KEY POINTS
- Alphabet and Marvell Technology are in active discussions to co-develop two AI processors — a memory processing unit and a new inference-focused TPU — sending MRVL shares up 6.3% in premarket trading.
- The partnership signals Google's intent to diversify away from Broadcom, which has been the sole designer of its Tensor Processing Units, pressuring AVGO shares.
- Watch for a formal announcement and design timeline; Google aims to finalize the chip architecture by 2027, with test production to follow.
Marvell Technology shares surged 6.3% in premarket trading Monday after The Information reported that Alphabet is in active discussions to co-develop two new AI processors with the semiconductor company, a deal that would reshape the competitive dynamics of the custom silicon market and put direct pressure on Broadcom.
The collaboration targets two distinct chips. The first is a memory processing unit designed to work alongside Google's existing Tensor Processing Units, addressing the memory bandwidth bottleneck that limits inference throughput on large language models. The second is a next-generation TPU built specifically for running AI models at scale — a chip that would compete directly with Broadcom's current role as Google's TPU design partner.
Why Google Wants a Second Supplier
Google has relied on Broadcom for TPU design since the program's inception. That single-vendor dependency has become a strategic liability as AI workloads have exploded. Hyperscalers are spending tens of billions annually on AI infrastructure, and any disruption to the chip supply pipeline — whether from manufacturing constraints, pricing leverage, or design delays — carries enormous financial risk.
Bringing Marvell into the fold serves multiple purposes. It creates competitive tension that could improve pricing and timeline commitments from both suppliers. It diversifies Google's design pipeline so that a stumble at one vendor does not stall the entire AI roadmap. And it gives Google access to Marvell's expertise in custom ASICs and networking silicon, capabilities that could prove critical as data center architectures evolve toward more tightly integrated compute-and-memory systems.
For Marvell, the deal would be transformative. The company has been building its custom silicon business aggressively, and landing Google as a flagship customer would validate that strategy and potentially add billions in revenue over the chip's production lifecycle. Marvell shares have underperformed the broader semiconductor index over the past year, and this partnership could serve as a re-rating catalyst.
The Broadcom Problem
Broadcom shares declined in Monday's premarket session on the news. The company has built a lucrative business designing custom AI accelerators for hyperscalers, with Google as one of its anchor clients. Losing exclusivity on the TPU program would not eliminate Broadcom's revenue — Google is unlikely to drop its existing supplier overnight — but it would erode the pricing power and strategic lock-in that have supported Broadcom's premium valuation.
The broader pattern is clear. Amazon has diversified its custom chip program across multiple design partners. Microsoft is building its own Maia accelerators in-house. Meta has expanded its internal silicon efforts. Google's move toward a multi-vendor approach fits an industry-wide trend of hyperscalers reducing dependence on any single supplier, whether that supplier is Nvidia for GPUs or Broadcom for custom ASICs.
The TSMC Connection
Both Google's existing TPUs and any new chips co-developed with Marvell would be manufactured by TSMC, which reported first-quarter profit up 58% year-over-year last week on surging AI demand. TSMC's sold-out advanced manufacturing capacity means that securing production slots is itself a competitive advantage. Google's willingness to fund multiple chip designs simultaneously underscores how aggressively hyperscalers are locking in TSMC capacity for 2027 and beyond.
ASML, which supplies the extreme ultraviolet lithography machines that TSMC needs to manufacture advanced chips, raised its full-year 2026 revenue guidance to €36–40 billion last week, citing sustained AI infrastructure investment. The entire semiconductor supply chain is running hot, and Google's expansion of its custom silicon program adds more demand into an already constrained pipeline.
Traders should watch for a formal announcement from either company confirming the partnership terms and timeline. Google plans to finalize the memory-focused chip's design by next year before moving to test production. The gap between MRVL and AVGO could continue to narrow if the deal progresses, but the real inflection point comes when production volume commitments are locked in with TSMC. Until then, this remains a design-phase catalyst with meaningful long-term implications.
