Hyperscalers Diversifying

The biggest cloud companies are trying to loosen Nvidia’s grip on artificial intelligence computing by testing more chip options inside their data centers. Google already sells both Nvidia graphics processing units and its own Tensor Processing Units through Google Cloud, while Marvell has been pitching custom AI silicon to hyperscalers. (cloud.google.com) (marvell.com) A graphics processing unit, or GPU, is a general-purpose parallel chip that can train and run AI models across many customers. A custom application-specific integrated circuit, or ASIC, is built for one buyer’s workload, the way a private-label engine is built for one car platform. (cloud.google.com) (marvell.com) Google has spent years building its own AI chips alongside Nvidia systems. Google said its first Tensor Processing Unit was deployed internally in 2015, and in November 2025 it introduced Ironwood, which it said delivers 10 times the peak performance of TPU v5p and more than four times better performance per chip than TPU v6e, known as Trillium. (cloud.google.com 1) (cloud.google.com 2) The spending behind that search is enormous. Bloomberg reported on February 6, 2026, that Amazon, Google, Meta and Microsoft together forecast about $650 billion in capital expenditures this year, much of it tied to new data centers and the equipment inside them. (bloomberg.com) Nvidia still dominates the market those buyers are trying to diversify away from. Nvidia reported on February 25, 2026, that fourth-quarter data center revenue reached $62.3 billion, up 75% from a year earlier, and full-year revenue hit $215.9 billion. (investor.nvidia.com) That concentration has pushed suppliers outside Nvidia to widen their pitch from chips to whole AI systems. Marvell said at its June 17, 2025 custom AI investor event that it sees a growing market for custom silicon, and in May 2025 it announced a tie-up with Nvidia’s NVLink Fusion program so customers could connect Marvell-designed custom processors into Nvidia-based AI infrastructure. (marvell.com 1) (marvell.com 2) The opening is not just in compute chips. Google’s AI Hypercomputer pitch includes networking, storage and host offload, and Marvell and Broadcom have both been promoting optical links, switches and packaging that move data between racks faster and with less power loss. (cloud.google.com) (marvell.com) (news.broadcom.com) Broadcom has been making the same case to large customers building their own accelerators. Broadcom said its 3.5D XPU packaging has been adopted by major consumer AI customers with earliest production scheduled for early 2026, and in March 2026 it said it was expanding an AI infrastructure portfolio built for “gigawatt-scale” clusters. (broadcom.com) (investors.broadcom.com) The shift does not mean GPUs are going away. Google’s cloud still offers Nvidia systems next to TPUs, and Nvidia’s own strategy now includes working with custom-chip vendors rather than blocking them from its networking stack. (cloud.google.com) (marvell.com) What changes next is likely to be the mix inside each data center, not a clean replacement. As hyperscalers spend hundreds of billions of dollars on AI infrastructure in 2026, every extra slot won by a custom accelerator, a central processing unit, a network interface card or an optical switch becomes a share shift worth chasing. (bloomberg.com) (cloud.google.com) (marvell.com)