The biggest names in tech have decided that buying AI chips off the shelf isn’t enough. Google, Meta, Amazon, Microsoft, OpenAI, Anthropic and Apple are pouring more money into custom silicon built around their own workloads.
The winners so far aren’t necessarily the companies whose logos go on the chips. Broadcom, Marvell Technology and TSMC are the ones doing the design work and the manufacturing, and the numbers show it.
The picks-and-shovels trade, silicon edition
The chips in question are application-specific integrated circuits, or ASICs. Unlike general-purpose graphics processors, an ASIC is built for one job. A company that runs the same kind of AI task billions of times can get a chip tuned precisely for that task.
Broadcom is the dominant player here. It holds a 60-70% share of the custom AI ASIC design services market as of October 2026.
Its financials match that position. Reported AI semiconductor revenue hit $16.7 billion in fiscal Q3 2026, up 221% year over year.
The pipeline looks just as full. Broadcom carries an estimated backlog of ~$73 billion in AI-related contracts as of October 2026.
Marvell plays a strong second fiddle
The company expects its custom business revenue to exceed $12 billion by fiscal 2029. It projects that figure could reach ~$30 billion by fiscal 2031.
At its Investor Day in October 2026, Marvell went further. It disclosed revenue targets of $70 billion to $90 billion by fiscal 2031.
A big piece of that optimism traces back to Google. Marvell’s partnership with the search giant includes potentially $120 billion in cumulative agreements as of August 2026.
TSMC prints the chips, and the money
TSMC’s revenue grew 53% year over year in August 2026, driven by demand for AI-related wafers.
Why Big Tech is going custom
Custom AI ASIC shipments are projected to grow ~44.6% year over year in 2026.
Custom ASICs are also expected to capture ~27.8% of the AI server market by 2026. That gain comes alongside a projected reduction in Nvidia’s share from around 70%.
What this means for investors and the chip landscape
Broadcom’s appeal rests on scale and visibility. A 60-70% share of design services plus an estimated ~$73 billion backlog gives it an unusually clear view of future revenue. The risk is concentration: a market dominated by a handful of enormous customers means losing even one relationship would sting.
Marvell’s forecasts imply dramatic expansion between now and fiscal 2031, and much of that rests on agreements described as potential rather than locked in.
TSMC’s 53% revenue growth in August 2026 reflects demand across the AI chip spectrum, not just custom designs.
For Nvidia, the projected slide from around 70% of the AI server market is worth watching, though it unfolds inside a market that is growing quickly.
The key signals to track going forward are hyperscaler capital spending plans, any changes to Broadcom’s backlog figure, and whether Marvell’s big agreements convert into booked revenue on schedule. TSMC’s monthly revenue reports also serve as a useful real-time gauge of how much AI silicon is actually being produced.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.