Is Nvidia going to monopolize the AI chips market for the next 10 years? Good question.
The answer is yes, probably, based on many months of reporting and Fierce Electronics’ interviews with a steady stream of analysts and industry officials. Nvidia will dominate in GPUs for at least two years, for sure, despite best efforts of AMD and Intel, who are distant followers. There are others building GPUs on their own, including hyperscalers and startups.
The Futurum Group this week released stats showing Nvidia held 92% of the market share for GPUs in early 2024, while noting that the GPU market makes up 74% of chipsets used in AI applications within data centers, a number expected to grow by 30% a year over the next five years. That market was $23 billion in 2023, a number expected to reach $102 billion in 2028.
So, by any measure, Nvidia is huge. Why?
The main reason Nvidia is so huge is that Nvidia has been effectively first to market, reaching back to its early days in using GPUs to boost gaming. First to market almost always wins the day. Also, customers see Nvidia as reliable, but the other side of the coin is that Nvidia makes CUDA software, a proprietary computing platform that some customers and analysts see as locking them into a long-term relationship with Nvidia.
A long-term relationship with Nvidia could be very beneficial,however, especially if a bank or insurance company found a way to use GenAI on Nvidia GPUs, like a monstrous Blackwell cluster, to be the first mover on a very beneficial AI application. Banks clearly see first mover advantage. They are willing to buy H100s or other GPUs in massive numbers, even 10s of thousands. (Or leasing that many for a short time from a big data center.) Cloud providers are doing the same.
Given Nvidia’s growth in the past year to the very top of market capitalization, customers are increasingly worried about the cost of all this. It is not uncommon to hear chip resellers charge $35,000 for a single GPU, with a higher end cost of $60,000. A single next-gen Blackwell chip is now priced around $70,000, although CEO Jensen Huang said back in March it would be as much as $40,000. (But Blackwells are not sold as solo GPUs and are only in pre-built systems, so pricing is difficult to establish.) Raymond James has said Blackwell B200s will cost Nvidia $6,000 or more apiece to make (that number would be BOM and not the intellectual property up front cost), which gives Nvidia plenty of room for profit.
TSMC has said, politely, it was charging Nvidia far too little for making Nvidia’s GPUs. “I did complain to Nvidia’s CEO Jensen Huang—the ‘three trillion guy’—that his products are so expensive," TSMC Chairman C.C. Wei said in June. “I think those products are really valuable for sure, but I am thinking about showing [him] our values as well.”
Smart customers of Nvidia are clearly aware they are paying plenty for Nvidia GPUs, but as with any newer tech, it is not the price that matters (within reason), but the value. Smart companies will create a plan to find alternatives to Nvidia to avoid long-term lock-in.
Yes, Nvidia deserves credit for advancing GPUs and CUDA (and networking and server racks with cooling), but reports have surfaced Nvidia may have gone too far. The New York Times reported in early June that the US Department of Justice and the Federal Trade Commission reached an agreement on how to investigate antitrust claims against Nvidia, OpenAI and Microsoft, with the DOJ taking a lead on investigating whether Nvidia has broken antitrust laws that oversee fair competition in business.
Given antitrust scrutiny is inherently complex and time-consuming, it is obvious Nvidia will keep marching ahead with market dominance for years to come. Why, you might ask, would it matter to the DOJ if Nvidia is a monopolist? Well, keeping competition at bay is one reason, but also because the Nvidia GPU treadmill makes everything dang expensive. Separately (but not completely separately), Nvidia is apparently still on course to alter its chips sold to China to get around US trade restrictions, despite vocal concerns of the US Commerce Department.
Nvidia rarely talks about trade or legal matters. While it holds its next earnings report on Aug. 28, don’t expect much to be said about antitrust or trade. Part of the focus could be on a three-month delay for Blackwell production, but the financial community is still impressed and lately the Nvidia share price climbed from below $100 on August 5 to $128 on Aug. 21.
Back to the main point: Nvidia is huge and destined to be even bigger in coming months but does that mean it is a monopolist? Ryan Shrout, president of Futurum's Signal65 largely agrees that Nvidia has had the early advantage, although he would only commit to saying the company has at least five years before anyone else comes close to competing for the same size of market share. Here's what Shrout said in total (and thanks Ryan for weighing in!) :
"Nvidia has more than a decade of R&D lead time on the competition here, moving from just graphics to general purpose compute on their GPUs, to ML and AI. That lead allowed them to develop the software stack, and the supporting infrastructure toolset like NVLink, that helps maintain the moat they have versus the competition.
"That doesn't mean they'll be the leader for 10 more years, but I would say it's at least 5 years before anyone, from AMD to custom silicon from the hyperscalers, is close to competing for the same kind of market share. As the AI compute standards continue to commoditize, continue to form into broader software standards and development tools, the value of differentiated hardware shrinks.
"Though Nvidia has the clear leadership position today, it seems improbable that it will hold this kind of market share leadership long enough to worry about being called a true monopoly."