Last week I warned about the energy use of Nvidia (NASDAQ:NVDA) AI chips.
The stock shrugged it off.
Shares that sold for $118 as recently as June 24 are trading at $134 per share today.
The worries are not being reflected in the stock price because they’re not yet hitting sales. Nvidia dominates the data center and it is taking almost half that revenue to the bottom line.
That’s why some analysts still consider Nvidia’s price-to-earnings (PE) ratio of 75 a bargain. If it can keep growing sales and profits at the current rate, it will be one-third of that value in a few years.
Changing of the Cloud Guard?
So long as the old cloud guard keeps spending, the thinking goes, Nvidia will keep thriving.
There is a second catalyst for Nvidia bulls. Nvidia chips are so powerful that companies that missed the Cloud Boom of the last decade can now become part of the new cloud guard. Oracle (NYSE:ORCL) is now a Cloud Guard member although its valuation has yet to reflect that. Elon Musk is making X, formerly Twitter, into a new Cloud Guard company. This is reflected in the stock price at Tesla (NASDAQ:TSLA), but it isn’t a member.
Around the world, companies that missed the cloud are reevaluating their positions. They’re wondering if they have enough cash to join the cloud wars. This includes application companies like Salesforce.com (NYSE:CRM) and cloud vendors like IBM (NYSE:IBM). It could include a host of what are now small European cloud players.
If anyone can build a cloud, using some Nvidia server racks put together by Super Micro (NASDAQ:SMCI) or Dell Technologies (NYSE:DELL), the technology world turns upside down.
This also plays directly into Nvidia’s hands. That’s because, as I wrote last November, Nvidia controls the cloud stack, thanks to its Cuda software. Cuda was developed almost 20 years ago for nascent robotics applications. It’s at the heart of AI model development today.
The AI War
There’s a war on right now. The good kind.
It’s the race to build compelling large language models (LLMs) that can replicate how thinking people think (as opposed to how stock analysts think).
Current models can cost $1 billion to train. Within a few years, models costing $100 billion will be in development.
As AI models become more powerful, the thinking goes, the hardware required to train them must also become more powerful. Chat GPT-4 will require 30,000 Nvidia GPUs to train. Tomorrow’s models will require even more.
How will anyone earn a profit from this investment? Anthropic CEO Dario Amodei said on a recent podcast they will do it as LLMs turn into automated general intelligence (AGI) engines that can think and not just do research.
It’s this rush to an automated brain, even an automated super brain, that has researchers and investors excited. IBM’s DeepMind was able to beat chess champion Garry Kasparov almost 30 years ago. What happens when it runs a CEO’s company better than the current C-suite can?
The Bottom Line
The real bottom line is paying the energy cost of running AGI models at scale and getting paid enough for the resulting services to make the whole thing profitable.
This is where AI will go through its “trough of disillusionment” and eventually find its “slope of enlightenment” and “plateau of productivity.”
How Nvidia data centers deal with their energy use will determine which clouds succeed and those that fail. Can they get enough power? How much of that power can they recycle, turning heat back into electricity?
Cooling data centers isn’t enough. If that energy can’t be recycled and put to use, Nvidia and the whole society will be in trouble.
Buying Nvidia gear is easy. Running it profitably will be harder. That’s what concerns bears, and it should concern bulls, too.
As of this writing, Dana Blankenhorn had a LONG position in NVDA. The opinions expressed in this article are those of the writer, subject to the InvestorPlace.com Publishing Guidelines.
On the date of publication, the responsible editor did not have (either directly or indirectly) any positions in the securities mentioned in this article.