According to industry rumors,After Microsoft cut off 20% of its orders in the third quarter, it was revealed that the NVIDIA GPU AI accelerator card had once again been cut off by a large cloud service provider. It is said that it was also Microsoft., NVIDIA also had to lower its order volume from TSMC in the second half of 2024. Regarding the reasons for order cancellation, the general view of the supply chain is thatAMD's newly released InstinctMI300 series is very competitive, and customers don't want to put all their eggs in NVIDIA's basket.

The AMDMI300 series includes the GPU accelerator MI300X and the CPU+GPU fusion accelerator MI300A. The latter is an exclusive technology with outstanding performance. Some indicators have far exceeded NVIDIA competing products, and the price is lower as usual.

At AMD's press conference, in addition to the two giants Google and Amazon, Microsoft, Oracle, Meta (Facebook), Dell, HP, Lenovo, AMD, etc. all supported the platform and related products were released immediately.

AMDMI300X has been shipped, and MI300A has also been put into mass production, which is expected to meet the industry's demand for AI large model training and inference.

NVIDIA currently holds 90% of the AI ​​​​acceleration market share, but product prices and costs are getting higher and higher, which makes many large companies unable to bear it and are looking for alternative solutions. The AMDMI300 series is undoubtedly the best choice at the moment, except that the ROCm development ecosystem is far worse than CUDA, and migration is a hassle.

AMD now predicts that the data center AI accelerator market will reach US$45 billion in 2023, with a compound annual growth rate of more than 70%, and will exceed US$400 billion in 2027, a significant increase from the previously expected US$30 billion, 50%, and US$150 billion.

In this market, even if AMD only captures 10% of the market, it will generate tens of billions of dollars in revenue every year.

However, some people believe that Microsoft's continuous order cuts are waiting for NVIDIA's next-generation chip B100 to be launched next year, which will be upgraded to 3nm process and Blackwell architecture, and its performance can be more than twice that of the current H200.