I picked up Stephen Witt's The Thinking Machine expecting to learn about the chips behind AI. I did learn about NVIDIA, but what stayed with me was a less tidy story: the company we now see in polished presentations was built through difficult changes of direction, long periods of uncertainty and an intensity that was not always comfortable for the people around Jensen Huang.
Witt begins with Huang's path to the company. Born in Taiwan, he moved to the United States as a child, spent time at Oneida Baptist Institute, studied electrical engineering at Oregon State University and later worked at AMD and LSI Logic. At Oregon State, he met Lori Mills, his future wife. These details make NVIDIA's origin feel less like a predestined technology legend and more like a life assembled through choices, work and relationships.
A company that had to change its mind
Huang, Chris Malachowsky and Curtis Priem discussed the idea for NVIDIA at a Denny's. Their ambition was to make computers better at graphics, especially for games. But the path from that idea to a successful product was anything but straight. The first chip, NV1, followed an approach that did not become the direction of the industry. NVIDIA had to move away from it while time and money were running short.
That is why the RIVA 128 matters to me. Released in 1997, it was not merely the next chip on a timeline; it represented a decision to abandon an earlier bet and build for the market that was actually emerging. NVIDIA says more than one million units sold in its first four months. The speed of that turnaround is exciting, but it also suggests how close a different outcome might have been.
From rigid linearity to fluid experimentation: perhaps that is the real story behind NVIDIA.
The bet that took years to make sense
The next important turn was not a dramatic rescue. It was a patient investment in possibility. NVIDIA's graphics processors were designed to handle many calculations at once; CUDA, introduced in 2006, gave programmers a way to use that parallel power for work beyond games. But releasing a tool does not mean researchers will adopt it, or that a large market will appear immediately. Witt's account of the long effort to make GPU computing useful makes the later AI success feel less like luck and less like destiny.
In 2012, the AlexNet image-recognition breakthrough showed how powerful GPUs could be for deep learning. In 2016, NVIDIA launched DGX-1, a computer that combined GPUs with software and development tools for AI research. One of those first systems went to OpenAI. The product was no longer just a graphics component sold to a PC maker; it was becoming a computing platform used by researchers and organisations.
Growth did not follow one product line
NVIDIA did not simply leave games behind. GeForce RTX brought real-time ray tracing to gaming in 2018, while the A100 in 2020 addressed AI training, inference and scientific computing in data centres. The H100, announced in 2022, was designed in part for the large models behind modern language AI. Blackwell, announced in 2024, extended the story further: the challenge was now to connect many processors, move data between them and operate complete systems at scale.
That growth has a different intensity from the early race to save a small company. A chip's performance alone is not enough. Software, high-bandwidth memory, manufacturing, networking, electricity and cooling all have to work together. NVIDIA's journey becomes a web of dependencies rather than a single rising line of products or share prices. The breakthrough of one stage creates the bottleneck of the next.
What I took from the book
Witt does not present Huang's drive as purely inspiring. The descriptions of his demanding, sometimes humiliating treatment of employees made the story harder to celebrate without reservation. Behind the smooth keynote is a workplace shaped by urgency and pressure. I think that tension belongs in any account of the company's success.
I also expected the final chapter to tell me more about the future of chips and AI. Instead, its intense exchange with Huang left me thinking about the person steering the company, and about the limits of anyone's certainty about where AI is going. That ending was not the technical answer I wanted, but perhaps it suits a story that has never moved in a straight line.
Looking beyond the book, NVIDIA's 2026 Vera Rubin platform shows how far the idea of a “product” has stretched: not merely a faster chip, but coordinated compute, networking and infrastructure. That is a later development, not something I attribute to Witt's 2025 book. It also reinforces the point I keep returning to: the destination is visible now, but the route was neither simple nor guaranteed.
Sources and context
This reflection began with my reading of The Thinking Machine. I checked the product chronology against NVIDIA's timeline, its longer historical timeline and Witt's interview about the book.
For the later products and systems: NVIDIA's accounts of DGX-1 and its delivery to OpenAI, GeForce RTX, A100, H100, Blackwell and Vera Rubin. The 2026 detail is context added after the book's publication.