2 Artificial Intelligence (AI) Stocks to Buy as Demand for Custom Chips Soars

Broadcom reported Q2 fiscal 2026 revenue of $22.2 billion, up 48% year over year, with AI chip sales rising 143% to $10.8 billion and adjusted EPS increasing 54% to $2.44. Marvell posted record Q1 fiscal 2027 revenue of $2.4 billion, up 28%, with adjusted EPS up 29% to $0.80 and management guiding for revenue growth to accelerate each quarter of fiscal 2027. The article argues that rising hyperscaler capex and demand for custom AI ASICs should benefit both companies, especially Broadcom, which already has major partnerships with Alphabet, Meta, and OpenAI.

Analysis

The key second-order effect is not simply “more AI capex,” but a structural shift in how hyperscalers allocate compute budgets: custom ASICs are increasingly a margin-arbitrage tool versus merchant GPUs. That favors the few suppliers with embedded design wins and long tape-out cycles, because once a chip is qualified the spend tends to recur across multiple generations, creating a multi-year revenue annuity rather than a one-off project. Broadcom looks best positioned to capture this because it is already sitting inside the largest platform roadmaps, which raises the probability that near-term growth is less about winning new logos and more about monetizing existing ones more deeply.

Marvell’s opportunity is real, but it remains more execution-sensitive. The market may be underestimating how much of Marvell’s upside depends on customer concentration and on the timing of hyperscaler deployment ramps; that makes the stock more sensitive to quarterly guide changes than Broadcom. The fact that custom silicon is growing does not automatically mean Marvell’s share of the wallet expands proportionally—if the larger platforms push more volume to the vendor with the broadest software, packaging, and networking ecosystem, Marvell can still grow fast while underperforming relative to the category.

The real loser is not Nvidia in absolute terms, but the marginal growth rate of the highest-cost accelerator segment. If ASIC adoption keeps rising, some AI workloads will shift from “best performance per chip” to “best total cost per inference,” which caps GPU pricing power at the margin and can pressure mix in future procurement cycles. That said, the bearish read on Nvidia is probably too early: this is a years-long rebalancing, not a sudden replacement, so any short should be sized as a relative-value trade rather than a directional collapse call.

Consensus may be overrating the immediacy of the monetization while underrating the durability of the cycle. The biggest risk is that the current enthusiasm front-runs 2027–2028 spend, leaving both AVGO and MRVL vulnerable to air pockets if hyperscalers digest prior capex or if custom chip yields slip. Near term, the cleaner trade is to own the winners on pullbacks rather than chase momentum after strong prints.

Key Decisions for Investors

• Long AVGO on any 5-8% post-earnings or post-rally pullback; thesis is multi-year embedded design wins with lower execution risk than peers, implying superior visibility into 2027 capex translation.
• Long MRVL only as a higher-beta satellite position, sized at roughly half the AVGO weight; better risk/reward if entered after a guide reset or market-wide AI selloff, because the stock is more exposed to customer-timing noise.
• Pair trade: long AVGO / short NVDA for a 3-6 month relative-value expression on custom ASIC share gains; this is a hedge against the market overpaying for GPU mix while ASIC content compounds.
• Avoid shorting NVDA outright; if expressing skepticism, use a defined-risk put spread 6-12 months out, since the replacement cycle is gradual and GPU demand can remain strong through the next capex wave.
• Monitor hyperscaler capex revisions into the next earnings season; if 2027 spending guidance keeps ratcheting up, add to AVGO before the market fully reprices the duration of the ASIC backlog.