Top 5 This Week

Related Posts

How Jensen Huang Turned a $97 Million Side Business Into a $5 Trillion Nvidia

Jensen Huang turned a $97 million data centre sideline into a $5 trillion Nvidia. The bets behind it, from CUDA to Mellanox, and the one that failed.

Nvidia reports its third-quarter results this week without Jensen Huang for the first time in its history, and the numbers will still be his. In the three months to January 2016, the company’s data centre business brought in $97 million, about 7% of quarterly revenue, according to Nvidia’s own filing. In the three months to July 2026, the same business brought in $89.0 billion out of $96.2 billion. Few chief executives have ever moved a company that far, and almost none did it by betting on a market that did not exist yet.

Huang stepped down on 12 November as president, chief executive and director, and the reasons have been covered in detail elsewhere on this site, including our explainer on the Megaspeed scandal. This piece leaves his private life to that coverage. Even Nvidia’s board, in the statement announcing his exit, thanked him for his “immeasurable contributions to the company and the industry”. The record behind that line is worth setting out on its own.

CUDA, the 20-year bet Wall Street did not ask for

Huang co-founded Nvidia in 1993 with Chris Malachowsky and Curtis Priem, and its first chip, the NV1, nearly sank the company. He has often said Nvidia was “30 days from going out of business” in those early years, a line he kept repeating long after the danger had passed. The RIVA 128 saved it in 1997, and in 1999 the GeForce 256 arrived as what Nvidia marketed as the world’s first GPU. The company went public the same year, and for most of the following decade it was known as a maker of graphics cards for gamers.

- Advertisement -

The decision that defines his tenure came in 2006 with CUDA. It turned every GeForce card into a programmable parallel computer, which meant spending chip area, engineering time and money on a feature most buyers would never use. Ian Buck, who had built a similar system as a Stanford PhD student, led the work, but it was Huang who kept funding it. Moreover, he kept doing so through the 2008 financial crisis, when investors questioned why a graphics company was building tools for scientists.

The payoff arrived in 2012, when the AlexNet image-recognition model was trained on two GeForce GTX 580 cards and won the ImageNet competition by a wide margin. Researchers had CUDA on their desks already, so deep learning grew up on Nvidia hardware by default. At GTC in March, Huang opened one part of his keynote by noting that “this is the 20th anniversary of CUDA”. By then, millions of developers and thousands of applications were built on it, which is the software moat rivals such as AMD are still trying to cross.

Nvidia’s official highlights from Jensen Huang’s GTC 2026 keynote, where he marked 20 years of CUDA.

Gamers felt those bets too, although they rarely got the credit. The RTX 20 series in 2018 brought real-time ray tracing and DLSS, an AI upscaler that used the same Tensor Cores Nvidia had built for data centres. That loop, with data centre ideas feeding back into GeForce, is one reason PC gamers are watching the succession closely, as our look at the future of GeForce after Huang explains.

Mellanox and the jump from selling chips to building AI factories

By the middle of the last decade, Huang was arguing that the computer of the future was a whole data centre rather than a single box. Nvidia launched the DGX-1 in 2016, an AI supercomputer in one chassis, and the first unit went to a small research lab called OpenAI. A year later, the V100 introduced Tensor Cores, hardware built specifically for the maths behind neural networks. Each step pulled Nvidia further away from being a component supplier and closer to selling complete systems.

The missing piece was the network that ties those systems together. In March 2019, Nvidia agreed to buy Mellanox for $6.9 billion, and it completed the deal on 27 April 2020 at $125 a share, about $7.0 billion in total, according to its filing with the SEC. Mellanox brought InfiniBand, high-speed Ethernet and network cards, the plumbing that lets thousands of GPUs train one model together. Huang summed up the logic at GTC this year: the data centre had to become a single unit of computing, “so Mellanox joined NVIDIA”.

From a side business to almost all of NvidiaData centre share of Nvidia’s quarterly revenue (%)Q4 FY2016quarter to Jan 20166.9%Data centre revenue: $97 million of $1.40 billion totalQ2 FY2027quarter to Jul 202692.5%Data centre revenue: $89.0 billion of $96.2 billion totalSources: Nvidia CFO commentary for Q4 FY2016 (SEC filing), Nvidia Q2 FY2027 results. Chart: TalkEsport
In early 2016, Nvidia’s data centre business brought in $97 million a quarter. Ten years later it earned $89.0 billion in three months. Chart: TalkEsport

The deal now looks like one of the best acquisitions in the history of the chip industry. Nvidia no longer reports networking separately, but The Next Platform estimates it brought in about $17.5 billion in the quarter to July, roughly a fifth of data centre sales. That means a single quarter of networking revenue is now worth about two and a half times what Nvidia paid for Mellanox. Everything else Nvidia sells, now reported as Edge Computing, made $7.2 billion in the same period, according to Nvidia’s second-quarter results.

Not every one of Huang’s big bets landed, however. Nvidia agreed to buy Arm for $40 billion in September 2020, then abandoned the deal in February 2022 after regulators in the US, UK and EU pushed back. Huang’s response says a lot about how he ran the company. Instead of giving up on CPUs, Nvidia designed its own Arm-based Grace processor, which now sits beside its GPUs in the Grace Blackwell systems that power much of the AI boom.

How Huang’s Nvidia compares with the companies it overtook

The market value tells the story in blunter terms. Nvidia became the first chipmaker worth $1 trillion in 2023, then the first company of any kind to reach $4 trillion in July 2025 and $5 trillion on 29 October 2025. It overtook Apple and Microsoft, two companies that had led tech for decades, while selling a product most consumers never see. Even after the selloff that followed his exit, which our chart of the scandal’s cost tracks day by day, Nvidia remains one of the most valuable companies in the world.

Huang also stayed in charge far longer than most founders do. He ran Nvidia for 33 years, from a booth at a Denny’s restaurant to an AI empire, through the dot-com crash, the 2008 crisis, the crypto mining boom and bust, and the AI wave. Under him, Nvidia went from $5.01 billion of revenue in the whole of fiscal 2016 to $96.2 billion in a single quarter this year. He also ran it with an unusually flat structure, with around 60 direct reports and almost no one-to-one meetings, by his own account.

Colette Kress, who is serving as interim chief executive while the board searches for a permanent successor, has already described her job as “building on the extraordinary foundation Jensen created over the last 33 years”. The candidates being discussed, which we covered in our report on who could replace Huang, inherit a company whose biggest bets he placed years before anyone else saw them coming. The results Kress presents this week cover the three months to late October, the last full quarter that Huang ran.

More from Mateo Rios

More on Business

- Advertisement -
Mateo Rios
Mateo Rios
Covers esports and gaming news from North and Latin America for TalkEsport.

Popular Articles