USA vs. China – The screw

USA looks to tighten further.

• The US Department of Commerce is looking to tighten restrictions on China still further by again lowering the performance bar above which exports to China of AI training chips will be restricted, but I am far from convinced that this will set China back that much.
• In October 2022, the USA set a limit of 640 GB/s on the speed at which AI chips could communicate with other parts of the data centre.
• This meant that the latest H100 series from Nvidia and the most advanced chips from AMD would not be able to be exported to China which remains a big market for both companies.
• As RFM has concluded on multiple occasions, AI is an area where China competes on a level playing field, but it does have some dependency on advanced chips.
• These chips improve both the speed of training and its cost-effectiveness, but they are not required to train the latest models.
• Hence, without these chips, China can still create AI, but it will take its developers longer and it will cost them more money.
• To get around this, Nvidia and AMD immediately came out with chips that were the same as their most advanced offerings, but with the communication speed throttled as per the requirements of the US Department of Commerce.
• These restrictions do not seem to have slowed China down in any meaningful way which is why I suspect the USA is looking at further restrictions that might have a greater impact.
• The problem here is that further restrictions are going to hurt the revenues and profits of Nvidia and AMD which is an unwanted side effect of this action, but in this regard, it looks like long-term national security is being put first.
• If the USA wants to really hit China, restrictions on the geometries of AI chips that can be shipped to China would be one way to do it but that would take a hammer to Nvidia particularly.
• Furthermore, this could be counter-productive as the rule of thumb that the industry has been using for the last 3 years is starting to break down.
• This was proposed by Kaplan et al (see here) who demonstrated a strong link between model size and performance which is why models have been getting bigger and bigger over the last few years.
• This has played directly to Nvidia’s advantage because bigger models need more infrastructure and more resources to support and train them.
• However, the cutting edge of the industry is moving in a different direction with training innovations implying that better performance can be achieved with smaller models that have been trained with more data.
• Furthermore, new techniques are allowing Meta’s 7bn parameter LlaMa foundation model to be fine-tuned on powerful laptops and PCs with high-end graphics cards.
• This is why there are a plethora of new models popping up that are based on LlaMa but have been re-trained by hobbyists to perform any number of functions.
• The upshot of this is that China may not need to acquire the latest and greatest chip if it is able to harness these innovations which have already been widely publicised in the academic literature.
• However, if the US was to restrict AI chip shipments below 14nm, then this would have a much more profound impact on China but would effectively end Nvidia’s AI business in China.
• Consequently, I don’t expect the Commerce Department to go that far but at the same time, I can’t rule it out completely given the recent history and the worsening relationship.
• The increased desire to contain China has led to the USA targeting far more than semiconductors and increased use of its leverage in the semiconductor industry to hamper China more widely.
• This is bad news for Nvidia which is now priced for global domination in AI with no hiccups making me very uncertain as to where the shares go from here.
• Hence, if I was holding Nvidia (which I am not), I would be tempted to take some off the table given that the shares are priced for perfection.
• I continue to prefer the cheaper end of the semiconductor sector including TSMC, Qualcomm and MediaTek.