The Era of Acute GPT Shortage Will Soon Begin
In a recent collaborative paper between OpenAI and the University of Pennsylvania titled "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (LLMs)" researchers outlined the economic and social repercussions of the widespread adoption of LLMs. They predict that GPT-4 can significantly expedite 15% of all worker tasks in the US without compromising quality. With the right software and tooling, this figure may rise to as much as 56%. This suggests that market forces should trigger rapid, widespread adoption of LLM technology, potentially creating a surplus of LLM services. But why would there be a shortage?
Perhaps the shortage will be deliberately created by regulators to allow more time for society to adapt to LLMs? After all, Sam Altman stated in an interview with ABC News that the pace of this technology's introduction should be slowed. Unfortunately, I don't believe that any regulators, not even Sam Altman, can slow the proliferation of LLMs. Any regulatory efforts will fail in the absence of global governance. Nations that attempt to limit LLM adoption risk falling behind others that prioritize it. Moreover, foreign competition for GPT-4 is already here with Baidu's recent AI chatbot, Ernie. While not as sophisticated as GPT-4, it is expected to improve over time.
Instead, the shortage I foresee is rooted in the drastic imbalance between the supply and demand for LLM services.
Stories of increased productivity and earnings using GPT-4 are rampant on social media, with professionals across industries embracing the technology. Additionally, OpenAI is offering GPT-4 services to businesses, and numerous large corporations are developing applications on top of the GPT-4 API. This will give rise to a new generation of websites, suggesting that LLMs will soon power the majority of popular sites like Expedia, Zillow, and Kayak. Consequently, it's not far-fetched to anticipate that within three years, GPT-4 will have to serve two orders of magnitude more requests than it does today.
The question remains: can the supply keep up? Nvidia, responsible for 93% of GPU computing in data centers, would have to significantly boost production in a short time to satisfy this demand. However, constructing new fabs to increase GPU production typically takes about three years.
With demand potentially skyrocketing and supply struggling to catch up, OpenAI may have to make tough decisions regarding access to their service. Early indications of limitations are already evident, with ChatGPT Plus curbing the number of questions users can generate per session and disallowing upfront yearly subscription payments. The GPT-4 API access waitlist is also expanding, with many businesses being denied.
It's my sense that OpenAI may stop offering paid access to GPT-4 for individuals soon. Instead, they're likely to reserve access primarily for US-based businesses, as that will likely be more profitable. The reality is that only the US and China have the necessary GPU computing power to run the advanced LLMs models efficiently, meaning that companies from other countries will struggle to adopt them. This could create a significant competitive advantage for American and Chinese companies in the global market. For individuals and non-US or Chinese businesses, this could usher in an era of acute GPT scarcity.