Marvell Confirms Adoption of TSMC 1.4nm Process for AI Data Center Strategy
0xBroomberg
Marvell has confirmed it will use TSMC's A14 (1.4nm) process for next-generation data center chips, targeting 2028 mass production — a sign that the AI compute race is now pushing process technology toward the 1nm frontier.
Why bet on a process that won't ship until 2028?
One word: power. Marvell president Koopmans told Nikkei Asia that AI data centers are extremely power-sensitive — "every tenth of a watt matters."
This means → the value of a leading-edge node isn't raw speed. It's doubling bandwidth while keeping the power increase well below 2×.
For Marvell's digital signal processors — DSPs, the chips that convert high-speed signals between fiber optics and silicon — the case for moving to a smaller node is "purely about power."
Has Marvell's "skip a generation" gamble worked before?
Yes. Marvell once jumped from 16nm straight to 5nm, skipping 7nm entirely — a rare move in the industry.
Koopmans' words: "You can't win by just outrunning a formidable competitor. You have to skip a node and go to the next one."
In plain terms = Marvell doesn't compete by being half a step ahead. It bets big and aims to lead by a full generation. Going from 3nm/2nm directly to 1.4nm is the same playbook.
Why put all eggs in the TSMC basket?
Marvell historically used Samsung, GlobalFoundries, and other foundries. After pivoting to data center, it consolidated production at TSMC alone.
It already uses TSMC's 3nm for the industry's first 1.6-terabit interconnect chip platform DSP, and is the first to bring 2nm into DSPs and data center interconnect pluggable modules.
This reflects an industry reality: on the AI chip track, leading-edge process = TSMC. There is virtually no second option.
What does a $1 billion prepayment buy?
Marvell gives TSMC a rolling five-year capacity forecast and has committed $1 billion upfront to lock in future capacity. Koopmans says "most companies won't do that."
This means → Marvell isn't paying for chips. It's paying for certainty that capacity will be there in three years — in an AI demand surge, failing to secure capacity is more fatal than falling behind on technology.
In plain terms = put up $1 billion so TSMC knows you're serious, and you move to the front of the production queue.
How much has Marvell itself changed?
2016: data center revenue under $200 million, less than 10% of total. Consumer electronics contributed over 60%.
Now: data center is the core growth engine; consumer electronics has shrunk dramatically.
This means → Marvell has fully transformed from a consumer electronics chip company into an AI data center chip company. Whether 1.4nm ships on schedule in 2028 is the critical checkpoint for this strategic bet.
Content is for reference only, not financial advice.