Marvell's Will Chu: skipping 7 nanometer fueled cloud AI lead

Marvell's decision to skip a process node was a departure from the industry's usual step-by-step progression, and Chu frames it as a deliberate bet on becoming a leader rather than a follower in data infrastructure. He noted that the company had products at 12 and 16 nanometer, and the next logical step would have been 7 nanometer — which it bypassed entirely. That choice, he said, underpinned its subsequent moves to 5, 3, 2 nanometer and 14 angstrom.

Lazovsky explained that roughly 85% of total data transmission happens inside scale-up networks, with only 15% across scale-out, which is why Marvell is pushing optics into processor-to-processor links. The company's August Photonic Fabric Memory Module and September 2-nanometer optical interconnect technologies are aimed at decoupling memory capacity from compute and reducing total power consumption, respectively. These advances sit alongside Nvidia's $2 billion investment in Marvell earlier this year and a recent agreement to design custom AI chips for Google.