The Thinking Machine summary

By the time Huang proposed to Mills, he was already making a name for himself in Silicon Valley. It was the mid-80s, and 20-year-old Huang was working for Advanced Micro Devices, better known as AMD, one of the Valley’s most influential chipmakers. He still remembers his starting salary – $28,700 – not just as a number, but as a sign of how far he would come.

Huang’s early days at AMD were a hands-on crash course in chip design, but he soon moved on to LSI Logic, where his intensity and work ethic quickly set him apart. This is where Huang came to know two important figures in his life: Curtis Priem and Chris Malachowsky. They were engineers who worked at Sun Microsystems, a demanding client of LSI Logic.

Part of the business plan made perfect sense. In the early ’90s, there was a growing consumer video game market that the big players like Sun Microsystems and Silicon Graphics were choosing to ignore. But it was also true that a whole stampede of startups were rushing to fill in this gap – to turn high-end graphics tech into something PC gamers could use. The key invention was the “graphics accelerator” – a plug-in circuit board with a graphics chip at its heart.

For Huang, this was an attractive challenge, and after a few meetings with Priem and Malachowsky in the back of the local Denny’s, tinkering with projections, he talked himself into the risk. There was, however, a personal benchmark: if they couldn’t clear $50 million in yearly revenue, it wasn’t going to be worth his time.

The mood at Nvidia in 1996? “Funereal” would be a nice way of putting it. When Huang was done cleaning house, all that was left was a skeleton crew of 35 engineers. Their survival was based on coming up with a last-ditch graphics card – and fast.

The NV1 disaster was the first time people got a taste of Huang’s leadership style. In a word, it’s explosive. And public. Huang doesn’t scream at you behind closed doors, he does it at meetings and in the open so that everyone gets the message. Even Priem was not safe from Huang’s wrath.

In the late ’90s, all Nvidia was trying to do was make the best graphics cards on the market. And by committing itself to coming out with updates every six months, it was, to put it bluntly, kicking ass. But in cracking the parallel computing hurdle, it unwittingly created a product for a whole new demographic: scientists.

Since the 1970s, researchers had been trying to create a neural net – a version of artificial intelligence inspired by the structure of the human brain. Many in the field thought neural nets were a waste of time, but in 1986, the tide turned. David Rumelhart, Geoffrey Hinton, and Ronald Williams unveiled the backpropagation algorithm – a way for neural nets to learn from their successes and mistakes and make adjustments, just like a real brain.

While most of the world shrugged at what CUDA was capable of, Huang was hiring academic legends like Bill Dally to turn Nvidia Research into a world-class R&D engine.

But Geoffrey Hinton’s AI group in Toronto was aware of CUDA’s power. Among his group was a quiet genius named Alex Krizhevsky who used CUDA and some Nvidia GPUs to build a neural network that could recognize images faster and better than anything that came before. The model was called AlexNet. The machine learning revolution had begun and Huang’s decade of persistence was about to start paying off.

The real judgment day arrived on February 21, 2024, the day of Nvidia’s earnings call. Over the past few years, Nvidia had expanded with the $7 billion purchase of the Israeli data center and networking company Mellanox, and it had signed a $100 million deal with OpenAI to train GPT-4. It now controlled 90 percent of the AI chip market, and its chips were seeing gross profit margins of over 90 percent. No one was harnessing raw computational power like Nvidia. But the earnings call would tell the true story.

Beginning just after the stock market closed, it was one of the most listened-to earnings reports in Wall Street history. Colette Kress, Nvidia’s CFO, opened with the news: annual revenues had more than doubled to $60 billion, and the company had an impressive 70 percent gross margin – beating even Apple. The net income was almost $30 billion, more than Nvidia had earned in the past 30 years combined. To put it another way: every employee at Nvidia was now generating a million bucks in profit.

The last thing to consider is the future of AI itself. Is the current progress really sustainable? 

Putting aside questions of whether AI could become self-aware and threaten humanity, there are already some very real threats to Earth to consider. Nvidia’s rise sparked a huge increase in global electricity demand. Companies like Google and Microsoft are seeing emissions skyrocket due to their ever-expanding data centers. Nvidia’s chips, powering most of them, are massive consumers of electricity. Just a few years ago, their DGX boxes used about the same power as a clothes dryer. Today’s systems are sucking up enough electricity to power entire homes.