- In a surprise move, outgoing President Joe Biden unveiled new rules governing AI diffusion, including chip shipments and model weight sharing
- The move clearly targets China, but one analyst warned it could backfire
- Other analysts noted the rules too narrowly focus on GPUs and will likely change under incoming President Donald Trump anyway
Parting gift or Pandora’s box? Outgoing U.S. President Joe Biden certainly left something for the artificial intelligence (AI) community on Monday in the form of surprise new regulations governing chip exports, compute purchasing and the sharing of model weights. Moor Insights and Strategy VP and Principal Analyst Matt Kimball told Fierce that while it’s hard to predict exactly how things will play out, the rules have the potential to backfire spectacularly.
You can read the White House fact sheet on the new regulations here. But the long and the short of it is the U.S. is implementing new rules in an attempt to control who has access to GPU compute power (surprise, there are no restrictions on 18 key allies!) and who can access sensitive information about the way closed-weight models perform. As we’ve noted before, model weights essentially function like dials you can turn to emphasize different things within a model, and access to model weights is considered by many (including the U.S. government) to be a security risk.
Though China isn’t mentioned at all in the fact sheet, the move is clearly directed at heading off the country’s AI advancements. But to hear Kimball tell it, the regulations could actually end up having the opposite effect.
How? Well, think back to what happened when the U.S. and others tried to cut Huawei (and ZTE) out of global telecom networks. Huawei didn’t wave the white flag. Instead, the government-backed company doubled down on R&D investments and launched an aggressive campaign to subsidize the use of its equipment in less-developed nations. This strategy was so successful that Microsoft recently called it out in a recent blog advocating for investment and deregulation to help make U.S. technology the de facto standard for AI.
“One of the questions we have to ask is whether we are truly protecting the U.S. and its allies with these protectionist measures or are we accelerating investments from adversary governments into AI that perhaps pay out longer term? It’s a tough question to answer,” Kimball told Fierce.
He also noted that while the rules target closed-weight models, Chinese companies like Alibaba have been building open-weight models like Qwen and trying to “gain a global footprint with these.”
And, as Microsoft put it in its blog, “The Chinese wisely recognize that if a country standardizes on China’s AI platform, it likely will continue to rely on that platform in the future.”
In an interesting twist of fate, the U.S. AI clampdown could actually end up benefitting – of all companies – Huawei. As Kimball explained, lack of access to U.S.-made chips could mean Alibaba and others turn to Huawei and its servers and storage, making the company the “platform of choice” for AI deployments. It’s exactly like Dell, HPE and Lenovo launching AI factories, but just….in China.
Will the rules even work?
As one might expect, the rules drew ire from U.S.-based AI chip giant Nvidia, which predictably went off about “government overreach” and how the move “threatens to derail innovation.” But neXt Curve founder Leonard Lee told Fierce the regulations overlook one critical fact: while no doubt important, GPUs aren’t the be-all, end-all of AI.
“The ‘chips’ (GPU) that go into these AI systems are having the least influence on the highly fluid definition and standard and scaling of these systems,” he pointed out in an email. “High-bandwidth memory (HBM), Interconnect and networking including inter-datacenter interconnect are playing a much bigger role especially for superclusters larger than 100,000 GPUs.”
Related
Biden proposes new rules on AI exports, drawing fire from Nvidia, much of chip industry
Plus, he said, a number of companies – like Cerebras – are coming up with novel approaches to AI compute that don’t necessarily involve GPUs. The rules don’t seem to address those.
Additionally, though the regulations require a license for orders of more than 1,700 GPUs, Lee noted that “post-training of large models can be done with systems with far less than 1,700 GPUs, precluding the need for massive superclusters to realize ‘advanced AI systems.’”
And finally, we come to the elephant in the room: Biden’s term will be over in just a few days, leaving open the question of whether incoming President Donald Trump will uphold the newly implemented rules.
According to AvidThink’s Roy Chua, Trump’s team will likely make some tweaks. And the timing of Biden’s move could delay any “tit-for-tat responses from countries like China.”
“I don't see that it will have much immediate impact since China will be taking a wait-and-see approach to the Trump's administration initial policies around protectionism. The incoming administration's much vaunted blanket tariff across all goods is a much bigger issue at hand,” he concluded.