Jensen Huang showed up unannounced at Marvell's keynote and called it the next trillion-dollar company. The fabric between GPUs is the bottleneck the market hasn't priced.
Jensen Huang doesn't appear at other people's keynotes. In two decades of running NVIDIA, he has headlined his own events, given his own demos, worn his own leather jacket on his own stages. On June 2, at Computex in Taipei, he walked onto Marvell CEO Matt Murphy's stage unannounced, stayed for ten minutes, and called Marvell "the next trillion-dollar company."
Marvell added roughly $45 billion in market capitalization in a single session.
NVIDIA disclosed a $2 billion equity stake in Marvell in March, alongside a broad AI partnership. Huang already had capital committed to the thesis. The Computex appearance was the public announcement of a bet he'd placed three months earlier.
The Next Bottleneck
This journal has followed the AI infrastructure stack layer by layer. The Second Customer documented NVIDIA's demand broadening beyond the five hyperscalers. The Input Tax staked a position on memory pricing power. The Interposer traced margin migration from chip design to physical packaging. Each entry identified the same pattern: as GPU supply expanded, value shifted to whichever component couldn't scale as fast.
Networking is next. Every GPU added to a training cluster requires switching capacity. An NVIDIA A100 needed 200 gigabits per second of network bandwidth. An H100 doubled that to 400. A B200 doubles it again to 800. The next generation will push past 1,200. Per-GPU bandwidth has roughly doubled every two years since 2020, and the physics demands it: larger models distributed across more chips generate more inter-chip traffic. A 100,000-GPU cluster at 800 gigabits per port produces petabits of aggregate traffic. The switches that move those petabits are the circulatory system of the data center.
Networking currently represents 10 to 13 percent of data center capital expenditure. That percentage has been climbing. It compounds with scale: the ratio of communication to computation increases as clusters grow, because the number of chip-to-chip connections grows faster than the number of chips. At large enough cluster sizes, the network becomes the binding constraint. GPU utilization drops because data can't move fast enough between processors.
The Duopoly
Two companies make the merchant silicon inside data center switches. Broadcom holds roughly 75 percent of the high-end market. Marvell holds about 10 percent. Together they control 86 percent of the merchant switching silicon deployed in data centers globally. Every Arista, Cisco, and white-box switch vendor is a customer of one or both.
Broadcom's Tomahawk 6 has been shipping in production volume since March 2026. It delivers 102.4 terabits per second on a 3nm process, doubling the prior Tomahawk 5 generation. Broadcom took TH6 from initial sampling to production in under three quarters. It is the current performance leader in merchant switching silicon.
Marvell announced the Teralynx T100 on June 1. It also delivers 102.4 terabits per second, built on a monolithic 3nm die, consuming under 1,000 watts. Marvell claims 25 percent lower power than the competitive benchmark at this bandwidth tier. It supports a 512-port radix, meaning fewer network tiers and fewer total switches in large-scale clusters. Samples are shipping to customers this quarter, with production targeted for mid-2027.
Marvell is not arriving first. Broadcom already holds the performance crown. What Marvell offers is something the hyperscalers want almost as badly: a second credible supplier at 102.4 terabits per second. When a single vendor controls 75 percent of high-end switching silicon, every customer is a captive buyer. A second vendor at parity bandwidth with materially lower power consumption gives procurement teams leverage they have not had in a decade.
The Transformation
Marvell arrived at this position through a decade of deliberate portfolio rotation. Ten years ago, data center revenue was less than 10 percent of the company's total. In the most recent quarter, Q1 of fiscal 2027, Marvell reported record revenue of $2.42 billion, with data center representing approximately three-quarters of the total. The company has guided Q2 at $2.7 billion and expects fiscal 2027 revenue of approximately $11.5 billion, a 40 percent increase over the prior year.
What makes Marvell's position unusual is breadth. The company competes across three categories that converge in AI data centers: merchant switching silicon, custom AI accelerator ASICs for hyperscalers, and optical interconnects. No other company spans all three. Broadcom covers switching and custom ASICs but lacks the same optical portfolio. The vendor that designs across all three layers can optimize for power, latency, and bandwidth simultaneously. That co-design advantage is what Huang was endorsing.
The Bet
Marvell will take meaningful share from Broadcom in AI switching silicon over the next 18 months. The data center networking layer will capture a rising fraction of AI infrastructure spend.
The thesis does not rest on arriving first. Broadcom got there in March. It rests on three structural advantages. Power: a 25 percent efficiency gap at 102.4 terabits per second translates to millions of dollars in annual operating costs across a hyperscale deployment. Breadth: no other vendor designs merchant switching silicon, custom AI accelerators, and optical interconnects under one roof. That co-design capability is what drew NVIDIA's $2 billion investment and Huang's public endorsement. Leverage: hyperscalers building 100,000-GPU clusters cannot afford single-vendor dependency for the network fabric. A credible alternative at the top of the performance stack changes every negotiation.
Three conditions would falsify it. Marvell fails to win design-ins at two or more major hyperscalers by the end of fiscal 2027, suggesting the efficiency and co-design advantages are not decisive. Broadcom closes the power efficiency gap to under 10 percent with Tomahawk 7 or firmware optimization. Or Marvell's data center revenue fails to reach $2 billion per quarter by Q4 fiscal 2027.
For four years, the market has watched the GPU race as though compute were the only variable. The fabric between the GPUs was always going to matter. Jensen Huang just told you who he thinks will weave it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)