Google TPU Orders Diversified: Marvell, Broadcom, and MediaTek Compete for Share
Claire Weston
Broadcom openly admits it cannot handle all of Google's TPU orders alone — the cloud AI custom-chip market has officially entered a multi-vendor era, giving MediaTek and Marvell a real way in and shifting the supply-chain power balance.
Why did Broadcom suddenly admit it can't do this alone?
Google's TPU business has grown beyond what Broadcom alone can supply. This is not modesty — it is a capacity ceiling.
This means → even as the largest supplier, Broadcom's fabs and supply chain can no longer keep pace with Google's expansion on their own.
Broadcom CEO Hock Tan called MediaTek and Marvell "ankle biters," but analysts read this as a careful framing of ASIC market scale, not dismissal — Google is already splitting orders in practice.
What have MediaTek and Marvell actually won?
MediaTek has secured Google TPU orders, investing heavily to ensure its SerDes — high-speed data transceiver interfaces — and other I/O technology meet Google's bar.
Google's next-generation product uses Intel's EMIB packaging — an advanced method that tiles different chips onto a single substrate. MediaTek must independently coordinate multiple parties for joint development and testing.
Marvell is building a partnership with Google, still at an early stage, but the direction is clear.
Why is this business so hard to break into?
Cloud AI ASIC work — designing custom chips for a specific hyperscaler — is far harder than outsiders expect: high technical barriers plus extreme dependence on advanced-node capacity and a stable supply chain.
IC design firms typically embed core engineers at the client site for technical integration. This means → it drains internal resources and risks top talent being poached by the client.
In plain terms = you send your best people to help the customer, and the customer may keep them. That is an open secret in the industry.
Why does Google need more than one supplier?
Two core drivers: spreading supply risk + securing enough chips for TPU deployment at scale.
Relying on a single ASIC vendor means that vendor's capacity bottleneck directly caps Google's expansion pace.
This reflects a deeper signal: once AI compute demand hits a certain scale, no single chip company can shoulder an entire hyperscaler's needs alone.
What decides who wins this race?
Public statements from both Broadcom and Marvell show the motive goes beyond revenue — it is about locking in a deep tie with a cloud giant and claiming strategic ground in the future cloud AI market.
Whether each vendor can keep up with Google's expansion on both capacity and technical integration will determine the final landscape.
In plain terms = whoever solves "can build it" and "can build enough" at the same time stays at the table.
Content is for reference only, not financial advice.