Dour warnings of an AI bubble have rocked markets in recent weeks. At least one big concern is misplaced, though.
Back in March, I told you about depreciation risks for some AI companies, including CoreWeave. In August, Jim Chanos, the guy who shorted Enron, shared similar concerns.
The big worry centers on GPUs, the chips needed to train and run AI models. As new GPUs come out, older ones get less valuable, through obsolescence and wear and tear. Cloud companies must use depreciation to reduce the value of these assets over a period that reflects reality. The faster the depreciation, the bigger the hit to earnings.
Investors have begun to worry that GPUs only have useful lives of one or two years, while cloud providers depreciate the value of these assets over five or six years. An accounting mismatch like this could set the AI industry up for a nasty earnings hit in a few years.
This view has become almost a consensus on Wall Street now. It’s one of the main pieces of evidence for the argument that we’re in a huge AI bubble. The problem is that it’s wrong: Even as Nvidia rolls out new GPU architectures every 18 months or less, GPUs aren’t aging out nearly as fast as some investors fear.
“GPUs can profitably run for about 6 years,” Stacy Rasgon, a leading chip analyst at Bernstein, wrote in a research report on Monday. “The depreciation accounting of most major hyperscalers is reasonable.”
Healthy margins
The cost of operating a GPU in an AI data center is “very low” compared to market prices for renting GPUs via the cloud. That makes the “contribution margins” of running old GPUs for longer quite high, Rasgon and his fellow analyst at Bernstein noted. (Contribution margins measure revenue left over after variable costs. It’s a common way product profitability is assessed and business decisions are made).
“Even with meaningful improvements in price/performance with each GPU generation, vendors can make comfortable margins on 5-year-old A100s, in turn implying a 5-6 year depreciation lifespan is reasonable,” the analysts added, referring to Nvidia’s A100 chips, which came out in 2020.
Seven to eight years
To find out why these GPUs are so valuable for so long, it pays to speak with the people who actually run these components at scale inside AI datacenters.
Matt Rowe, senior director of strategic business development at AI cloud provider Lambda, said recently that the effective lifespan of GPUs can stretch to seven or eight years.
While most firms still use a six-year depreciation schedule for accounting purposes, warranty extensions and redeployment strategies are extending their useful life, he told Bernstein.
Warranty contracts are often overlooked by observers worrying about depreciation, Rowe explained. These warranties typically last five years, so if GPUs fail, they are replaced with new ones, extending the life of the overall GPU fleet.
He also noted that Amazon Web Services offered very early generations of GPUs, such as K80s, P100s, and V100s. These all lasted well beyond six years.
Nvidia’s H100 GPUs, which debuted in 2022, are still running well inside Lambda data centers. Utilization is above 85% and Lambda hasn’t cut its on-demand public cloud pricing for this GPU in more than 12 months, Rowe noted.
“We all think seven to eight years is possible,” Rowe said.
Crusoe’s experience
I chatted this week with Erwan Menard, SVP of product management at Crusoe, which is developing the huge Stargate data center complex in Texas. Before joining Crusoe, Menard helped build Google’s Vertex AI cloud service, so he’s a real hands-on expert.
Menard described a lifecycle where GPUs migrate from cutting-edge AI model training jobs to less demanding inference workloads.
When creating a new state-of-the-art model, you need the latest and greatest GPU from Nvidia.
Then, you have to run these top models, a process called inference. That requires powerful GPUs, but not the latest ones.
Beyond that, there are thousands of different, valuable AI workloads that can run well on older GPUs, according to Menard. That means there are many GPUs that are multiple years old in Crusoe’s fleet and are still actively used and profitable.
“Because there’s a large diversity of models to solve many different problems, there’s a lot of room to use GPUs for a long time, just transitioning them from one type of job to the next,” Menard told me. “It’s actually a widely accepted view in the industry.”
Free versus paid
AI cloud companies consider user expectations and budget to help them decide which GPUs to use. To illustrate, Menard described an example of an AI service that has a free tier and a paid version.
“You may decide that for the freemium version you’re going to use an AI model that can be inferenced on older, cheaper hardware with lower performance,” he said.
That’s likely good enough to create an initial experience for users. Then, some customers might migrate to the paid version. At that point, you tap into a more powerful AI model that requires newer GPUs to deliver a superior user experience.
“We see a lot of these opportunities,” Menard said. “Not everything is a nail requiring one single mega-model running on the latest and greatest GPU.”
Open-source + older GPUs
Some AI services are less compute-intensive and can be run on open-source models, such as Alibaba’s Qwen, DeepSeek, or Meta’s Llama offerings. One example is speech-to-text services (such as the transcription service I used to transcribe my interview with Menard).
Older or less-capable models can be run on older GPUs, while still providing valuable intelligence for AI services that customers will pay for. (Business Insider pays for those transcriptions, for instance).
As more startups embrace cheaper open-source models, older GPUs could actually be used even more. “An open model may be absolutely great and give a more cost-competitive structure,” Menard said.
Older GPUs are cheaper
Older GPUs use more energy to produce the same amount of intelligence, so another investor concern is that newer GPUs will always be preferred—aggravating this depreciation problem.
That’s actually not true either, according to Menard. Older GPUs are cheaper to buy, so the fact that they consume more energy doesn’t change the fact that older GPUs are often cheaper to run, when all costs are taken into account.
“The driver for a given GPU is going to be cost, first and foremost,” he explained. “So we go to the older ones because they’re cheaper.”
What’s an L40?
So, I asked Menard for an example of an old GPU that Crusoe uses. He described new modular data centers Crusoe has built that are powered by recycled EV batteries from the startup Redwood Materials.
“I can put L40s from Nvidia in these data centers,” Menard said. “Because the whole deployment is energy-first in its design, I’m going to be able to make an impact.”
I hadn’t heard of L40s and had to ask him what they were.
“That’s an old GPU,” he said, laughing.
Sign up for BI’s Tech Memo newsletter here. Reach out to me via email at [email protected].