Broadcom aims to sell 1 million 3D-stacked AI chips by 2027, reshaping data center market

Broadcom said it expects to sell at least 1 million of its new 3D-stacked AI chips by 2027, a production goal that signals a rapid push to make advanced stacked-die accelerators mainstream in enterprise data centers. The company, a major designer of semiconductors and networking equipment, is positioning the parts as a step up in performance density for customers building large-scale AI services.

Broadcom executives, including Harish Bharadwaj, vice president of product marketing, have outlined the sales target as part of a broader effort to capture share in the fast-growing market for AI accelerators. The company’s announcement underscores a shift in hardware design toward vertically integrated chip stacks that place compute layers and high-bandwidth memory in close physical proximity to boost throughput and reduce energy per operation.

Three-dimensional stacked-die designs use advanced packaging to stack multiple silicon dies and link them with dense vertical interconnects. That architecture typically increases memory bandwidth, shortens signal paths, and raises compute density compared with conventionally packaged chips. For large AI models and inference workloads that depend on sustained memory throughput, those technical advantages can translate into lower latency and higher effective performance per rack of servers.

Delivering 1 million units of this class of accelerator within roughly two years will require upgrades across Broadcom’s supply chain. Manufacture of stacked chips depends on advanced packaging services and high-yield foundry processes; volume production also tightens demand for testing, assembly and thermal management solutions. If Broadcom meets its target, suppliers of advanced packaging and system integrators are likely to see a notable lift in orders and hiring.

The move will also intensify competition with incumbent accelerator suppliers. Leading GPU and accelerator makers have steered significant engineering resources into memory and interconnect optimizations; Broadcom’s sales target signals it expects customers, particularly large cloud providers and hyperscalers, to consider alternatives based on 3D stacking. For data center operators, the arrival of high-volume stacked accelerators could create leverage to renegotiate procurement and push for denser server configurations.

There are trade-offs. Stacked-die chips pose thermal and reliability challenges, and their benefits vary by workload. They can reduce power per operation in memory-bound AI tasks, but they require system-level redesigns for cooling and power delivery. Broadcom will need to demonstrate consistent yields, software ecosystem compatibility and clear total-cost-of-ownership advantages to persuade broad adoption beyond early deployments.

Broader implications include supply-chain geopolitics and environmental impacts. Rapid scaling of advanced packaging concentrates production at a small set of specialist suppliers and increases demand for energy-intensive manufacturing steps. At the same time, higher-performing, more efficient accelerators can reduce overall energy use per model run, making the net environmental outcome dependent on deployment patterns.

Achieving the 1 million-unit target will be a test of both market appetite and industrial scale. If Broadcom hits that milestone by 2027, it would mark a turning point in which 3D-stacked accelerators move from experimental offerings into a standard option for large-scale AI infrastructure, reshaping how companies build and power next-generation machine learning services.