Published: April 20, 2026
Category: Technology / AI / Semiconductors
A new report suggests that Google is in talks with Marvell Technology to develop specialised artificial intelligence (AI) chips marking a potential shift in its long-standing chip strategy.
Expanding Google’s AI Hardware Strategy
Google is reportedly exploring the use of custom-built chips known as ASICs (application-specific integrated circuits), designed specifically for AI inference the stage where AI models generate responses for users.
This move reflects a broader effort to:
- Make AI services faster
- Reduce operational costs at scale
- Strengthen infrastructure for products like its AI models and services
Impact on Broadcom
For years, Broadcom has been Google’s primary partner for its Tensor Processing Units (TPUs). While a partnership with Marvell could raise concerns, Broadcom’s position appears stable for now.
- Google and Broadcom already have an agreement extending through 2031
- Adding Marvell could give Google more bargaining power and reduce reliance on a single supplier
Why This Matters
The potential partnership highlights a key shift in the AI industry:
- Companies are investing heavily in custom hardware
- The cost of running AI at scale (power, compute) is rising rapidly
- Diversifying chip suppliers is becoming a strategic priority
Google’s approach also reflects the intensifying “AI arms race,” where companies like Microsoft and OpenAI are pushing rapid advancements in AI capabilities.
Broader Industry Context
The report also notes that Google has opened its TPU infrastructure to other companies, including Meta and Apple, further positioning itself as a central player in AI infrastructure.
Meanwhile, partnerships with firms like Anthropic indicate growing demand for high-performance AI compute, with new capacity expected to come online in the coming years.
The Bigger Picture
If the Marvell deal goes through, it won’t replace Broadcom but it signals a more diversified and competitive approach to AI chip development.
In a market where AI performance increasingly depends on hardware efficiency, controlling the chip ecosystem could be just as important as building the models themselves.
Author.Adigun Adedoye.