Hyperscale Cloud Providers Shift Toward Multi-Vendor Custom Chip Co-Design
nashnova research
Hyperscale cloud providers are breaking up custom chip design from single-partner projects into multi-vendor contests where suppliers compete on specialized IP for individual blocks inside the same chip. Google's expanding supplier roster is the clearest case so far, and the shift is rewriting how profits flow through the custom-silicon food chain.
Why is the single-partner model fading?
Cloud giants used to hand an entire custom chip to one design house, start to finish. Now they unbundle the project by function and assign each block to the strongest specialist.
The technical enabler is chiplet architecture — splitting an AI accelerator into separate compute dies, memory subsystems, and high-speed interconnect modules, then packaging them together.
This means → once a cloud provider's in-house engineering team matures, it no longer needs a turnkey vendor. It acts as its own general contractor, picking best-in-class subcontractors by IP strength.
What does Google's supply-chain expansion reveal?
MediaTek's path is the clearest template: it entered Google's TPU supply chain through a specialized-function chiplet, then worked toward full-chip integration contracts in later generations.
Google recently added Marvell to its custom-chip supplier ecosystem and is exploring a potential technical collaboration with AMD — the goal is to assemble an optimal platform from each vendor's strongest IP.
Industry sources say Marvell's likely entry point is memory-controller and memory-subsystem interface IP. Its high-speed SerDes IP — the technology that moves serial data between chips at very high rates — puts it in direct competition with MediaTek for the next-generation interconnect socket.
Who captures the margin — and who doesn't?
Suppliers holding high-value proprietary IP command premium gross margins. If your technology is hard to replace, pricing power stays with you.
Partners handling the bulk of integration and tape-out work control project execution but face lower overall gross margins because of high pass-through costs.
In plain terms = inside the same chip, "the one with the secret recipe" earns more than "the one doing the heavy lifting." Profit is migrating from the execution layer to the IP layer.
Does system-level integration still matter?
Industry veterans stress that despite clear IP premiums, system-level integration remains a critical competitive moat.
When cloud providers make the final call, three factors come first: cost efficiency, full integration, and time to market.
This means → a supplier that can span multi-die packaging or even rack-scale scale-out architectures is positioned to capture a larger revenue share. Owning a single piece of IP is not enough — you also need to prove it works inside a running system.
Why are top suppliers diverging on the "full-rack" question?
Broadcom and Marvell have explicitly said they will not expand into rack-level system integration. They are betting on IP specialization.
NVIDIA and Qualcomm are moving the opposite way, seeking to deliver complete rack-scale infrastructure.
MediaTek has not publicly stated its long-term roadmap for full-rack systems.
This reflects two distinct moat strategies: IP specialization vs. full-stack system ownership. Which path ultimately commands the higher premium in cloud procurement decisions is the key test of this custom-chip landscape reshaping.
市场有风险，内容仅供研究参考，不构成投资建议。