MaxLinear launches a power chip to save up to half a watt per AI optical module, its answer to Marvell

MaxLinear▼−7.53% — MaxLinear, down 7.53 percent today (Nasdaq: MXL) launched Jeju (MxL77904) on November 10↗, a power-management chip for the pluggable optical modules that move 1.6 trillion bits a second between AI servers ¹. The engineering is real, but the chip is aimed at the wrong line of the income statement: power-management ICs sell for a small fraction of the price of the digital signal processor (DSP) they sit beside ², so Jeju's job is not its own revenue.

Its job is to help sell Rushmore, MaxLinear's 1.6T DSP, into the same module ¹. The shares rose more than 13% the next day, in a broad rally across AI-related technology stocks that the launch also fed ³.

The last half-watt is what a sealed module cannot spare

The newest optical DSPs, built on 3- and 4-nanometer processes, run below one volt while drawing more current, exactly the load where conventional power chips lose efficiency ¹. Jeju is designed to save up to half a watt per DSP at those load points ¹. Inside a sealed module, that half-watt is thermal margin: less heat, higher manufacturing yield, lower cooling cost ¹.

MaxLinear expects combined 800G and 1.6T module volumes near 200 million units this year, so small watts multiply ¹. "At 1.6T and 3.2T, every watt matters," said Amit Bavisi, the senior vice president who runs the analog business ¹. The 95% peak-efficiency figure is a datasheet number; what relieves heat in a data center is efficiency across a changing workload, which only deployed modules can show ².

Two chips in one socket is MaxLinear's answer to Marvell

A week before the launch, Marvell, the chipmaker buying Lumentum, agreed to pay $115 billion in stock for the laser maker, putting chip design and laser manufacturing under one roof ⁴.

Buyers facing a dominant supplier want a second source, and that is the door Jeju is built for: it is tuned for Rushmore but is designed to support 1.6T module designs across the industry, whatever DSP sits at their center ¹. A module MaxLinear does not win with Rushmore can still carry its power chip. Management expects the 1.6T generation to more than double the silicon dollars in each module↗ ⁵, which is why holding any line in that socket counts.

The proof lands in 2027, not this quarter

Everything growing in MaxLinear's results today is the current 800G generation: third-quarter revenue was $218.1 million↗, up 72% from a year earlier, with AI-infrastructure revenue up 210% ⁶. Management expects its optical DSP share to reach the teens within two years, neck-and-neck with Marvell▼−7.21% — Marvell, down 7.21 percent today or Broadcom▼−6.53% — Broadcom, down 6.53 percent today, whose 1.6T parts are already shipping or announced ⁵⁷.

Jeju samples in February 2027, and Rushmore production deployments begin early next year, ramping harder in the second half ¹⁵. What limits the outcome now is not silicon but how many module makers MaxLinear can certify parts with at once, alongside tight DSP wafer supply ⁵.

The rally repriced a company whose 1.6T revenue does not start until 2027 ⁷. Jeju's half-watt is real engineering on a small line; the thesis it serves asks module makers to take two chips from one supplier instead of one. February's samples open that test, and next year's qualifications settle it.