Industry contemporary
Nvidia - the GPU company that runs the fabric
CUDA's decade-early bet, the AlexNet ignition, Mellanox - the network's biggest customer became one of its vendors.
Nvidia named the GPU, made it programmable a decade before the world needed it, and became the platform of the AI era. For this site's readers the 2020 Mellanox acquisition is the hinge: InfiniBand, Spectrum-X Ethernet, and BlueField DPUs make Nvidia simultaneously the most demanding workload networks carry and a top-tier network vendor - both sides of the AI-fabric argument.
The profile covers the 1993 founding, RIVA-to-GeForce survival and naming, CUDA, AlexNet, the Mellanox networking turn, the trillion-dollar ascent, and the NVLink/InfiniBand/Spectrum-X fabric wars.
Founding stories
Nvidia
Three engineers bet in 1993 that graphics would become the hardest and most valuable problem in computing. The first chip nearly killed the company; the RIVA 128 saved it; the GeForce 256 gave the product category its name - GPU. The decisive move came in 2006: CUDA made the graphics chip programmable for anything, a decade before the world needed exactly that. When deep learning arrived (AlexNet, 2012, trained on gaming cards), Nvidia owned the substrate - and its 2020 Mellanox acquisition made it, quietly, one of the most important NETWORKING companies on earth.
The timeline
- RIVA 128
After the NV1's near-fatal stumble, the RIVA 128 wins the 3D accelerator market on performance-per-dollar - the survival chapter every Nvidia history starts with.
- GeForce 256
October 1999: hardware transform-and-lighting, marketed as the world's first GPU - the acronym sticks, the category is named, and the IPO the same year funds the run.
- CUDA
The bet that defines the company: general-purpose programmability across every GeForce shipped, plus a decade of unprofitable ecosystem-building - so when researchers needed massive parallel math, the tool was already on their desks.
- AlexNet
A neural network trained on two consumer GPUs crushes the ImageNet benchmark - the deep-learning ignition. The datacenter business that follows turns Nvidia from a components maker into the platform of the AI era.
- Mellanox - Nvidia becomes a network vendor
The ~$7 billion acquisition closes: InfiniBand, Spectrum Ethernet switching, ConnectX NICs, BlueField DPUs. The insight - the datacenter is the computer, so the NETWORK is the backplane - makes fabric performance Nvidia's own product problem.
Deal figure per the public record at closing, April 2020.
- The trillion-dollar ascent
May 2023: first chipmaker to a $1 trillion market cap; $2T follows in February 2024 and $3T that June, with stretches as the world's most valuable company - the market pricing compute, and the fabrics that feed it, as the era's scarcest resource.
Milestone dates per the public record.
- Blackwell and the fabric wars
NVLink scale-up domains, InfiniBand scale-out, and Spectrum-X Ethernet pitched against the merchant-silicon world - the AI cluster's network becomes the industry's central architecture argument, with Nvidia on both sides of the buyer's table.
Flagship products and solutions
- GeForce / RTXThe consumer graphics line that funded everything - and named the GPU category.
- CUDAThe programming platform that is the actual moat - the software layer a generation of AI is written against.
- InfiniBand, Spectrum-X, BlueFieldThe Mellanox-heritage networking portfolio: the fabrics and DPUs of the AI datacenter.
Key innovations
- The GPU as general-purpose computeCUDA converted a graphics part into the era's defining processor - the platform bet that paid off a decade later, at civilization scale.
- The network as the backplaneNvidia's fabric thesis - NVLink up, InfiniBand or Spectrum-X out - made network topology a first-order compute design parameter; every AI-cluster diagram on these pages reflects it.
Main markets
Nvidia dominates AI compute and is a top-tier networking vendor by revenue through the Mellanox lineage - simultaneously the network industry's biggest customer, its most demanding workload, and one of its vendors.
Analyst standing
- First of the contemporaries by design: no company more directly sets what today's networks must carry - and its InfiniBand-versus-Ethernet position is the live argument every AI-fabric design this site's readers touch must answer.