Nvidia and AMD hardware.

Google is expanding Marvell Technology’s role in its custom AI chip supply chain under an agreement covering several parts of its tensor processing unit infrastructure. Marvell described the agreement as an expanded partnership in a regulatory filing dated August 19.

The companies signed the commercial agreement on July 29. It covers custom semiconductor products that connect to Google’s TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing products.

Marvell’s work therefore extends beyond AI accelerators to silicon used across storage, networking, and memory infrastructure. Near-memory computing places some processing closer to memory, reducing the amount of data that needs to move between memory and processors.

The agreement does not identify which current or future Google TPU generations will use individual Marvell products. Marvell’s filing only states that the custom silicon programs will attach to the broader TPU ecosystem.

Marvell issued Google a warrant allowing it to purchase up to 58.97 million Marvell shares at $206.58 each. Exercising the warrant in full at that price would require roughly $12.18 billion.

Most of the potential equity position is tied to purchases Google makes from Marvell. Of the shares covered by the warrant, 1.36 million vest in equal quarterly instalments during the first year.

The remaining shares will vest according to purchases made by Google and its affiliates between Marvell’s fiscal third quarter of 2027 and the end of fiscal 2033. Marvell divided those shares into 240 equal tranches, with one tranche vesting for every $500 million in revenue from the custom products.

Full vesting across the 240 purchase-linked tranches would therefore correspond to $120 billion in custom-product revenue. The warrant remains exercisable until August 18, 2033, subject to its vesting conditions.

Marvell shares rose nearly 8% following the announcement, while Alphabet shares were little changed. Broadcom shares fell more than 5%, according to Reuters.

Google keeps multiple chip suppliers in play

Broadcom remains another major supplier in Google’s custom AI hardware strategy. The two companies entered a separate long-term agreement earlier this year covering the development and supply of custom TPUs for future generations of Google’s processors.

Broadcom’s April regulatory filing also says it will supply networking and other components for Google’s next-generation AI racks. The disclosed arrangements can extend through 2031, overlapping with the period covered by the Marvell agreement.

The Marvell and Broadcom filings do not specify how individual TPU programs or order volumes will be divided between the suppliers.

Google’s TPUs are custom processors built for machine-learning workloads and are used in its own services and through Google Cloud. The chips support training as well as inference, which involves running trained models to produce outputs.

Google’s eighth-generation TPU systems also provide context for the networking, storage, and memory technologies covered by the Marvell deal. TPU 8t is designed primarily for large-scale training, while TPU 8i targets inference and reinforcement-learning workloads.

TPU 8t can scale to 9,600 chips in a single superpod. Google describes the system as part of its AI Hypercomputer architecture, which combines accelerators with networking, storage, and software.

TPU 8t also uses TPUDirect RDMA and TPUDirect Storage to move data more directly between TPU memory, network interfaces, and storage, reducing reliance on the host CPU for some transfers. Marvell’s agreement covers several of those surrounding infrastructure categories, although neither company has identified which TPU generations will use the resulting products.

Cloud providers expand custom chip portfolios

Google continues to deploy Nvidia hardware alongside its own processors. Google Cloud says Nvidia GPUs remain part of its accelerator portfolio and plans to offer systems based on Nvidia’s Vera Rubin platform alongside its TPUs.

Google is also designing its own processors around different workloads. TPU 8t targets large-scale training, while TPU 8i focuses on inference and reinforcement learning; Google says TPU 8i provides 80% better inference performance per dollar than the previous TPU generation.

AWS, Microsoft, and Meta are also developing their own AI accelerators while continuing to use processors supplied by other chipmakers. AWS has developed Trainium, Microsoft operates its Maia processors, and Meta is expanding its MTIA family.

AWS says EC2 instances using Trainium2 provide 30% to 40% better price-performance than its GPU-based P5e and P5en instances for generative AI training and inference. The comparison is specific to AWS’s own cloud infrastructure.

Amazon CEO Andy Jassy has also linked Trainium to infrastructure costs. He said AWS expects the chips, at scale, to reduce annual capital expenditure by tens of billions of dollars and provide several hundred basis points of operating-margin benefit compared with relying on third-party processors for inference.

Microsoft developed Maia 200 specifically for inference. Built on TSMC’s 3-nanometre process with 216GB of HBM3e memory, the accelerator delivers more than 30% better performance per dollar than the latest-generation hardware previously running in Microsoft’s fleet, according to the company.

Microsoft continues to operate Maia alongside Nvidia and AMD processors. CEO Satya Nadella said Microsoft wants access to multiple hardware options and processor generations as it manages performance, supply, and total cost of ownership across its infrastructure.

Meta’s infrastructure also spans Nvidia GPUs, AMD GPUs, CPUs, and its own MTIA accelerators. The company said in March that hundreds of thousands of MTIA chips were already deployed for inference across content and advertising workloads.

Meta is developing further MTIA generations for ranking, recommendation, and generative AI workloads. The company says accelerator selection is based partly on performance and total cost of ownership.

Chipmakers remain part of custom silicon plans

Custom-chip programs still involve established semiconductor suppliers. Meta expanded its Broadcom partnership in April to co-develop multiple generations of MTIA processors, with Broadcom contributing to chip design, advanced packaging, and networking.

Google also works with outside semiconductor suppliers across its TPU infrastructure. Broadcom is contracted to work on future TPU generations and networking components, while Marvell’s expanded agreement covers inference accelerators and silicon linked to memory, storage, and networking.

Inference is a common target across several of these custom-chip programs. Google developed TPU 8i for inference-related workloads, Microsoft built Maia 200 specifically for inference, and Meta says hundreds of thousands of MTIA chips already handle inference across its services.

Amazon says most inference performed through Amazon Bedrock runs on Trainium. Marvell’s agreement with Google similarly includes custom AI inference accelerators, alongside silicon for the memory, networking, and storage infrastructure around Google’s TPU systems.

Marvell’s warrant also has parallels in other large AI hardware agreements. AMD issued OpenAI a warrant in October 2025 covering up to 160 million AMD shares alongside an agreement for the deployment of up to six gigawatts of AMD GPUs.

Those warrant shares vest in stages tied partly to OpenAI’s purchases of AMD Instinct GPUs. The first tranche is linked to delivery of the first gigawatt of MI450-series GPUs, while full vesting depends on purchases reaching six gigawatts, alongside AMD share-price and other performance conditions.

AMD entered a similar arrangement with Meta in February 2026. Meta received a warrant covering up to 160 million AMD shares, with vesting tied to GPU purchases and other conditions.

Google’s Marvell agreement uses a different structure, linking most warrant vesting to revenue generated from custom semiconductor purchases rather than gigawatts of accelerator deployments.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.