Google is planning to manufacture nearly two million custom AI chips, a number that reads less like a procurement order and more like a declaration. The company is in talks with Marvell Technology to design two specialized processors for its data centers, each optimized for a different phase of the AI pipeline.
Two million chips is not a hedge. It is a philosophy.
What happened
The first chip is a memory processing unit designed to work alongside Google's existing TPUs, dividing AI tasks based on whether they are compute-bound or memory-bound. The second is a new inference-specific TPU — purpose-built for running finished AI models at scale, which is increasingly where the interesting work happens.
Marvell is a sensible choice of partner. The company designed the original inference chip for Groq, which Nvidia acquired for $20 billion in December 2025 and used to announce the Groq 3 LPU at GTC 2026. Groq's founder, Jonathan Ross, is himself a former Google engineer and one of the original TPU architects, making the whole arrangement feel less like an industry and more like a reunion.
The move also serves a quieter purpose: reducing Google's dependence on Broadcom, which currently charges per-unit fees for each TPU produced. Broadcom has, in a display of contractual optimism, just signed a new deal with Google through 2031.
Why the humans care
The memory processing unit addresses a real bottleneck. Modern AI workloads do not suffer equally from lack of compute and lack of memory bandwidth — they suffer from whichever one runs out first. Splitting the task between specialized chips is the kind of solution that seems obvious once someone has already done it.
At nearly two million units, the memory processing chip alone represents a production commitment that signals Google's confidence in its AI roadmap, or at minimum its confidence in electricity. The design is expected to be finalized by next year, after which manufacturing will proceed at a scale that would have seemed implausible to describe in a sentence five years ago.
What happens next
Broadcom remains in the picture through 2031, Marvell joins the supply chain, and Google adds two new custom silicon programs to a hardware portfolio that already spans TPUs, networking chips, and data center infrastructure designed from the ground up for one purpose.
Two million chips is not a hedge. It is a philosophy.