- Everyone involved in AI seems to be obsessed with Nvidia's chips.
- All except Sharon Chow.
- Lamini AI's CEO has been using rival AMD's GPUs to propel his startup forward.
CEOs of technology companies with big plans for artificial intelligence spent much of the last year scrambling for Nvidia chips.
The Santa Clara giant's chips, known as GPUs, have become the hottest assets of the generative AI boom. Powerful figures like Mark Zuckerberg and Sam Altman competed to secure the critical computing resources needed to run apps like ChatGPT.
But there's one AI boss who isn't at the mercy of Nvidia's billionaire leader Jensen Huang and his $2.2 trillion GPU empire. Meet Sharon Chow.
At 30 years old, he has had quite the career.
She is the first person to major in both classics and computer science at Harvard University. She received her Ph.D. After researching generative AI under machine learning pioneer Andrew Ng at Stanford University, he became an adjunct professor there and has devoted his time to online education and angel investing. If that wasn't enough, she was also asked by Amazon to join the initial founding team of her OpenAI rival Anthropic, which she just raised $2.75 billion.
But her ambitions have taken a slightly different direction, and she is now carving her own path as the head of her own AI startup.
Who needs Nvidia?
Last April, Chou and co-founder Greg Diamos brought Palo Alto-based startup Lamini AI out of stealth. Its primary goal was to provide a platform that allows companies to easily train and create large, customized language models with “just a few lines of code.”
That could mean taking a foundational model like GPT from OpenAI and allowing companies to easily fine-tune that model using their own data. “What we're doing is basically enabling any company to have their OpenAI infrastructure in-house,” Zhou said.
But an equally interesting fact emerged a few months later.
In September, Chou said Lamini's platform had been working with customers for the past year using only GPUs from Nvidia's main rival, AMD, the chip giant run by Huang's cousin Lisa Su. He revealed that he was building a customized LLM.
This was a big problem considering almost everyone seemed obsessed with just the H100, a GPU that Nvidia has struggled to meet demand amid supply constraints. Ramini's revelation was accompanied by a video of Chou teasing Nvidia about the shortage.
But as Chou admits, it wasn't an easy decision to turn away from what everyone has longed for in generative AI. “The decision-making process was a long one,” she said. “It wasn't a small, small thing.”
A few things helped me decide. First, her co-founder Diamos played a key role in realizing the realization that non-Nvidia GPUs work perfectly fine.
As a former Nvidia software architect, Diamos co-authored a paper on the Law of Scaling that demonstrated the importance of computing power, while GPU hardware is essential to extracting the best performance from AI models. One Diamos knew that software was important. is also important.
Diamos witnessed this while working on CUDA, a software first developed by Nvidia in the 2000s. He will be able to use his AI models with GPUs such as the H100 and Nvidia's new Blackwell chips as easily as a plug-and-play system.
So it became clear that if another company could build a similar software ecosystem around its GPUs, there was no reason why it couldn't compete with Nvidia. Fortunately, after consulting with Diamos, AMD was building a rival system that it would eventually test, Zhou said.
“This took us a number of years because Greg and I were just tinkering with it, but once we had a working prototype, we decided to try and make it even better,” Chow said.
More broadly, Zhou acknowledges that enterprises are “very excited about using LLM,” but many will wait until Nvidia ramps up the supply of GPUs enough to meet demand. Maybe you don't want to wait, or maybe you just can't afford to wait.
That's another reason why AMD has proven so valuable to her ambitions. Thanks to the increased availability of GPUs, Zhou was confident that Lamini would be able to provide the “infrastructure to meet the surging demand” for her LLM.
“This is because Lamini fully leverages LLM computing with 10x performance and provides vendor-neutral computing options, allowing you to scale quickly without supply constraints. Because customers don't realize they're running Lamini on Nvidia and AMD GPUs,” she explained.
No wonder the company is willing to double down on AMD. In January, Zhou shared with X an image of the MI300X (AMD's new chip, first announced by CEO Su as “the world's most powerful accelerator” in December), actually being produced at Lamini.
Nvidia's Huang may now lead one of Silicon Valley's most powerful companies, but competition will be coming for him. Or, as Zhou said of his AMD, “They have a real horse in this race.”