Alphabet (GOOGL) Turns to Marvell (MRVL) for AI Chips as It Looks Beyond Broadcom (AVGO)

Alphabet is reportedly in talks with Marvell Technology to design new AI chips, including a memory unit and a live-app TPU aimed at inference workloads. The move suggests Google is diversifying beyond Broadcom while still maintaining a working relationship through 2031, potentially improving cost efficiency and performance at scale. GOOGL shares rose 1.68% to $341.68, and the article cites a Strong Buy consensus with a $385.97 average target, implying 12.96% upside.

Analysis

The market is starting to price an inference-capex supercycle rather than a one-off training boom. That matters because inference economics are dictated by latency, memory bandwidth, and chip-level specialization, which structurally favors vendors that can attach themselves to the data-movement layer rather than only the compute core. If Google successfully modularizes TPU design, the value pool shifts from a few dominant ASIC design partners toward a broader stack of memory, interconnect, packaging, and custom silicon suppliers — a slow but durable reallocation over the next 12–24 months.

Relative winners are likely to be the companies with exposure to custom inference architectures and high-margin adjunct silicon, while the biggest structural loser is any incumbent whose pricing model depends on being the sole indispensable design partner. The second-order effect is that hyperscalers will increasingly dual-source chip design to reduce vendor leverage, which compresses bargaining power across the ecosystem and can cap long-duration margin expansion for legacy partners. This is less about lost revenue immediately and more about less pricing power at renewal cycles as custom-chip volumes scale.

The setup also reinforces a broader portfolio implication: inference winners can outperform even when AI demand remains strong, because cost-per-query becomes the key battleground. That makes the trade more nuanced than “long AI semis” — investors should prefer names with clear share gains in custom logic, memory subsystems, or AI networking over pure beta to the theme. The contrarian point is that this may be more about supply-chain optimization than incremental AI spend, so the headline is positive but the earnings delta could be smaller than the narrative suggests unless design wins convert into meaningful unit volumes.

Key Decisions for Investors

• Long MRVL on a 3-6 month horizon: buy into any post-news consolidation, targeting continued re-rating as the market prices custom-inference content. Risk/reward is attractive if the thesis expands beyond a single design win into broader hyperscaler custom-chip share.
• Reduce AVGO overweight / hedge with a short against MRVL or GOOGL over 2-4 weeks: the risk is not near-term revenue loss, but multiple compression if investors extrapolate weaker lock-in and lower pricing power at future TPU cycles.
• Pair trade: long GOOGL / short basket of broader AI semi beta over 6-12 months. The cleaner way to play this is via the platform owner with direct inference monetization, while shorting names more exposed to the market’s over-optimistic AI hardware duration assumptions.
• Add NVDA on pullbacks only, but size smaller than broad consensus: inference is supportive, yet the incremental benefit is being competed away by custom silicon adoption. Keep this as a tactical long, not a core overweight, unless model deployment growth reaccelerates over the next 1-2 quarters.
• Watch for follow-on order flow in memory/interconnect names over the next 1-2 quarters; if hyperscaler custom inference broadens, the higher-probability second-order winners may be the suppliers enabling data movement rather than the headline ASIC designers.