Nvidia’s graphic processing units have set off a gold rush of the AI age
Jaspreet BindraMarch 2024 was a big month for Sunil Gupta, the founder-CEO of then little-known Indian cloud services provider Yotta. The fast-growing company was competing in the market with global giants in the cloud space, including the hyperscalars, and trying to carve out its own niche. Its ambitions were big, reflected in its name itself – a Yottameter is the term for 10 to the power 24, or 10 followed by 24 zeros. March 18 2024 catalysed that ambition as the company opened its gates to welcome the first cluster of over 4000 GPUs from Nvidia, making it the first Indian cloud provider to have these and offer a high performance AI cloud to eager customers. Word in the market was that it took conversations at the highest political levels with top management of Nvidia to make this happen.
If AI is a gold rush, Nvidia’s GPUs are the gold. Much has been written of the astonishing rise of Jensen Huang’s company, which has rocketed up from a $400bn company in 2021 to the world’s most valuable company at $3.25tn last month. This astonishing rise, unprecedented in technology, has been driven by one product: the Graphic Processing Unit (GPU). Nvidia first made these high-performance processors for gaming, as that industry boomed over the last decade. After dominating that world, Nvidia pivoted smartly when it realised that its chips could be the workhorses for crypto mining, as that industry exploded off the block about five years ago. But it was its second smart pivot, realigning its GPUs to feed off the AI and Generative AI frenzy, which rocketed the company past the $3 trillion mark. Today, Nvidia owns more than 80% of this fast-growing market, is far ahead of the competition with new version releases every year. It also makes its secret weapon: CUDA (Compute Unified Architecture), the software which manages and runs their GPUs.
This absolute dominance has made Nvidia’s GPU a scarce commodity, giving it absolute prizing power and leading large customers to hoard GPUs – both for their own future needs and to deny its competitors, and give them unparalleled competitive advantage. The market measures a company’s AI prowess with how many of the very latest A100 or Grace Hopper or Blackwell GPUs they have. Meta and Microsoft have spent a combined $9bn last year to buy 150,000 of them each, with Meta reported to have accumulated 600,000 of them, and Microsoft aiming to amass 1.8mn GPUs by end-2024. Google and Amazon bought 50000 each of them last year. The Chinese are denied the latest chips by US law but have been buying the earlier versions in hundreds of thousands, and smuggling in the latest versions through third-party countries.
It is not only BigTech; countries are amassing chips the way they used to buy and hold strategic reserves of oil. The US has placed restrictions on them towards unfriendly countries, so that it always leads the AI race with the most chips. Singapore drove 15% of Nvidia’s revenue in Q3, 2023, as it rushed to build AI data centres. Even India has ordered 10000 of them as a part of its India AI Mission. The list goes beyond countries. One of the most intriguing customer is the venture capital firm, Andreessen Horowitz
(A16Z), which has built of a stash of more than 20000 Nvidia GPUs. Kate Clark writes in The Information (https://bit.ly/3zKbVg8 ) about how it is using these H100 processors to bag the hottest AI startups as its investee companies. These small startups find it near impossible to find and buy this scarce commodity, without which they cannot build their AI models. A16Z steps in to invest, sweetening the deal with discounted or free access to the GPUs they have, thus making it a deal the startup cannot refuse. Case in point was super-hot GenAI startup, Luma AI, which chose A16Z for its chips, despite other investors offering them higher valuation. Other VC firms have cottoned on, with Index Ventures and Daniel Gross also procuring the ‘new gold’ to attract startups. The scarcity and resulting demand for these chips have reached ridiculous levels: a few years back and entire truck full of Nvidia GPUs was waylaid on a highway, and its contents stolen. Months later they turned up in Vietnam. There have also been instances of them being stolen from warehouses, forcing customers to deploy special security measures for them.
This is reminiscent of how gold was stored in the Fort Knoxes of the world as strategic reserves, and countries creating large strategic reserves of another precious commodity – oil. As data becomes the new oil, and AI the power that unleashes intelligence and information from this data, it is the chips and Graphic Processing Units, currently from one company, which is becoming the new gold in the age of AI.