The Operating System
15 Platform Play

Own the Whole Stack, Not Just the Chip

To sell data centers, Nvidia had to make more than processors: software, systems and networking became one product.

if you want to be a data center company, building the processing chip isn't the way to do it.

Huang’s argument, on Acquired in 2023, was that a chip is not what a data-center customer buys: “if you want to be a data center company, building the processing chip isn’t the way to do it.” Nvidia announced the Mellanox acquisition on 11 March 2019 at $6.9 billion in cash ($125 a share) and closed it on 27 April 2020 at about $7 billion. The press release said: “Addressing this demand will require holistic architectures that connect vast numbers of fast computing nodes over intelligent networking fabrics to form a giant datacenter-scale compute engine.” Huang later called Mellanox “one of the best strategic decisions I’d ever made.”

The GPU Where Nvidia started: graphics chips since 1995, the "GPU" label coined for GeForce 256 in 1999.
CUDA software The programming model, launched 2006, which is what developers actually lock into.
Systems DGX-1 in 2016 (about $129k list) turned chips into a complete AI computer.
Networking Mellanox, bought for ~$7B in 2020; networking revenue hit $14.8B in a single quarter (Q1 FY2027).
The AI factory Whole data centers sold as one system, the unit Huang now pitches.

Doesn’t transfer at small scale, and it has limits even at Nvidia’s. Full-stack ownership needs the balance sheet to buy a Mellanox and the market position to make customers accept a bundle. It also invites regulators: China’s SAMR has preliminarily found Nvidia breached the conditions attached to the Mellanox approval, and the $40 billion Arm deal was abandoned under regulatory pressure. Integration is a moat until it becomes a case file.

Across the chapters