Elon Musk is reportedly seeking to raise funding for a new artificial intelligence startup that would compete with OpenAI LLC.
The Financial Times reported the development today, citing people familiar with the matter. An earlier report from Business Insider suggests that Musk intends to build a large language model. Many of the most advanced AI systems on the market, including OpenAI’s GPT-4, are large language models.
The news comes weeks after Musk co-signed a public letter that urged a pause on the development of neural networks more advanced than GPT-4. In the letter, the signatories asked the tech industry to halt development for at least six months.
Musk reportedly began planning the startup launch earlier this year. Since then, the Financial Times’ sources detailed, the Tesla Inc. Chief Executive has recruited a half dozen AI engineers to the venture. The hires are said to include Igor Babuschkin, a former employee of Alphabet Inc.’s DeepMind machine learning unit.
Musk is reportedly searching for not only AI engineers but also investors. According to one source, he is currently holding fundraising discussions with investors who have previously backed SpaceX Corp. and Tesla.
It’s unclear how much capital Musk is seeking to raise for the startup, but the sum is likely to be significant. Developing large language models such as ChatGPT is an expensive endeavor that requires large amounts of hardware. Earlier this week, Amazon.com Inc. Chief Executive Officer Andy Jassy stated that “really good” AI models can take billions of dollars to develop.
It’s believed that Musk’s AI venture may use content from Twitter to train its neural networks. According to today’s report, the training will be carried out using “thousands” of Nvidia Corp. graphics processing units. A recent report put the number of chips purchased by Musk at 10,000 GPUs.
It’s unclear what type of graphics card the purchase involved. There are significant performance differences between Nvidia’s GPUs. The chipmaker’s newest data center graphics card, the H100, can train large AI models up to nine times faster than the previous-generation A100 chip.
GPUs can’t be used on their own, but rather have to be deployed with supporting components such as central processing units and cooling equipment. Assembling such infrastructure requires a significant amount of effort. To spare customers the hassle, Nvidia sells appliances that combine GPUs with supporting components in a pre-packaged module.
Standalone GPUs can be more difficult to operationalize than appliances. However, not using pre-configured hardware building blocks may enable Musk’s AI venture to customize its AI infrastructure to a greater degree. Tesla already uses custom AI infrastructure to support its operations.
As of last year, the automaker operated not one but two supercomputers optimized for AI workloads. The first reportedly featured 7,360 of Nvidia’s previous-generation A100 graphics cards as of last August. Tesla’s other AI supercomputer is based on a custom chip, called the Dojo D1, that combines elements of a CPU with circuits optimized for machine learning.
The latter supercomputer also features other custom elements. Those elements include a proprietary networking protocol, dubbed the Tesla Transport Protocol, for sharing data between Dojo D1 chips. It’s possible that Musk’s new AI venture may implement some of the technologies developed by Tesla in its AI infrastructure.
The startup’s expected launch would increase competition in what is already a crowded market. OpenAI, the most prominent player in the nascent generative AI segment, recently received an investment from Microsoft Corp. that is reportedly worth $10 billion. Google LLC and Meta Platforms Inc. are also developing generative AI systems along with multiple well-funded startups.