Marvell Wins the Custom-AI Design. TSMC Owns the Choke Point.

In two weeks, Marvell's stock fell 10% on a record quarter and then jumped 6% on a long-term guidance update. Same facts, opposite reactions — which is the tell that investors are arguing over what the numbers actually mean, not whether Marvell is winning the custom AI chip business. It pretty clearly is. The question worth asking is who gets to keep the economics.

Marvell designs custom AI accelerators for the biggest cloud buyers. It reported record fiscal second-quarter revenue of $2.74 billion, up 37% from a year earlier, with the data-center business climbing 46% and now about four-fifths of the company's sales. Management then lifted its multi-year outlook, calling for roughly $12 billion of revenue in fiscal 2027 and $18 billion by fiscal 2028. The stock is up about 250% over the past year and trades near $236 with a market value around $207 billion.

That is the visible win. The chokepoint lens starts one layer below it.

The chip goes where the bottleneck is

Marvell designs the chip. It does not build it. Along with Broadcom, it is one of two dominant designers of custom AI silicon — the two together control an estimated 95% of the market, with Broadcom at around 70% and Marvell the clear number two — to customers including Amazon (Trainium), Microsoft (Maia), and Alphabet. The leading-edge versions of those chips — like the AI accelerators and custom ASICs that dominate hyperscaler demand — are overwhelmingly manufactured and then packaged by a single company: TSMC.

The narrowest node is not the design socket. It is TSMC's advanced packaging, called CoWoS, which stacks compute and memory onto one package and has become the physical constraint on the whole industry. CoWoS lines are booked out with lead times of roughly a year, Nvidia controls maybe 60% of the allocation, and TSMC dominates about 95% of leading-edge packaging and most of advanced fabrication. Broadcom did something unusual in March: it publicly named TSMC capacity as the ceiling on its own AI chip growth. Marvell faces the identical constraint — it must win a slot in the same queue to ship the chips whose revenue it is already booking in its guidance.

This is the useful reframe. Marvell's custom business is booming precisely because hyperscalers want an alternative to Nvidia's GPUs — custom ASIC shipments are growing about 44% a year, roughly triple the rate of merchant GPUs. But routing around Nvidia does not route around TSMC. Every detour still crosses the same bridge, and the bridge collects a toll no matter whose name is on the chip.

Who captures the rent

That is why the competitor framing — that TSMC may be the safer way to play Marvell's upside — is more than a slogan. The two companies sit at different points in the chain with different amounts of pricing power.

TSMC gets paid regardless of which design wins. Whether Nvidia, Broadcom, or Marvell sells more, every accelerator runs through TSMC's fabs and packaging lines, and demand for all of them is rising at once. TSMC holds ~90% of advanced AI fabrication, its 2nm capacity is reported booked into 2028, and its gross margin runs around 64%. The growth of the entire custom-ASIC category is simply more volume flowing through the same chokepoint — wind at TSMC's back even if Marvell and Broadcom are trying to take share from Nvidia.

Marvell's position is different and thinner. It must win design sockets against a rival with roughly three times its custom-chip share, and its chips depend on manufacturing and packaging capacity it does not own and must queue for. The economics also get diluted as custom work grows: the jump in lower-margin ASICs is steadily pushing Marvell's blended gross margin down, with non-GAAP gross margin at 58.9% last quarter and falling. And the multi-year revenue targets are guidance, powerfully suggestive but not binding bookings — hyperscalers can reshuffle orders among Nvidia, Broadcom, Marvell, and their internal teams.

None of this means Marvell is a bad business. It means the market has priced it as if flawless execution is already guaranteed. Marvell trades at roughly 78 times trailing earnings and about 22 times sales. TSMC trades at about 33 times trailing earnings and 16 times sales, with a far wider economic moat described by the very capacity that caps Marvell's growth.

Where the comparison lands

The honest caveat: TSMC is not risk-free. Its leading-edge capacity is geographically concentrated in Taiwan, its expansion is capital-intensive — it spent about $46 billion on capex over the past year — and a slowdown in AI spending would hit it too, just with more cushion from pricing power and a more modest valuation.

So the map points one way. Marvell has been a remarkable story of design execution, and its custom-AI franchise is real. But the dependency that ultimately decides how many of those chips can ship sits with TSMC — a company that profits from the wave no matter which designer wins it, and whose stock has not yet asked the market to pay for perfection. When the bottleneck behind the apparent bottleneck is this concentrated and this indispensable, the safer expression of Marvell's own thesis may be the company Marvell has no way around.