Omdia’s latest research highlights the rapid growth in demand for Google’s Tensor Processing Unit (TPU) AI chips, a trend that may be strong enough to start chipping away at NVIDIA’s market dominance in GPUs. 3Q results from Broadcom, whose Semiconductor Solutions division acts as a custom chip outsourcing partner for Google, Meta, and several other AI players, give our analysts some insight into purchasing trends and information that is typically kept under wraps – how many custom processors is Google buying?
Broadcom’s CEO Hock Tan has repeatedly revised up his target for AI semiconductor revenue, aiming for $12bn for this year. Based on this, it is estimated that Google’s TPUs could account for somewhere between $6bn – close to Omdia’s existing estimates – and $9bn, depending on the breakdown between compute and networking devices. This figure includes a substantial number of Meta’s MTIA chips, with a project for a mysterious third client expected to ramp up in 2025.
Alexander Harrowell, Principal Analyst at Omdia noted: “Even though there is some uncertainty in the exact ratio between compute and networking devices, shipments of TPUs, even at the lower $6bn estimate, are growing at a pace fast enough to take share from NVIDIA for the first time. It’s worth noting that Google’s Cloud Platform business continues to grow as a share of Google’s revenue while improving profitability. This could indicate that TPU-accelerated instance types or AI products running on TPUs are driving the growth at Google Cloud, especially as accelerator instance types are high-margin products.”
