In 1993, three engineers met at a Denny's diner in California to found a graphics chip company in a market that already had dozens of competitors. Thirty years later, that company would become — at several points — the most valuable on the planet. This is NVIDIA's history: from the flops that nearly killed it to the happy accident called CUDA that put it at the center of the AI revolution.
1993: three founders and a diner table
Jensen Huang (formerly LSI Logic and AMD), Chris Malachowsky and Curtis Priem (both formerly Sun Microsystems) founded NVIDIA in April 1993, with US$ 40 thousand and the thesis that video games' 3D graphics would drive the future of computing. Jensen was exactly 30 — and made a commitment he maintains today: he is the semiconductor industry's longest-serving founding CEO.
1995–1997: the flop and the rescue
The first product, the NV1 (1995), was an instructive failure: it bet on quadratic surfaces while the world (and Microsoft's new DirectX) standardized on triangles. The company reached 1996 on the brink of bankruptcy — it laid off more than half its staff and had enough cash for one last gamble. The famous internal motto dates from that period: “our company is thirty days from going out of business”. That gamble was the RIVA 128 (1997), a fast, cheap DirectX chip that sold a million units in four months and saved the company.
1999: the birth of the term “GPU”
In October 1999, months after its IPO, NVIDIA launched the GeForce 256 and coined the term we still use today: Graphics Processing Unit. The real innovation was hardware T&L (transformation and lighting), taking work off the CPU. The following year, it bought the remains of the legendary 3dfx — the rival that defined the Voodoo era — and secured the graphics chip contract for the first Xbox.
2006–2007: the happy accident called CUDA
The G80 architecture (GeForce 8800, 2006) brought unified shaders — and with them came the most important decision in the company's history: exposing those programmable cores for general-purpose computing. CUDA, released to the public in 2007, was ridiculed by analysts as an unnecessary cost in video game chips. Jensen persisted for years, putting CUDA in all GeForce cards — unknowingly creating the world's largest installed base of parallel computing.
2012: AlexNet changes everything
In 2012, three Toronto researchers (Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton) trained a neural network on two gaming GeForce GTX 580 cards and crushed the ImageNet image recognition competition. AlexNet proved that deep learning + GPU was the formula — and a decade of CUDA on store shelves meant every laboratory in the world already had the hardware. In 2016, Jensen personally delivered the first DGX-1 supercomputer to the headquarters of a startup called OpenAI.
2018–2022: RTX, Arm, and the H100
In 2018, the Turing architecture brought real-time ray tracing (RTX) and Tensor Cores to the desktop. In 2020, NVIDIA tried to buy Arm for US$ 40 billion — a deal blocked by regulators in 2022. That same year, the H100 (Hopper) reached the market months before ChatGPT launched: perfect timing to become the world's most sought-after hardware.
2023–2026: the world's most valuable company
The generative AI race sent data center revenue soaring. NVIDIA crossed US$ 1 trillion in market capitalization in 2023, US$ 3 trillion in 2024 with the arrival of the Blackwell architecture, and since 2025 has competed for the position of the world's most valuable company — controlling around 92% of the enterprise GPU market, as we covered in the post about CUDA's moat.
The (complicated) relationship with Linux
For our world, NVIDIA is an ambiguous character: quality proprietary drivers, but closed — which earned it the most famous middle finger in free software history in 2012, courtesy of Linus Torvalds. The tide turned: in 2022 the company opened its kernel modules, and today the NOVA driver in Rust and Nouveau/NVK are advancing with official support — we covered the details in the post about Linux kernel 7.3. It makes sense: practically every AI cluster on the planet runs green GPUs on Linux.
Timeline
- 1993 — Founded by Jensen Huang, Chris Malachowsky, and Curtis Priem
- 1995 — NV1, the first flop (quadratic surfaces)
- 1996 — Near-bankruptcy; more than half the team laid off
- 1997 — RIVA 128 saves the company (1 million chips in 4 months)
- 1999 — IPO in January; GeForce 256 coins the term “GPU” in October
- 2000 — Acquisition of 3dfx assets; Xbox contract
- 2006 — G80/GeForce 8800: unified shaders
- 2007 — Public release of CUDA
- 2012 — AlexNet trains on GeForces and ignites the deep learning era
- 2016 — First DGX-1 delivered to OpenAI
- 2018 — RTX/Turing: real-time ray tracing
- 2020–2022 — Attempted Arm acquisition (US$ 40 billion), blocked
- 2022 — H100 released; Linux kernel modules opened; ChatGPT ignites demand
- 2023 — US$ 1 trillion market capitalization
- 2024 — US$ 3 trillion; Blackwell architecture
- 2025+ — Competes for the world's most valuable company position; ~92% of the accelerator market
NVIDIA's lesson is one of strategic persistence: CUDA lost money and drew jeers for almost a decade before becoming technology's most valuable moat. In the next posts in the series, its rivals: the history of AMD and Intel’s history.