Google in talks with Marvell to build new AI chips for inference, The Information reports

Google is reportedly in talks with Marvell Technology to develop two AI chips: a memory processing unit for its TPU system and a new TPU designed to run AI models more efficiently. The companies are said to aim to finalize the memory processing unit design as soon as next year before test production. The news supports Google's efforts to make TPUs a more viable alternative to Nvidia GPUs and could be a modest positive for Google’s cloud and AI strategy.

Analysis

This is less about a single chip program and more about Google trying to reprice the AI infrastructure stack from a pure compute race into a systems-design race. If Google can commercialize a memory-centric accelerator alongside its TPU roadmap, it could reduce dependence on Nvidia's software moat by attacking the bottleneck that increasingly matters in inference: memory bandwidth, not just raw FLOPs. That shifts bargaining power upstream to custom silicon designers and packaging partners, while compressing the addressable premium for general-purpose GPUs in cloud deployments over the next 12-24 months.

Marvell is the cleanest second-order beneficiary because it sits at the intersection of custom silicon, interconnect, and high-speed memory subsystems. The market may still underappreciate how quickly hyperscaler custom chips can scale once the first design wins are validated: a modest ramp at Google can convert into repeat orders and design spillover at other cloud customers within 2-3 product cycles. The bigger strategic signal is that AI capex is fragmenting into a multi-vendor ecosystem, which should support suppliers with IP in chiplets, SerDes, and advanced packaging even if headline GPU growth normalizes.

The risk is timing. Design announcements are not revenue; the first meaningful earnings contribution is likely months to years away, and test-production slippage is common. Near term, the trade can still reverse if Google decides to keep the program internal, if performance-per-watt gains disappoint versus Nvidia's next platform, or if cloud demand softens enough that hyperscalers cut custom silicon budgets.

The contrarian take is that this may not be bearish Nvidia in the medium term. A successful Google chip stack expands total AI deployments by lowering unit economics, which can increase inference volume enough to sustain GPU demand elsewhere. In other words, the immediate displacement risk is real, but the bigger 2025-2026 effect could be accelerated AI consumption, benefiting the broader semiconductor supply chain more than it hurts the incumbent.

Key Decisions for Investors

• Long MRVL vs short NVDA in a 3-6 month pair trade: use the market's tendency to overreact to 'GPU replacement' headlines while positioning for custom-silicon attach rates to re-rate; target 10-15% relative outperformance if design momentum is confirmed.
• Buy MRVL on any pullback of 5%+ after the headline, with a 6-12 month horizon; risk/reward favors upside because even a single hyperscaler design win can expand the multiple, while downside is cushioned by existing networking/optical exposure.
• Reduce or hedge NVDA exposure with call spreads 6-9 months out rather than outright shorts; the thesis is not structural collapse but slower share gains in hyperscaler-specific deployments, so upside should be trimmed not eliminated.
• Accumulate GOOGL on weakness as a strategic AI infrastructure compounder; if custom silicon improves cloud margins by even low-single-digit percentage points, that creates a multi-year valuation tailwind that the market usually discounts late.
• Watch for follow-on beneficiaries in packaging and interconnect names; if the program advances to tape-out, rotate a portion of AI exposure into the picks-and-shovels layer where revenue visibility tends to arrive earlier than the headline chip winner.