America Last? Relegating U.S. national security to leading from behind.
It seems we’ve lost the plot.
The United States has correctly determined that beating China in artificial intelligence, semiconductors (GPUs) and underlying cloud infrastructure (networking, compute) is the defining economic and security issue of our time. There is no daylight between the U.S. technology industry and so-called “China hawks” on this point. But if we have reached that conclusion, the CCP has as well.i And we would be naïve to assume that China won’t pour all resources necessary to take leadership in these critical technologies from the U.S. That’s why the decisions we make today are so critical and why Biden’s “Export Control Framework for Artificial Intelligence Diffusion” (“Biden Diffusion” or “Diffusion”) proposal is so dangerous. This rule will not advance U.S. national security interests because it is an over-prescriptive collision of jargon meets market allocation.
The recent announcement of DeepSeek’s R1 model makes the point we made several weeks ago and referred to as Mario’s Law here.ii,iii,iv Initial reports indicate that DeepSeek used lower-performance GPUs with an optimized LLM training approach to develop a current generation model on previous generation hardware. The fundamental problem with the Diffusion Framework is that if U.S. technology is not available, global customers will turn to alternate suppliers for an AI stack based around models like DeepSeek, which are served by any number of Chinese cloud providers, and are unfettered by the Diffusion Framework. The DeepSeek announcement is a vivid reminder that the Diffusion Framework relies on the underlying premise that the U.S. lead in GPUs and AI is large enough that the Chinese can be constrained by our limits on global supply of critical technology. But we don’t really need to have this debate because: 1) The Diffusion rule is simply incomprehensible and unworkable; and 2) U.S. industry does not oppose greater export controls in this space.
The Diffusion Framework was pitched as a “small yard, high fence,” which feels to us like playing right into the hands of “belt and road.” There has never been a moment in history where U.S. industry wins by competing in “small yards.” America leads the Chinese in AI, GPUs and cloud because U.S. industry has out-worked and out-innovated our Chinese counterparts. Notwithstanding DeepSeek, we intend to keep it that way with projects like Stargate, because we now understand that DeepSeek was far more evolutionary than revolutionary. And we suspect there’s more behind DeepSeek’s apparent success than was publicly advertised by the Chinese. We argue here that the best way to maintain U.S. leadership is to ensure U.S. firms flood the global market with U.S. technology, cutting off the air supply for our Chinese competitors. We argue here that the Diffusion Framework regulates global commercial cloud—not AI—for the first time, released as a Final Rule just five days before a change in administration and without any industry consultation.
We thought we had seen the worst of this ill-conceived idea. But we were wrong. The entire Diffusion Framework is not only bad policy, but also unworkable.
Here’s the deal. Export controls have, and will, play an important role in restricting China from obtaining mass quantities of latest-generation GPUs. Starting under the first Trump Administration, and continuing in the Biden Administration, controls on key chokepoints of GPU manufacturing capabilities have helped to confound the CCP’s goal of domestic independence from western control of advanced GPU design and manufacturing technologies.v We should use all levers of national power to constrain our adversaries, and notably the CCP, from obtaining U.S. GPUs to accelerate the development of new frontier models. But somehow the Biden Administration thought that the best way to constrain the Chinese is to constrain the U.S. more.
Export controls or arbitrary market allocation?
Let’s start with the basics, because the Atari regulators who drafted this proposal need a dictionary.vi “Diffusion” literally means to disperse something more widely, which is the exact opposite of what the Biden Diffusion does.vii The Biden Diffusion is not an export control policy; it is an arbitrary set of market allocations seemingly picked from thin air. Or picked from the ivory tower of RAND, the Center for Security and Emerging Technology (CSET), or the Center for American Progress. It then applies a set of complex, academically and commercially questionable formulas to manage, monitor, and manipulate the global distribution of GPUs according to their quarter-baked view of who should get chip allocations and when.
This rule does not diffuse anything, rather it attempts to suppress U.S. global leadership by allocating chips to a few certain markets and not others, thereby creating massive artificial demand for products from China by turning export controls upside down.
Buckle up for an explanation.
To be fair, from the first version of the proposal leaked a few weeks ago, the Biden Diffusion shrank from 200 pages to a mere 168 pages of mind-numbing market allocation regulations. We could call that progress, but between drafts, it appears the Biden Diffusion only got more efficient in handicapping U.S. technology leadership because the substance of the proposal got worse.
The Biden Diffusion constrains the shipment of GPUs and AI processing power globally, creating a three-tier system. There’s a list of 18 “Tier 1” countries (down from 20 in the prior proposal). We will call these countries the AI-18. Then there are 23 “Tier 3” countries like China, Russia, Iran, and North Korea in which any GPU deployment is appropriately banned outright. That leaves the remainder of the 143 or so countries which fall into “Tier 2.”
With a dose of market reality, Tier 1 is entirely too small, and Tier 2 is entirely too large with seemingly no cognizable set of standards for being on one list or the other—and it is certainly inconsistent with the Commerce Department’s established Country Groups.viii To underscore the arbitrary nature of these tiers, the Biden Diffusion places Yemen in the same category as Israel. We are unaware of any national security policy that embraces the concept that Yemen and Israel be treated similarly. And between the draft Biden Diffusion leaked a few weeks ago to the final Diffusion, the AI-20 became the AI-18. Somehow, we missed the news of what U.S. allies Switzerland and Poland did to the Biden Administration to fall off the list in the intervening two weeks, yet Denmark and Finland are OK. French Guiana, totally fine. Saudi Arabia and the UAE, not included. Huge markets like India and Brazil didn’t make the cut. Mexico out.
The Biden Diffusion then creates a new class of approved data center operator exporters, called Universal Validated End Users (“UVEU”). Entities such as Google, Amazon, Microsoft or Oracle—and there will be others—must apply to be designated as “trusted” exporters of GPU technology.ix You might think that sounds relatively straightforward. Trusted parties attesting to approved end users and end use. But, no. Before a UVEU can deploy any GPUs outside the U.S., it must first calculate its global deployed Total Processing Power (“TPP”), another term made up out of thin air but one that ultimately drives the entire global market allocation scheme.
TPP purportedly measures the total performance of a processing unit based on the maximum theoretical performance of the unit (as described by the manufacturer) and the bit length of the operations being used (for example, many LLMs are trained using 16-bit floats). Each vendor’s total TPP will be different based on its total deployment of GPU’s and the mix of those GPUs, even if those GPUs may have nothing to do with frontier models or national security. For example, if large numbers of GPUs are deployed to support AI Agents at scale in the U.S. for SaaS applications, those GPUs increase a vendor’s TPP and therefore increase the number of GPUs which can be exported. Or if a large number of GPUs are deployed to support gaming at scale in the U.S., those GPUs increase a vendor’s TPP and therefore increase the number of GPUs which can be exported. The problem is that the TPP calculation is not tethered to a specific kind of GPU or workload, creating enormous disparities among vendors and creating an export control policy that is not linked to an actual number of GPUs that might cause concern. Rough translation: If you increase the number of apples used in the U.S., you increase the number of oranges available for export. The other more obvious problem is that a certain number of GPUs exported to a country or region are either a problem or not, yet this rule creates an elastic set of ever-changing controls based on the unrelated use of GPUs in the U.S.
Assuming a UVEU eventually calculates its total TPP, it must then calculate how to allocate TPP globally, using what we’ll label as the “50/25/7 percent” rule. Under this Diffusion, no more than 50% of a UVEU’s total TPP may be located outside of the U.S. The Diffusion then becomes even more prescriptive in its market allocation, limiting how that 50% outside of the U.S. can be deployed. While a UVEU is permitted to deploy up to 50% of their overall TPP in AI-18 countries, a UVEU must not deploy more than 25% across 143 Tier 2 countries, collectively. If a UVEU were to deploy 25% across all Tier 2 countries, that leaves 25% of its TPP to deploy in AI-18 countries. In practice, this means many AI-18 countries are not guaranteed any GPUs, because they must play their own version of Hunger Games, competing against each other for the 25% TPP ration that a UVEU may deploy.
But that’s not all. UVEU authorized companies may not allocate more than 7% of their TPP in any single Tier 2 country. Now factor this into the overall 25% market allocation of TPP that a UVEU can deploy in 143 Tier 2 countries (again, collectively, not in each). Suddenly we run into scenarios where the 7% limit effectively shrinks 143 countries down to three and a half.x Now it’s Hunger Games meets Squid Game.
The intended consequence of this rule means that most of the countries around the world stand to receive zero GPUs from American sources operating under a UVEU authorization. That’s a pretty small yard.
Then there’s the fine print of the “trusted” Universal Validated End Users. Turns out UVEUs aren’t so trusted after all because, in order to qualify, U.S. companies are forced to sign a prenuptial with the government requiring that all global data centers relying on UVEU status (including co-located data centers, sovereign data centers, cloud at customer data centers) comply with FedRAMP High—a set of U.S. government security requirements meant for U.S. government information—as well as a slew of brand new mandates, including third party audits to confirm compliance. Ironically, one area where there is “universal” agreement is that the entire FedRAMP approach for federal data is broken, yet we now adopt it as the centerpiece of our export control policy.
This mandate applies to all countries, including the AI-18. So, when the Biden Diffusion states that “no restrictions apply” to the AI-18, that is, let’s say, a liberal reading. How many global data centers are FedRAMP High compliant today? Zero. Because to achieve FedRAMP High status, the High Baseline SA-9(5) control contains a security requirement that data centers are located in “U.S. Territories or geographic locations where there is U.S. jurisdiction.”xi So, in one Diffusion, the proposal has managed to dramatically shrink the global market (small yard) and radically increase costs (high fence).
And then there is the irrational over-reach of the rule that applies to lots of GPU uses that cannot possibly be a threat to national security and can’t possibly be diverted or aggregated by our enemies. Remember, in a cloud environment the GPUs are owned and operated by the end user, which is the hyperscale cloud provider, not the customer. We know exactly where these GPUs are and how they are being used. GPUs deployed for a national health care system? Not a threat. GPUs deployed for a bank’s anti-money laundering efforts? Not a threat. GPUs deployed for AI agents in SaaS applications, like finance, HR, supply chain? Not a threat. GPUs deployed in consumer applications? Not a threat. GPUs deployed for gaming? Not a threat. All these GPUs are “on rails”, meaning they are tied to the system for which they are implemented. They cannot be diverted or aggregated for other purposes. Rather than have a regulation that eats the entire market, how about we decide where the threats lie and write a rule to address those threats?
The DeepSeek Wake Up Call.
The inconvenient truth of this export control policy rests on the Biden Diffusion’s simple assumption that the Chinese cannot fill market demand for those restricted from U.S. technology. We believe there is a lot more to learn about DeepSeek in the coming months. DeepSeek’s advances were an expected evolution in model refinement, and nothing publicly available so far leads us to believe that the U.S. does not maintain its AI leadership position.
But it is also true that demand for AI chips is only increasing, and those countries not on the AI-18 list or left begging because of the 50/25/7 percent rule will not wait on the U.S.—they will turn to alternative suppliers. Major economies around the world are not going to sit out the AI revolution; they will simply go elsewhere.
Because of Mario’s Law, less capable GPUs can be accounted for by simply adding more, less performant GPUs to the cluster. And while China’s alternative technologies may be inferior for now, the Biden Diffusion creates a global vortex of capital and demand that will relieve the PRC from needing to directly subsidize Chinese manufacturers as a result of the Small Yard’s market allocation rules.
There is no disagreement on the need to beat the Chinese and to maintain U.S. leadership on these critical technologies, but the Diffusion is plainly ill conceived, primarily because there was no industry consultation. If the idea here is to out-compete China by confining the U.S. to a small yard with high fences while China occupies a large yard with no fences, the outcome will be clear.
There is no AI crisis that can’t wait 90 days, DeepSeek notwithstanding. The Biden Diffusion should be pulled entirely, and a Notice of Proposed Rulemaking should take its place wherein there will be meaningful engagement with industry.
Let’s understand what DeepSeek did and how they did it. Let’s create a rational export control policy—not arbitrary market allocation policy—that accomplishes our agreed-upon goal of ensuring U.S. dominance while constraining China’s ambitions.
i Xi Jinping states: “Accelerating the development of a new generation of artificial intelligence is an important strategic tool for us to win the initiative in the global technological competition and is an important strategic resource for promoting the leapfrog development of science and technology, industrial optimization and upgrading, and overall leap in productivity in my country.” In “Xi Jinping on Artificial Intelligence: An Important Strategic Tool to Win the Initiative in Global Technological Competition,” People’s Daily and Communist Party of China News, 1 November 2018, http://cpc.people.com.cn/xuexi/n1/2018/1101/c385476-30376558.html.
ii https://www.technologyreview.com/2025/01/24/1110526/china-deepseek-top-ai-despite-sanctions/
iv https://x.com/stevesi/status/1883746880536072375?mx=2
vi https://en.wikipedia.org/wiki/Atari_Democrat
vii Diffusion is the process in which particles (or chips and AI processing power in this case) move from areas of high concentration to low concentration. However, in its application, the Biden Diffusion will instead keep a high concentration of chips and AI processing power in the U.S.—the exact opposite of diffusion. See https://www.britannica.com/science/diffusion
viii Current as of January 24, 2025 as identified at this link: https://www.bis.gov/ear/title-15/subtitle-b/chapter-vii/subchapter-c/part-740/supplement-no-1-part-740-country-groups
ix Applications for UVEU authorization are evaluated by the End User Review Committee (ERC). Chaired by the Department of Commerce, the ERC is composed of representatives of the Departments of State, Defense, Energy, and other agencies as deemed appropriate. The ERC is responsible for determining whether to add to, to remove from, or otherwise amend the list of VEUs and associated eligible items.
x Showing our work, [Total TPP available for all Tier 2 countries] / [Total TPP allowed in one single Tier 2 country] = [Total TPP*25%] / [Total TPP*7%] = 3.57.
xi https://www.fedramp.gov/an-update-to-fedramps-high-baseline-sa-95-control/