Last month, the latest inference chip for cloud-based large models introduced by the American artificial intelligence startup Groq has sparked widespread attention in the industry.
The chip adopts the innovative Tensor Streaming Architecture (TSA) and boasts ultra-high-bandwidth SRAM, resulting in over a 10x improvement in inference speed for large models, even surpassing NVIDIA's GPUs.
This has also piqued the interest of some customers in Groq's AI chip.
However, according to recent reports, Groq CEO Jonathan Ross has accused NVIDIA of hindering fair competition in a recent interview.
"Many people we've talked to have said that if word gets back to NVIDIA that they met with us, they would be in trouble."
"The issue is that you have to pay NVIDIA upfront a year in advance, but the delivery might be a year or more out. They'll say, 'Oh, you went with someone else, I guess our delivery might slip.'"
Jonathan Ross remarked.
This seems to imply that even tech giants like Microsoft, Google, and Amazon, who are developing AI accelerators, would insist they don't intend to compete with NVIDIA.
NVIDIA alone earned over $9.2 billion in revenue from Google, Microsoft, and Amazon last quarter, which accounted for over a quarter of their capital expenditure during the same period (approximately $35 billion).
Microsoft has stated that its custom chips complement NVIDIA's and are not intended to be replacements, aiming to provide customers with the best options for price and performance.
Insiders familiar with the matter also revealed that Google is offering both its own chip and NVIDIA chip solutions, allowing customers to choose based on their budget and technical requirements.
While Amazon has its own AI chips, it still emphasizes its expanded long-term partnership with NVIDIA in 2023, claiming to offer the widest variety of chips for cloud services.
