The "shortage" of GPUs is starting to ease, with some early buyers who rushed to purchase now beginning to sell off their "high-priced inventory".

Wallstreetcn
2024.02.24 07:56
portai
I'm PortAI, I can summarize articles.

Over six companies using AI chips have stated that renting NVIDIA's most advanced chip (H100) from cloud providers has become relatively easy.

There are signs that the year-long GPU "shortage" issue is easing.

According to a report from The Information on Friday, more than six companies using AI chips have indicated that renting NVIDIA's most advanced chips (H100) from cloud providers has become relatively easy. As a result, some buyers who were frantically buying GPUs early on are now looking to sell off some of their inventory because holding onto them without using them would incur very high costs.

At the same time, current GPU buyers are reducing orders and becoming more selective about chip rental prices.

Miguel Solano, CEO of VMind AI, stated, "Three months ago, you couldn't get the H100 from Alphabet-C Cloud, but now you can." VMind AI is a company that helps large-scale companies reduce computing costs and improve chip efficiency.

Several factors may be contributing to the decrease in demand. According to sources, Amazon has introduced a new service that allows customers to rent GPUs on a daily or weekly basis. Previously, customers could only rent GPUs for longer contract periods or on-demand. A spokesperson for Alphabet-C stated that the company can "meet almost all customer needs and continuously provide new capacity."

Additionally, as businesses become more familiar with chip usage, some companies may lower their expectations of how many GPUs they actually need. An employee of a GPU rental company mentioned feeling more pressure to ensure their business is economically viable.

All of this contrasts sharply with last year when startups and investors were scrambling to get NVIDIA GPUs.

Sarah Guo, founder of Conviction, a venture capital firm focusing on AI startups, said, "There was a sense of panic and scarcity at that time. By mid-2023, the big model community was all talking about one term - 'GPU Poor'."

Of course, not everyone has resolved the GPU "shortage" issue. Companies like OpenAI, which train large models, are still struggling to obtain enough AI chips.

During the earnings call on Wednesday, NVIDIA CEO Jensen Huang stated that the supply constraints faced in recent quarters are "improving." However, in a recent interview, he mentioned that there will be a supply shortage of GPUs this year and even into next year.

It is worth noting that NVIDIA's competitor Intel's CEO Pat Gelsinger stated on Wednesday that while he does not feel any slowdown in demand for AI chips, he expects the company to be more cautious in its investments in AI chips.

Gelsinger said:

"You can't invest hundreds of billions of dollars in these large data centers without starting to build an economic model to support these data centers."