USA vs. China – No Contest pt. III

Huawei is already active in Malaysia but not in AI.

  • The media is reporting that Malaysia is considering using Huawei in its data centres, but there is already a 50MW Huawei installation up and running but for regular compute rather than AI.
  • Bloomberg (see here) is reporting that Malaysia is considering using Huawei in its AI data centres, but there are elements in the article that don’t make any sense.
  • The story states that Telecom Malaysia is running the project and that it is considering using Huawei’s Ascend 910C in its AI data centres, which doesn’t make sense, as I think it has been out of stock for a long time.
  • This is because all of the indications are that the Ascend 910C was manufactured by TSMC and illicitly obtained by Huawei through shell companies outside of China.
  • This explains its superior performance over the more recent, in-house-manufactured Ascend 950, but all of the indications are that the TSMC wafers ran out over a year ago and Huawei cannot get any more.
  • I met Huawei last week, and it told me that it had 50MW of capacity already up and running in Malaysia (of which it was displaying a model) with the 910C, which makes far more sense.
  • Furthermore, the 50MW data centre is not for AI, but regular compute, but could be upgraded to AI at some point.
  • This could be Huawei, Nvidia, AMD or anyone else, as the data centre has been designed to be upgradeable to anyone’s silicon, not just Huawei’s.
  • Hence, it is quite possible that Malaysia ends up with Western silicon housed and supported by Huawei shell, core and power.
  • This would make the most sense for Malaysia because the Ascend 950 and its future variants are very uncompetitive compared to Western chips in terms of performance and cost to run.
  • Hence, using a Western variant will greatly improve Malaysia’s cost per unit of compute produced, even when paying Nvidia 75% gross margins to acquire its silicon.
  • There is no sense in Huawei bidding to change the 910C to the Ascend 950, as by Huawei’s own data, the 950 is a 20% downgrade when it comes to performance.
  • Hence, I suspect Huawei is bidding for new infrastructure in Malaysia, which the USA and other Western offerings should be able to win easily when taken solely on the merits of the offering.
  • However, China is making an effort to win the support of the Global South, and so it may be offering very attractive terms to tempt Malaysia to stick with Huawei as it moves into AI.
  • I still think that this is unlikely to be successful, as Huawei’s offerings for 2026, 2027 and 2028 remain so uncompetitive that it costs 3x as much to build the equivalent amount of compute compared to Nvidia, which also consumes 9x the electricity.
  • Hence, no rational actor would ever deploy it outside of China, as there are far better options available which are orders of magnitude cheaper to build and operate.
  • This remains by far China’s biggest weakness, as when it comes to models and open source, it has already proven itself to be very competitive.
  • However, this disadvantage is so large that I continue to think that the USA maintains the edge in the AI race.

RICHARD WINDSOR

Richard is founder, owner of research company, Radio Free Mobile. He has 16 years of experience working in sell side equity research. During his 11 year tenure at Nomura Securities, he focused on the equity coverage of the Global Technology sector.

Leave a Comment