Jensen (and his wife, Lori) almost died in a car accident in 1984, years before he co-founded Nvidia:

Jensen hit the transparent layer of black ice that coated the freeway and sent the Supra into an unrecoverable glide. The tires spun without purpose, and the vehicle drifted onto the shoulder before rolling off the road. Jensen and Lori were momentarily inverted. Then the car hit the ground with an awful crunch before banging along to a stop.... The Supra was totaled, and the couple was trapped inside. Lori, wearing her new engagement ring, was mostly unharmed. Jensen was bleeding, and his neck was twisted bad.... When the first responders eventually arrived, they had to cut the couple out of the car. Jensen required stitches in multiple places and had to wear a neck brace for several months thereafter. When I asked him about the incident years later, he mostly expressed regret for the Supra. "Incredible car," he said.

Nvidia went public in 1999 at a $600M valuation, but there wasn't a party, or popped champagne. One employee saved an email Jensen sent him the day after the IPO:

The TNT2 team needs to do whatever it takes to get over the finish line. They're fighting for every single minute within Dell and Compaq for the motherboard business. S3's Savage 4s are working well on Camino, but we're still struggling. There's no more time. Do what it takes to get it done. We need design wins to take share from ATI and keep S3 down, and take Nvidia to the next level. Remember, there are three priorities: one, two, and three. We're counting on 250,000 units of TNT2 shipments by April to make our Q1. Do that and we've got to cert risk wafers accordingly. Get it done.

Nvidia hit some stumbles in the early 2000s, including an accounting scandal, a lost deal with Xbox, and the famous "dustbuster" GeForce FX with a deafening fan. The latter event was one of many occasions for Jensen tearing into employees in large meetings:

Huang arranged a meeting in which the product managers presented, to a few hundred people, every decision they had made that led to the fiasco. Huang then screamed at them, near the top of his voice, for nearly an hour. "'Terrifying but cathartic' is how I would describe it," said Sharon Clay, one of the engineers responsible for quality control. Huang's tirades inspired as much guilt as fear, and he often described, in detail, how in letting their customers down, Nvidia employees had let one another's families down as well. ("I think I'm driven as much by guilt as anything else," Huang told me.)

Despite his temper, Jensen would rarely fire people. Tenure of Nvidia employees was significantly higher than many other companies in Silicon Valley. Jensen would make personal appeals to even low-ranking employees to keep them at Nvidia.

Compute Unified Device Architecture, or CUDA, is a proprietary Nvidia platform that originally enabled developers to use the GPU, previously built exclusively for 3D graphics, for general purpose computing. This is what eventually allowed Nvidia chips to accelerate the field of machine learning. But it was a longshot bet, and led to an activist investor attempting to force Jensen out as CEO.

Eventually, two longshot subfields of computer science proved correct: parallel processing and machine learning. Had either of these not materialized, Nvidia would not be the company it is today.

Nvidia's gains in processing speeds across the 2010s and into the 2020s really came more from software innovations than hardware advancements:

Only a small portion of the performance gains now came from the classic strategy of packing more transistors into the chip—Moore's Law was dead. The rest came from... Nvidia scientists accelerating matrix multiplications with numerical magic tricks. Nvidia engineers taught the GPU new instructions that acted like speed solvers on a Rubik's cube. They replaced the processor's native language with ugly but effective lo-fi data types, akin to switching from calligraphy to shorthand. They trimmed 'dead' synapses from the matrices, essentially deleting unproductive information from the neural net in the manner of the forgetting machine in Eternal Sunshine of the Spotless Mind. Between 2012 and 2022, Nvidia achieved a thousand-fold speedup in single-chip AI inference performance, which was far in excess of anything that Moore's Law had ever achieved. A mere 2.5x of that speed-up came from transistors; most of the remaining 400x came from Nvidia's mathematical toolbox.

Great book, 10/10 would recommend.