
Perhaps no document has ever had a more appropriate title than the “Interim Final Rule on Artificial Intelligence Diffusion” announced by the Biden Administration and the US Department of Commerce today. It would only be sillier if the document was written in Python – as in Monty Python.
But AI is a serious business and the title of the new policy for controlling the export of AI technologies is meant to lay out malleable framework for how AI hardware and software technologies will be disseminated around the world to the allies of the United States and will not be distributed to China, Russia, and twenty other military adversaries of the US and its allies.
The irony of this ruling – and there are many ironies – is that Nvidia, AMD, Intel, and other makers of AI compute engines have been arguing that AI technology is so strategic that the 180 countries carved up onto the surface of the Earth all need to have their own strategic compute engine reserves. This is what all of talk of “sovereign AI” is all about, and it is a sales tactic as well as a first principle (and an obvious one) that is immediately self-evident. Sovereign AI is a way to get each country to build AI clouds and to foster the use of AI routines in all aspects of commercial, consumer, and political life of the people in those 180 countries. And the tactic has been working to expand the reach of generative AI in particular beyond the big hyperscalers and cloud builders in the United States and China.
But that sovereign AI strategy cuts both ways. If you argue that AI is so strategic to the militaries and the economies of the world, you can’t just let anyone sell it to anyone else as long as there are enemies of the nation states that control AI or at least are trying to keep it out of enemy hands. China may have some of the best AI algorithm creators in the world, but they do not have large numbers of Nvidia and now AMD AI compute engines on which to run them or, thanks to prior chip controls from the Biden administration, access to foundries to create their own.
Such export controls were imposed on AI accelerators in the past for supercomputers running traditional HPC simulation and modeling applications, and still China created its own DSP-based matrix math accelerators and beat the United States to exascale-class supercomputing with at least two machines by around a year and a half. Chinese companies have several different datacenter GPUs in the works, and even with inferior chip fabrication technologies it will be able to brute force its way into the upper echelons of AI just as it did many years ago with exascale HPC.
No export controls can stop the development of Chinese AI that rivals that created in the United States, the chip vendors no doubt argue, but they can make the Chinese government work harder and pay more money for equivalent systems, and that has a strategic military value as much as supporting Ukraine helps deplete the Russian military. We have to ask whether or not that depletion, which moves us closer to the brink of World War III, is worth it, and generals and politicians weigh these things. Just as they are weighing the sweeping new rules that the Biden administration is proposing to control the diffusion of AI technology from US-based companies around the world.
Dropping this regulation bomb a week before the Trump Administration begins is clever in that returning President Trump has no desire to look weak to China or any other country in the world but also shows no desire to ever agree with outgoing President Biden about anything. So even if the Interim Final Rule on Artificial Intelligence Diffusion, which you can read about in a fact sheet released by the White House this morning, was essentially good, even with a 120-day review period for input by AI hardware and software vendors and other interested parties and a year before it goes into effect, it is hard to believe that this IFR doesn’t get scrapped by a new Commerce Secretary.
Still, even if and when that happens, the issue remains that the United States has a national security interest in restricting advanced AI technologies from disseminating to its adversaries and to be preferentially be deployed to its allies. And the Trump administration will have to weigh the obvious benefits of a free market to companies like Nvidia and AMD against those national security interests without looking like it is sacrificing one for the other.
This seems an impossible task for either the Biden administration or the Trump administration. Sometimes equality as well as equity means that everyone is equally unhappy as measured by the length of the vectors, all pointing in different directions.
No company – including Nvidia and AMD – has a divine right to make as much money as it can any way that it can. Corporations are issued by states which are part of the national framework of business, and just because the United States has not always curbed the sale of technologies to foreign powers who also turned out to be enemies – think of IBM selling punch card tabulating machines to the Nazis in Germany – perhaps it should have. But such curbs, as we have pointed out time and again, only buy time. Eventually, if something has been proven to be possible, then someone else with enough resources will be able to invent it. This is why Google does not really have a permanent advantage in any technology it has invented. But it does have the advantage of time.
So what is that time worth? And can the United States really ride the coattails of Nvidia and its smaller rivals – AMD, Cerebras Systems, SambaNova Systems, and Groq are the big ones, Intel could get there if it ever gets “Falcon Shores” out the door and it is not terrible – and claim a strategic advantage for AI for indigenous businesses, hyperscalers and clouds, and the military?
According to a report in the New York Times just before Christmas, when rumors of the new Biden administration rules were leaking out and when Nvidia was expecting about $10 billion in sales outside of the United States in its fiscal 2026 year, which roughly coincides with calendar 2025, against what will probably be around $175 billion in datacenter sales over the same period. (We expect for Nvidia to rake in $110 billion to $120 billion in datacenter sales in fiscal 2025, which roughly coincides with calendar 2024.) The accelerators that Nvidia is selling into China are crippled, and still because of intense demand, it can charge a hefty premium against their limited capabilities despite the export controls that were put in place in 2022 and 2023 by the Biden administration. This is a good business for Nvidia, in that sense. But under the new rules, there doesn’t seem to be a way to sell large systems into China. (There is a limit of machines with around 1,700 accelerators before the export controls kick in under the new IFR.)
If the United States is in an economic and technical war with China and AI is a new weapon, stopping the flow of GPUs and other kinds of accelerators into the Middle Kingdom makes perfect sense. We didn’t sell uranium or plutonium or military equipment to the USSR during the Cold War, although the United States did buy radioactive fuels from Russia after the Berlin Wall came down in 1989 and the Soviet Union dissolved in the early 1990s to keep them off the market. So the predictable outrage from the AI vendors and the cloud vendors at the new IFR is a bit much. These companies benefit from the sanctuary provided by the political will and military might of the United States.
But as we have said many times before, China is big enough and strong enough to create its own AI, just as it has created its own nuclear missiles and its own exascale HPC systems to help design them. All we are doing is buying time. We strongly suspect that with that time, the North American and European militaries will enhance AI and perfect its use in robotic warfare. If we were writing this plotline, this is how the movie would go.
An interesting aside we have not seen anyone talking about: It is not clear how the Interim Final Rule affects manufacturing facilities and warehouses, which are obviously part of Nvidia’s push into what it has called “physical AI” and which it is attacking with its compute engines and its Cosmos world foundation model.
There are 10 million factories in the world and 200,000 warehouses and just under 1 billion workers in each, as we noted last week in covering the keynote by Nvidia co-founder and chief executive officer Jensen Huang. It will eventually take a lot of GPUs to automate these, and we would not be surprised to see AI compute engine sales curbed for massively distributed factory automating as well as tightly coupled AI learning systems. The war machines of the future will no doubt be built by the factories of the future. It remains to be seen what role people will play in that future war.
AI Command And Control
The IFR revealed today is interesting is the way that it is controlling access to AI accelerators and to model weights for closed source AI models.
With the former, so-called National Validated End Users, entities that are located in countries other than those with arms control provisions and outside of the United States, will be able to build datacenters in other countries but they will have to keep at least 75 percent of their AI compute engines in the United States and specified allied countries. (The list was not available as we went to press.) The NVEUs will not be allowed to install more than 7 percent of their AI compute engines in any single country. Universal Verified End Users (what we call hyperscalers and clouds) in the United States will be able to get a single authorization to deploy in datacenters around the world (obviously not in countries subjected to arms controls) and are required to keep at least 50 percent of their AI compute engines in the US.
From our reading of the documents describing the IFR – we have not been able to locate the actual IFR – it looks like hyperscalers and cloud builders are going to have an easier time getting GPUs and deploying them than entities overseas – even America’s strong allies like European countries and Japan – and that it will be easier to build datacenters with AI accelerators in the United States than elsewhere. It will be hard for the Trump administration to argue against this, and we are pretty sure that President Trump will want to broker each deal himself and have a press conference about it. (Like the $100 billion pledge from Softbank founder Masayoshi Son committed to the US market just before the holidays that citizen Trump had a press conference.)
Moving on to the second point. With closed source model weights, the IFR prohibits the distribution of those weights that have been calculated with 1 x 1026 operations or more. GPT-4 was said to take on the order of 2.15 x 1025 operations to train. Argue amongst yourselves if having a cap on weights being exported that is more than four times greater than the cap for training GPT-4 is a low ceiling.
Given that the model weights are the useful part of the model – if you have the weights, you don’t need to train, you just need to tune – we are frankly surprised that the politicos didn’t set the bar at 1 GPT-4 unit of compute. That took 25,000 Nvidia “Ampere” A100 GPUs to accomplish in 90 days. With the “Hopper” H100s, that would have only take around 4,000 GPUs to train in 90 days, and with the “Blackwell” B100s that would only take around 870 GPUs. Add more GPUs, you decrease the time proportionately.
Oracle has been railing against the IFR before it was announced, and Nvidia joined the fray this morning as the IFR was being unveiled. The Semiconductor Industry Administration followed suit, and so did the Information Technology & Innovation Foundation. There was much complaining that in the past you just blocked access to a technology to a country and that was that, but now we are talking about levels of compute capacity and AI model weights with specific capacity against them. While this is annoying and new, you have to admit, if you wanted to keep someone from getting a smart AI model, the algorithm in the IFR is how you would have to keep the ones being developed outside of your allies dumb.
This is far from over, but the Biden administration certainly has given the AI industry and the Trump administration a lot to think about as a parting shot. It will be interesting to see how the next president and his cabinet responds, and how much input industry will have.
Selling our military adversaries machines with up to 1700 advanced GPUs each is completely nuts. Every supercomputer made in the last 30 years has been a cluster of smaller machines. This Interim Final Rule is not nearly as strict as it should be.
There should be a total ban on sales of all products with a CPU or GPU more powerful than a cellphone processor to all countries that are military adversaries of the US and our allies. For example, sales of all NVIDIA GPUs, including those for embedded systems, and sales of all computers with Apple’s Pro, Max and Ultra chips should not be allowed to military adversaries. There should also be a total ban on sales and maintenance of all semiconductor manufacturing equipment, not just the most advanced equipment, to our military adversaries.
I agree with Timothy Prickett Morgan that we can’t prevent China from eventually developing similar technology but we can slow them down. A total ban on sales of any product with a chip more powerful than a cellphone processor would slow China down enough that China might come to its senses and decide to remain part of the civilized world instead of invading Taiwan. The war in Ukraine has caused between $500B and $1T of damage in addition to enormous human suffering. A war in Taiwan would be even worse. Taiwan’s GDP ($800B) is about 4x bigger than Ukraine’s GDP ($200B) before Russia invaded.
Thank you for sharing! To explore further
https://www.360iresearch.com/library/intelligence/artificial-intelligence-in-military
“There are 10 million factories in the world and 200,000 warehouses and just under 1 billion workers in each” – that would be 10.2 million billion workers, about a million times the current human population of the planet.
Also – that’s a ratio of 50 factories per warehouse. I have no idea what the author’s definition of a ‘factory’ or a ‘warehouse’ is, but I would have said there were probably more warehouses than factories in the world.
Nobody actually proof-read this article before it as published, did they?
Not my definition. Jensen’s. Perhaps those are freestanding warehouses, like those we see for distribution of products. That is how I read it.
I stopped reading, and discounted entire article, at this point: “There are 10 million factories in the world and 200,000 warehouses and just under 1 billion workers in each, as we noted last …”