Nvidia crossed a $1 trillion market cap in 2023, but what truly cemented its throne as the king of AI compute isn't faster chips — it's the software stack the company began laying down in 2006 and that no one has dared swap out for twenty years. Understanding this is the prerequisite for grasping why AI compute costs what it does in the AI era.
What this is
Nvidia was founded in 1993 at a Denny's diner in California, initially making graphics cards that rendered 3D images for personal computers. It went public on Nasdaq in 1999, launched the GeForce 256, and coined the term GPU (Graphics Processing Unit) for this class of chip.
The real turning point came in 2006 with CUDA — a software platform that let programmers use graphics cards for general-purpose computing. Investors saw it as burning cash at the time, but Jensen Huang held the course for a decade. In 2012, the University of Toronto used two Nvidia gaming GPUs to train the image recognition model AlexNet, crushing the competition. AI researchers realized: training neural networks needs exactly what GPUs excel at — massive amounts of simple computation done in parallel. CUDA's decade-long bet paid off.
In 2017, Nvidia added Tensor Cores; in 2020, it acquired networking company Mellanox, pivoting from "selling chips" to "selling full compute systems." When ChatGPT ignited the world at the end of 2022, the H100 was impossible to get, and Nvidia's market cap crossed $1 trillion in 2023.
Industry view
The dominant narrative is that "CUDA's ecosystem is unbeatable": after nearly twenty years of accumulation, almost every AI codebase on Earth runs on CUDA, and the switching cost is prohibitive. Combined with tight integration across chips, interconnect (NVLink), and rack-scale systems, Nvidia sells not just a chip but a complete "compute factory."
But we're also hearing dissenting voices.
One concern is customer concentration. The five biggest buyers of Nvidia GPUs globally — Microsoft, Google, Meta, Amazon, and Oracle — are all developing their own chips. Google TPUs and Amazon Trainium are already usable; OpenAI is working with Broadcom. Once in-house silicon matures, the "CUDA lock-in" could be bypassed.
Another risk is TSMC dependence. Nvidia doesn't manufacture its own chips, outsourcing everything to TSMC, leaving geopolitical shifts as a sword hanging over the company. There's also a cooler-headed take: CUDA's moat is essentially "best available today" — once open-source alternatives (such as AMD's ROCm or Triton) mature, or model architecture shifts again (say, inference replacing training as the main battleground), the moat could be redrawn.
Impact on regular people
For enterprise IT: AI product iteration speed increasingly depends on whether you can secure GPU compute. For most companies, renting cloud compute is more realistic than building a data center.
For your career: the AI tools you use daily (chatbots, translators, writing assistants) all run on scarce GPU resources. When internal AI projects at companies get delayed, the problem usually isn't the model — it's "not enough cards."
For consumer markets: gaming GPU supply and pricing remain squeezed by AI demand. Gamers may end up subsidizing, indirectly, the boom in cloud AI.