How much is an old AI chip worth?
The answer to that question has significant financial and technical implications. Cloud-computing player CoreWeave recently suggested it could be a lot more than many people have been assuming.
The company recently signed a deal renting out Nvidia's A100 chips at "an attractive price" in a deal lasting into 2029, CoreWeave finance chief Nitin Agrawal said in a call with analysts last week. Those chips came out in 2020 and were produced for about three years. "Increasingly, we are seeing longer utilization at higher prices," he said, "offering the potential for significant further upside."
That could be a big deal because companies typically depreciate their chips over four to six years. That reduction in value has to run through their income statements, hitting profit.
Yet the drag on the bottom line ceases after chips are fully depreciated. This means that selling access to old chips could come with high profit margins-as long as the price companies can charge doesn't go down too much.
A nine-year-old chip still running and generating sales also gives some credence to the idea that chips are viable assets to lend against. Banks and asset managers might be willing to extend credit to CoreWeave and other cloud-computing operations at cheaper interest rates, confident their chip collateral won't become worthless for quite a while.
Of course, there are counterpoints to the notion that chips can have lasting value.
The first is that many AI chips fail. Data on how frequently they fail is scarce, but a Meta Platforms report in 2024 on a 54-day training run for its Llama 3 model suggested the failure rate could be around 9% a year. Assuming that rate accelerates as time goes on, a nine-year-old AI computing cluster could easily be well below half its initial size because of failures.
The viability of an old chip is also increasingly questionable as the rush to build out AI data centers runs into power constraints. Newer generations of Nvidia's chips are orders of magnitude more energy efficient than the A100.
That means that if the availability of electricity comes into question, abandoning old chips may become a no-brainer. Companies may judge it more important to harvest more AI computational work from new chips than to boost profits by running old, less efficient ones.
CoreWeave's statement about its A100s also leaves an important question unanswered: What, exactly, is it able to charge? While Agrawal described the price as attractive, it is unclear what he meant by that. "Attractive," after all, is in the eye of the beholder.
And that, more than anything, will determine what value older chips retain.
This is an edition of the WSJ AI & Business newsletter, a weekly digest to help you make sense of AI's impact on business with news, insights and data from our global team of technology journalists. If you're not subscribed, sign up here.
Nvidia Hands OpenAI a $105 Billion Lifeline
Nvidia is in the business of selling chips. It is also increasingly in the business of making sure customers can afford to buy its chips.
That appears to be the purpose behind the company's residual-value backstop for OpenAI's lease of a huge data center complex being built in Ohio. OpenAI is leasing the facility for 20 years and plans to fill it with Nvidia's chips. But OpenAI isn't a profitable company, and its creditworthiness is too questionable for the project's developer-a unit of Japan's SoftBank called SB Energy-to go forward without some outside support.
And that is where Nvidia comes in, agreeing to be on the hook for the lease if OpenAI defaults-to the tune of up to $105 billion.
The Number
Value of leases not yet started among tech's four largest AI spenders, an amount not reflected on their balance sheets.
What the Humans Are Saying
AI in Charts
Open-weight models are getting smarter.
That is one of the messages to come out of the recent proliferation of AI models from Chinese companies that offer strong performance coupled with cheaper access than equivalents from Anthropic, OpenAI and others.
But while open-weight models' ascendance has caused a stir among investors worried about their impact on AI spending, it remains to be seen whether they will have a real impact.
One problem with more advanced open-weight models is that they cost a huge amount of money to train, just like closed-weight counterparts that aren't freely available to use. So how do open-weight model developers recoup those outlays? That is less clear than in the closed-weight world, where access to models is entirely dependent on the developer.
Apart from business-model concerns, open-weight models from China face political headwinds. The U.S. has sought to crimp China's development of advanced AI systems, and the emergence of a new breed of cutting-edge models from the People's Republic could invite a clampdown.
AI in the Wild
The U.S. manufacturing sector is enjoying the fruits of the AI boom. Companies like Caterpillar and Cummins are raising production of generators critical as backup-power options for data centers springing up all over the world. Ford Motor is redirecting excess electric-vehicle battery production to the data-center boom. Eaton, a maker of circuit breakers and other electrical equipment used in data centers, is seeing a sales bump as well.
Other Highlights From the Week in AI
The U.S. is urging Apple not to buy memory chips from China.
Uber Eats plans to start drone deliveries.
Chinese AI developer DeepSeek is raising prices fourfold.
About Us
WSJ AI & Business is a weekly look at AI's transformation of the business world. This newsletter was curated and edited by Asa Fitch. Reach him at asa.fitch@wsj.com (if you're reading this in your inbox, you can just hit reply). Got a tip for us? Here's how to submit.
Comments