Nvidia is betting big on the future, and that future is data centers built for AI, or “AI factories,” as Nvidia now calls them. Even a cursory glance at Nvidia’s latest finances makes a clear case for why: Its data center business made up almost 90% of its revenue in Q4 2024. Gaming, at less than 9%, is barely a footnote.
But Nvidia isn’t just looking to sell components to existing data center owners, it’s looking to help them build new ones, and claims there are already 100 of its newly envisioned AI factories currently under development around the world. These data centers, we’re told, will have the ability to run the latest AIs across a range of models and be equipped to scale up to future challenges.
This is the future Nvidia is pitching, and it wants that future to have an Nvidia logo, describing it as something that every enterprise will need to survive.
Redefining the datacenter
The data center industry is well-established, with major corporations like Amazon, Microsoft, and Google operating a range of highly capable, extremely expensive, and power-hungry farms of servers. While these facilities are designed to handle a range of tasks, though, Nvidia’s AI factories are said to be more focused.
Although Nvidia metaphorically describes its AI factories as “manufacturing intelligence,” what it’s really suggesting is that these massive data centers of the future will train and run AI exclusively. Although those two tasks often have different requirements of the hardware that powers them, the core component needed in both cases is masses of raw compute power. That means CPUs and GPUs, and Nvidia is at the forefront of both of those when it comes to powering AI.
Nvidia’s hardware brings power, but it also brings versatility, we’re told. Nvidia pitches these AI factories as capable not only of running one AI model, but multiple models, and even many at once. This would allow these factories to utilize the most capable AI model at any given time, or to train new models tailored to their own specific needs.
As Ian Buck, vice president and general manager of NVIDIA's Accelerated Computing business, said in a recent conference call, “That's why NVIDIA has got to work with every single AI company to make sure that our platform is constantly innovating.”
Nvidia doesn’t want to sell the claims to the gold rush; it’s selling the shovels. Or, perhaps more accurately, the automated drilling and processing machines, which it hopes will accelerate the extraction of wealth from the burgeoning AI industry. And the company says it’ll have a better one next year, too.
Performance at a price
Although gaming isn’t Nvidia’s main focus any longer, it was its bread and butter for the best part of three decades, and it’s benefited from an industry cadence that meant shiny and new products launched almost every year. That guaranteed a constant churn of hardware and customers, keeping it going long before it became one of the world’s most valuable companies.
It’s that revolving door of GPUs and AI hardware that Nvidia is hoping will usher in its next few decades of prominence. In much the same way as Bitcoin mining has been a digital gold rush, Nvidia believes its hardware can do the same for its customers looking to make the most of AI.
“There are two pieces of information that are very important,” Nvidia CEO Jensen Huang said during a recent chat at Computex. “The first is the reason why we upgrade every year. It's because in a factory, performance equals cost, and performance equals revenues.”
Whereas with gaming, Nvidia had to sell a product where the performance had to equate to something tangible – a new lighting feature, or support for higher resolutions – AI factories are Nvidia’s attempt to convert raw computing power into money. And Huang believes it can do that with equal parts power and efficiency.
“If your factory is limited by power, and our performance per watt is four times better, then the revenues of this data center increase by four times,” Huang said. “So, if we introduce a new generation, the customer's revenues can grow, and their costs can come down.”
He went on to discuss how software efficiency improvements give Nvidia hardware a long tail, or fine-wine-like performance and efficiency improvements – a little carrot to go with the threatening stick of competition outspending your enterprise to AI dominance. The CUDA infrastructure has long been Nvidia’s way to near-monopolize professional hardware-driven tasks, but it could prove a keystone of its AI factory strategy, too.
But while Nvidia will surely benefit from the consistent evolution of its technologies, always promising more performance, which for AI factories could mean more money one day, this isn’t some one- or two-year project. Nvidia believes it’ll be living and breathing the AI infrastructure rollout for the next half a century.
50 years of challenges
“AI infrastructure will cover the planet, just as internet infrastructure has covered the planet," Huang said. “Eventually, AI infrastructure will be everywhere. We are several hundred billion dollars into a tens of trillions of dollars AI infrastructure buildout that will take five decades.”
He mentioned this in a discussion about Taiwan’s infrastructure and Nvidia’s long-term plans to build and power such AI factories in the country. It has also announced partnerships on new datacenters with TSMC and Foxconn in Arizona and Texas, and with HUMAIN, in Saudi Arabia. It’s also announced other partnerships with Gigabyte, Asrock, Rack, Asus, Pegatron, Supermicron, Winstron, and many others, in on-premises and AI- systems using Nvidia GPUs. Those might not all be full AI factories, but the basics will be much the same.
These developments will all be built around tens of thousands of some of Nvidia’s most advanced hardware, like its GB200 and GB300 systems that combine Nvidia GPUs with Nvidia CPUs. But that will just be the start. It won’t be long until the Nvidia upsell begins, with next-generation Rubin graphics architectures promising the performance of multiple Blackwell GPUs, as part of its Kyber rack system.
But as with those Bitcoin mines, power and cooling will be a major concern. That’s why Nvidia is looking to revolutionise that, too. Nvidia plan is to switch away from evaporative cooling to closed-loop systems. The lower power draw of such a system could make that new hardware more efficient, saving power and allowing for greater net profits for the AI workloads.
That’s a hell of an upsell. Especially since Nvidia is hoping to make that pitch within the next few years. But it's one that will come at a huge cost, one Nvidia hopes its customers will pay time and again.
A bright, shiny, green future
Jensen Huang might be an engineer at heart, but he’s become an effective hype man for his company’s efforts over the years, and he’s no stranger to hyperbole and a hard sell. That 5070 never did match the 4090 performance claims, did it?
So while Nvidia promises such an exciting AI-driven future, and is pushing hard to build out the ground floor of it (with Nvidia-branded concrete that will only support future Nvidia office furniture), it’s important to remember that this future is a prediction, not a prophecy.
AI is still incredibly expensive and requires super expensive (largely Nvidia) hardware, except when Chinese companies do it for less, and some premier AI companies are operating at incredible losses in the 10-figure range. While there is no doubt that AI is going to revolutionize a number of industries, and may well one day (perhaps closer to the end of those 50 years) be the everything-tool, money-printing machine that Nvidia wants, we’re not there yet.
And that future is far from certain. Nvidia is selling a vision of it, and has made some very big and bold promises, and many of its partners and countries around the world are racing to get there first. But many feel there’s still a whiff of Dotcom bubbles, of NFT JPGs, of Blockchain promises, about it all.
Mr. Huang and his company are wagering everything on this vision. Branding the data centers its hardware largely powers as “AI factories” is one more chip it’s sliding onto the table.