As Big Tech Designs Its Own AI Chips, the Companies Behind the Scenes Win

Custom AI processors are gaining ground: TrendForce estimates their market will grow 45% this year, versus 16% for GPUs, as cloud companies deploy more specialized chips alongside GPUs. Broadcom forecasts AI semiconductor revenue of $56 billion this year and more than $100 billion next year; TSMC reported Q2 sales up 36% to over $40 billion, while Marvell received a $2 billion Nvidia investment. The article favors TSMC, Broadcom, and Marvell as ways to benefit from the expanding market, while noting uncertainty around Intel's foundry business, which lost $2.1 billion in Q2.

Analysis

The important second-order effect is not a clean GPU-to-custom-chip substitution: cloud providers can add cheaper, workload-specific inference capacity while retaining GPUs for training and flexible workloads. That expands total compute demand but shifts some economics from merchant silicon toward cloud providers’ own cost structures—and gives custom designers revenue exposure without guaranteeing attractive margins. The key bottleneck may therefore be advanced manufacturing and packaging, favoring TSM, while design wins at AVGO and MRVL remain dependent on customer ramp timing, yield, and program economics.

Near term (days to 1–3 months), treat the theme as established rather than a fresh catalyst; avoid extrapolating company forecasts into realized earnings. Verify AVGO’s AI revenue conversion and the economics/concentration of its long-dated Alphabet relationship, plus MRVL program ramps. Over 6–18 months, inference growth supports custom accelerators, but NVDA’s software stack, networking and integrated systems may preserve pricing power; custom chips are more likely to cap GPU growth in selected workloads than displace the platform outright. TSM is the broadest volume beneficiary, but its geopolitical exposure and capacity constraints are material counterweights. INTC’s foundry losses make it a weaker exposure absent evidence of external-customer scale.

Contrarian read: custom-chip growth rates can look impressive off a smaller base, while cloud capex, not chip design wins, ultimately determines realized demand. No supplied data establishes AVGO is cheap or that its forecast is consensus-beating. Prefer relative exposure over chasing a single-name narrative.

Key Decisions for Investors

• Prefer a measured TSM overweight on broad-market weakness rather than a momentum entry: it participates across GPU and custom-silicon architectures. Reassess if advanced-node/packaging capacity or external-customer demand disappoints, or if geopolitical risk rises materially.
• Keep AVGO and MRVL as watchlist/position-sizing opportunities, not automatic buys. Before adding, verify customer-specific ramp timing, revenue conversion, gross-margin contribution, and whether announced commitments are firm enough to support guidance.
• Avoid shorting NVDA solely on custom-chip growth; its software and systems position may cushion share loss. Monitor relative GPU versus custom-accelerator demand and cloud-provider capex over the next 1–3 months for evidence of actual substitution.
• INTC is a relative underweight versus TSM while foundry economics remain unproven. Falsify the view if external foundry revenue scales and losses narrow sustainably; account for the risk that policy support or new customer wins trigger sharp short-covering.