The “What” and “Why” of AI Chip Regulations
A decade ago, Artificial Intelligence (AI) with human-like capabilities may have seemed like something from a science fiction novel. But over the last few years, advancements in the technology behind AI and the release of widely used models like OpenAI’s Chat-GPT have changed the prevalence of AI in day-to-day conversations. For many, human-level AI is no longer a sci-fi conception but a likely outcome. As excitement and concern about advances in AI gain traction in public discourse, policymakers are facing greater pressure to establish a governance strategy. Increased national security threats from AI will also serve to motivate government action. This article aims to explain how regulating access to computational resources, or “compute,” may be a promising initial approach to reducing risks from ungoverned, powerful AI models.
Why might we want to govern frontier AI?
Arguably the most powerful emerging technology, AI has the potential to contribute great value to society. But, like other emerging technologies, its potential for good is coupled with potential for harm. Notably, AI has grown — and will grow — exponentially more capable every year. In 2011, language models were unable to produce grammatical sentences; now, models like ChatGPT can quickly compose intelligible and accurate responses to a broad range of prompts. So, when thinking about the benefits and risks posed by AI, it would be a mistake to restrict our thinking to the capabilities of today’s models.
We can think of two broad categories of dangers posed by AI: misuse and misalignment. Misuse occurs when an individual or group uses AI for malicious ends. For example, an authoritarian regime could use AI to spread misinformation and monitor dissent, or a terrorist organization could access instructions to build dangerous bioweapons through AI. Misalignment describes any scenario where a model fails to pursue its intended objective. You may have heard of AI models tasked with hiring employees or predicting disease risk exhibiting bias against minority groups. Researchers have also observed trends of misaligned models hacking digital games when tasked with winning — a phenomenon called “specification gaming” — or misleading their trainers to receive positive feedback. Such examples illustrate difficulties with training models in ways that reliably produce the desired result — a problem which, as models become increasingly powerful, will lead to greater risks of harm.
The risks AI poses, the forces magnifying these risks, and the speed at which the technology is progressing has led many experts to speculate that AI poses an extinction-level threat.
Moreover, the above risks are exacerbated by two dynamics. First, companies that release frontier AI models before others are rewarded with large profits, creating incentives to cut corners on safety and expedite the development process. Recognizing this dynamic, leaders in computing technology, including Elon Musk, Steve Wozniak, and thousands of others, signed a letter that unsuccessfully called for a pause in frontier AI development in March. Second, as models become more capable, AI will provide large advantages to companies and governments that use it. Companies and governments will not only be able to perform tasks more cheaply and quickly by using AI models instead of human labor — they will also generally receive higher quality output. As a result, companies and governments will use AI to handle greater portions of their work, as well as more of the most important tasks. To ensure the AI completes these tasks successfully, they will have to entrust AI with resources, including data, budgets, and labor. AI’s more central role will also give it increased influence over how companies and governments act. Greater access to resources and a greater weight in decision-making will amplify the risk that unsafe and misaligned models pose to society.
The risks AI poses, the forces magnifying these risks, and the speed at which the technology is progressing have led many experts to speculate that AI poses an extinction-level threat. Such claims are controversial, but AI at least seems worthy of rigorous and extensive governance efforts. While a variety of strategies have been proposed for governing AI, this article will limit itself to one such strategy: compute governance.
What is compute, anyway?
For our purposes, we can think of compute as computing power provided by computer chips. These chips are the brains of electronic devices: they are responsible for executing operations and storing memory.
The amount of compute a computer system has depends on the number of chips it uses and the quality of these chips. Like brains, computer chips can be faster or slower. We can think about the speed of a computer chip in terms of the operations it is able to carry out per second, often called floating point operations per second, or FLOPS. Just as brains can increase efficiency by working together, so can computer chips — and some can communicate more efficiently than others. The speed at which a computer chip can communicate with other computing chips is its interconnect bandwidth. When we talk about an increase in compute, then, we’re either talking about an increase in the FLOPS of individual chips, an increase in the speed at which these chips can communicate, or an increase in the total number of chips.
The training process for AI models is especially compute-intensive. Compared to the chips found in your computer, specialized AI chips are ten to 1000 times faster. And, while a typical laptop will have less than 10 chips, AI models are often trained using thousands. According to Towards Data Science, GPT-4, OpenAI’s frontier AI model, was trained on 25,000 specialized AI chips. Over the past decade, AI systems have used exponentially more compute for training, with compute usage doubling about every six months. In fact, OpenAI estimated that between 2012 and 2018, the compute applied to training frontier models increased by a factor of 300,000. It is widely recognized that these increases in compute have led to otherwise inaccessible improvements in model performance. Compute’s centrality to model development and progress mean that it would be a costly mistake to overlook compute governance as a policy lever.
Why might compute be the most promising policy lever?
Along with compute, data and algorithms also play important roles in the capability of a model. Data refers to the datasets on which AI models are trained, the quality and quantity of which affect model performance. Algorithms allow models to learn from datasets and adjust to produce more consistent, accurate outputs. Compute, data, and algorithms are often called the AI Triad. Thinking about compute as one of three factors in AI development is useful for understanding why it might be the best target for AI governance.
OpenAI estimated that between 2012 and 2018, the computing power applied to training frontier AI models increased by a factor of 300,000
Perhaps the most important aspect of AI governance is ensuring that certain groups, such as oppressive governments or terrorist organizations, are unable to train and develop powerful frontier AI models. To this end, compute currently seems like the best factor to use as a bottleneck. Data and algorithms are both digital and are thus relatively easy to make digital copies of. While cybersecurity and regulatory efforts could attempt to limit theft and prevent diffusion of these resources, data and algorithms cannot be tracked and secured in the same way that a physical product like a computer chip can. The supply chain for chips used in training frontier AI models is dominated by a few large companies, located in the U.S. and allied countries. The U.S. government can, as a result, more easily coordinate with these companies and their home countries to exclude certain groups’ access to their products.
But as promising as compute governance may be, it faces limitations. It currently seems like the most capable AI models are those with access to the most advanced chips. However, improvements in algorithm efficiency and falling costs for current state-of-the-art AI chips will likely reduce the leverage compute governance currently has. In this sense, compute governance may not offer a sustainable solution to threats posed by AI. But further efforts to regulate compute could still prove highly valuable. For example, excluding unwanted groups from access to frontier AI may provide time for the U.S. and its allies to task frontier models with finding safer training methods and other solutions to outstanding problems in AI development.
Perhaps the most important aspect of AI governance is ensuring that certain groups are unable to train and develop powerful frontier AI models. To this end, compute currently seems like the best factor to use as a bottleneck.
What is already being done to govern compute?
Governments have already taken action to limit access to compute. In 2022, the U.S. government established export controls on China that aimed to limit China’s ability to purchase state-of-the-art AI chips and to design and manufacture these chips on their own. In order to avoid restricting exports on other powerful computer chips, including chips used for gaming, the AI chip export specified a threshold for FLOPS and interconnect bandwidth below which chip exports to China are permissible.
Alongside the U.S., other countries, including the Netherlands and Japan, have also imposed restrictions on exports to China. These restrictions are especially significant given that these two countries, along with the U.S., account for 90% of global sales of chip manufacturing equipment.
Also of note is Biden’s recent executive order, issued in late October, that took a variety of steps toward mitigating AI risk. These steps included requiring companies that surpass given FLOPS and interconnect bandwidth levels to report information about training plans, cybersecurity efforts, and results of model safety evaluations. While such regulations do not govern compute per se, they do direct greater scrutiny toward models trained on large amounts of compute.
As frontier AI continues to make exponential gains in capability and pose ever-larger risks, compute governance provides an effective short-term strategy for controlling who has access to the most powerful models. Policymakers should look elsewhere for long-term solutions, but they would be mistaken to overlook further compute regulations or fail to update current regulations in the meantime.
Matt Mande ‘25 studies in the College of Arts & Sciences. He can be reached at m.mande@wustl.edu.