Nvidia’s Aggressive Sales Tactics Will Backfire, Says Rival
The Takeaway
• Tenstorrent CEO says gigantic AI data centers might not be a winning move
• Companies could get stuck with unnecessarily expensive hardware
• Nvidia’s Cuda software is a ‘swamp, not a moat’
The U.S. Department of Justice is investigating accusations by customers and rivals of Nvidia, including Advanced Micro Devices, that the artificial intelligence chip designer is behaving in an anti-competitive way, including in how it bundles its products.
One longtime rival argues that while Nvidia’s sharp edged tactics don’t appear to be illegal, they will eventually hurt its business.
“When people start complaining to you about what you’re selling them, that seems like a bad thing,” said Jim Keller, CEO of AI chip developer Tenstorrent and a former chip executive at AMD, Intel, Tesla, Broadcom and Apple.
Customers feel pressured to buy Nvidia’s networking gear in order to guarantee themselves access to the company’s vaunted AI server chips, for instance, causing resentment, he said.
“I’ve talked to many people who are seriously not happy about that, because it’s not what they wanted,” Keller said in a conference call with The Information subscribers on Tuesday. “Fundamentally, you should be in business to make the best technology…and then people love to pay for that.”
Still, he defended Nvidia’s “enviable position, which isn’t” illegal.
“They invested for 10 years before AMD even thought about” graphics-focused chips, which now power everything from automated driving software to ChatGPT, he said.
An Nvidia spokesperson said in a statement that “Nvidia wins on merit, as reflected in our benchmark results and value to customers, and customers can choose whatever solution is best for them.” An AMD spokesperson didn’t immediately respond to a request for comment.
Keller’s current company, Tenstorrent, builds chips and software for training and running AI models. In contrast with Nvidia’s graphics processing units, Tenstorrent’s chips are built with multiple small central processing units in each chip that can independently decide what data to process or deprioritize. That means they are cheaper and less power-hungry than Nvidia’s GPUs, Keller has said.
Additionally, Tenstorrent differentiates itself from other chip designers by using RISC-V, an open-sourced chip design that competes with those made by Arm and Intel. The startup also licenses its technology to companies such as LG and the Leading-Edge Semiconductor Technology Center, a Japanese government-sponsored research center.
Founded in 2016, the startup has raised more than $380 million to date, and was recently in talks with Samsung to raise at least $300 million at a $2 billion pre-investment valuation.
But Tenstorrent and its ilk face a tough battle against Nvidia. Some of Nvidia’s dominance comes from its software for writing machine learning apps, Cuda, which only works with Nvidia’s own chips—making app developers less likely to switch to alternatives.
Keller said Cuda’s permanence is far from guaranteed.
Cuda “is a swamp, not a moat,” he said, because the software has become so complicated to use.
Cuda has “lots and lots of libraries and code, and so when people say, ‘well, just run Cuda,’ it’s like, well, have you ever looked at it?” Keller said.
As a result, he said, Cuda could go the way of Unix operating systems, which were replaced by easier-to-use, open-source Linux software in the 2000s. That move contributed to the downfall of leading companies including IBM, HP, Sun Microsystems and Silicon Graphics, he said.
Already, startups like Modular have emerged to try to weaken Cuda’s grip and make it easier for developers to run and train AI models on different types of chips.
Supercomputer Race
During the subscriber call, Keller also questioned the effectiveness of some of the AI-focused data centers, also known as supercomputing clusters, that companies such as Microsoft, Meta Platforms and OpenAI are planning to launch as a way to improve the capabilities of AI.
Although he agrees with AI “scaling laws”—a belief that the more compute power and data used to train AI will lead to faster improvements—he said that gigantic, expensive data centers might not be the winning move.
Rapid improvements in AI hardware and software could soon lead to new AI models that require less computational power, leaving companies stuck with unnecessarily expensive hardware, he said.
“There could be a model published next month that takes computation [costs] down ten times,” Keller said. Decades ago, “Sun, Digital Equipment and Silicon Graphics all made small computers that were easy to use and they got bigger and bigger and more and more expensive, and then the technology shifted and [the companies] literally all went bankrupt.”
With growing skepticism about the returns AI developers will generate from their investments in developing it, he said the industry could also end up in a similar situation to the internet bubble, in which millions of miles of fiber-optic cables were left gathering dust.
“Semiconductors are cyclical,” Keller said. “Every time there's an upcycle, sooner or later, somebody builds too much capacity. And then when there's a crash, everybody says, ‘you should have seen that coming.’”
He added that “this kind of [asset] bubble, the shape is recognizable which means change will come, but the thing no one can predict is when,” he said.
The biggest developers are also likely willing to lose a lot of money in an attempt to capture as much market share as possible, which could also artificially prolong the AI bubble, he said.
“There’s already a gap between what people are paying to train models versus what they're charging [customers to use AI products]...they’re already over their skis.”