So, in recent years, especially during this golden age of AI, it seems like a new superhero has emerged. It’s a bird! It’s a plane! No, it’s a bespectacled man in a leather jacket! It’s Jensen Huang, CEO and founder of Nvidia.
And while Nvidia was founded in 1993, in just a couple of years, it’s the most valuable publicly-traded company in the world. It was able to get there, because Nvidia designs and manufactures advanced GPUs (Graphics Processing Units). You may have heard of the word GPU without actually understanding what it means. Imagine a Pringles chip that could help render images, train AI models, process videos or help your device display graphics smoothly. Obviously, the Nvidia chips are not Pringle-shaped, that’d be cool, but they’re more rectangular or square. Since AI became all the rage, Nvidia became a darling of the world and Huang was a new emperor.
Nvidia has a chip called Blackwell AI. These are said to be 30x faster than usual at tasks, like providing responses from chatbots. How is a Pringle chip even doing that? In case you’ve used an LLM, it is trained to perform super complex math computations quickly and efficiently by going through patterns in substantial datasets. It’s not like the LLM is going to Google and copy-pasting an answer; it’s processing the input humans give, calculating probabilities for potential responses and then, generating the most likely or most accurate response. It’s a lot of work, apparently. What the Pringles chip does is perform those calculations super fast using all kinds of tech mumbo-jumbo. For something this cool, the customers were said to be small-time companies, like Microsoft, Google and Meta.
Though, in late 2024, something intriguing happened. The Blackwell chips were said to have overheating issues. There were, also, initial production delays, which were said to have frustrated its customers, who wanted to implement Blackwell capabilities into their data centres. So, what’s with the overheating? Nvidia is said to keep the chips in a server rack: up to 72 of them. But, unfortunately, when the chips are in their server racks, heat may not be effectively managed. So, Nvidia may be looking into how their racks are designed to deal with the overheating.
Is this just part of the iteration process? Is this how experiential learning takes place? Where failures are not setbacks, but stepping stones? Nonetheless, this is something that would leave Nvidia customers worried, because to build a data centre, it requires substantial investment. Plus, Nvidia’s customers are all in the middle of an AI race and each wants to get ahead. Nvidia’s GPUs are the de facto benchmark and go-to to power this AI race. Unfortunately, it looks like market dominance doesn’t make you immune to the occasional operational challenge.
But, the aspects that make Blackwell revolutionary, like its computational power and design, those might, also, make it vulnerable to challenges and new issues. The more powerful the chip, the more energy it consumes. The more energy is consumed, the more heat is generated. So, there’d have to be advanced cooling solutions. Bitcoin mining is facing a similar issue.
While Nvidia is probably working overtime to fix the overheating issue, could it create a temporary vacuum in the AI hardware market globally? Does India have a shot here? They only get one shot, they can’t miss their chance to blow, this opportunity comes once in a lifetime.
It’s said that if AI is a gold rush, Nvidia is selling shovels. Now, if servers represent the gold rush in 2025, could thermal management or liquid cooling tech be the new shovels? Something meant for high-density server environments? Or are there platforms out there that could monitor or predict or manage the thermal performances of servers live?
So, if Nvidia is temporarily having ACL issues, who’s going to be the new quarterback? Who’s stepping up? Or will Nvidia rise from the ashes again? Flame on.
