NVIDIA CEO Jensen Huang

The RIVA 128 launch was a survival decision. NVIDIA had roughly six months of cash, had already chosen an architecture that proved incompatible with Microsoft’s DirectX, and faced dozens of new competitors. The chip returned from fabrication supporting only eight of DirectX’s 32 blend modes. Huang jokes that the other 24 were unimportant, but admits the team should have implemented them. The practical task was to make the product viable anyway by giving developers a reason to use the supported features.

Huang describes RIVA 128 as a reset because NVIDIA abandoned its prior direction and aligned with DirectX rather than continuing to fight it. The team built a fully hardware-accelerated 3D pipeline, pushed the chip and memory to the limits of what was physically feasible, and priced it above what competitors were likely willing to attempt. The strategy depended on a real customer segment: PC enthusiasts who wanted the fastest available graphics, even at a premium. Huang argues that 3D graphics remained attractive because users continually wanted more performance.

The apparent gamble of ordering production without a physical prototype was enabled by emulation. NVIDIA bought an emulator from a company that was shutting down, then ran the chip’s full software stack on it. The team tested drivers, software, games, and VGA applications before fabrication. Huang says the simulation could take an hour to render a frame, but the slow feedback was still preferable to the conventional cycle of fabricating a chip, writing software, finding bugs, and fabricating again—a process that would have consumed the company’s remaining cash.

Huang’s lesson is not to bet the company blindly whenever a founder feels conviction. He says NVIDIA could proceed because it had pulled future risks forward: the team had virtually prototyped the chip and tested the software before committing to one tape-out. Once the company believed the design was ready, it also prepared production, marketing, and developer relations in parallel. The principle is to compress uncertainty through advance work, so a final commitment rests on evidence rather than hope.

NVIDIA’s early work on graphics also created a path toward general-purpose GPU computing. Before CUDA, the company explored CG and abstraction layers that could expose GPU capabilities through higher-level languages. Programmable shaders were already massively parallel, and the team had begun testing uses such as CT reconstruction and image processing. These were signs of an opportunity, but they did not yet prove that a large machine-learning market would emerge.

AlexNet prompted Huang to return to first principles: if deep learning could leap ahead of decades of computer-vision work, why did it work, and could it scale to other problems? He points to the ability to learn predictions from large collections of examples and to train increasingly deep neural networks layer by layer. This led him to imagine a general-purpose function approximator with applications beyond the first visible use cases. The discussion is an account of the reasoning that shaped NVIDIA’s investment, not a claim that the full AI market was already predictable.