Compute Constraints and the New Economics of Scarcity in AI

AI now makes it possible for laymen to get simple answers to the most complex questions, analyzing countless sources of information.  It also allows users to create images, audio clips, and even videos through simple descriptions.  No programming experience is needed.  Soon, digital products will be created almost at user whim - or so AI proponents declare.  Because anyone can be a digital creator with AI, new limits arise where old ones have disappeared.

Compute Constraints as AI Limit

A major goal of AI is to make computers more accessible to humans, which has evolved over the last few years to allow humans to create complex queries with common language.  However, as humans come to ask for more robust digital creations, including entire websites and full-length movies, limitations arise in the form of computing power.  AI uses a tremendous amount of computing power, which restricts which users can receive their desired output in a timely fashion.

Economics of Computing Power

High Barriers to Entry in AI Market

Entering the AI market has very high fixed costs due to the expense of computer servers and necessary infrastructure, such as storage and computer cooling facilities.  The high energy usage of computer servers requires dedicated electrical supplies.  Essentially, there are high barriers to entry in the AI computing market, which limits competition and keeps prices relatively high.  

Fixed costs aside, there are also frequent regulatory hurdles for building data centers.  Data centers are often large and very loud, drawing complaints of nearby residents.  Local and regional governments are ambivalent about these data centers, which bring in lots of financial investments…but not many jobs.  The lack of local employment benefits for allowing a data center has drawn many skeptics of their construction, especially since their high water and electricity needs allegedly raise rates for all other users.

Paid Subscription Tiers and Economies of Scale

The growing costs of building data centers must be recouped at some point.  Many AI platforms have begun restricting use to freemium users and requiring paid subscriptions to answer more than a handful of queries per day.  This revenue will help AI firms build more data centers, but only if enough users subscribe.  As AI firms with paid subscribers expand, they will likely experience economies of scale and be able to charge lower subscription rates than smaller rivals or start-ups, heightening further the barriers to entry.  Potential entrants into the AI markets will find their average fixed costs higher than established firms that are expanding.

Supply Constraints

Additionally, AI computing servers require specialized inputs, such as advanced microchips, that are in limited supply.  Because of this persistent shortage, start-ups are unlikely to be able to purchase the necessary capital goods to build their own servers.  All available chips have likely long been purchased, thanks to long-term purchasing contracts, by major AI firms like Google, Microsoft, and Amazon.  This is yet another barrier to entry that limits competition and keeps AI costs relatively high.

Result of Compute Limits on AI

Because of supply limitations and high fixed costs, compute constraints are limiting the expansion of AI, especially for freemium users.  An oligopoly market for AI has emerged, which may improve quality but result in underproduction.  Because of high barriers to entry, this oligopoly market will likely persist, potentially leading to long-term ramifications when it comes to consumer access to AI.

AI Access as an Equity Issue?

The growing demand for AI processing by corporations and other large firms may be “crowding out” individual users, with AI firms reducing their freemium access and increasing subscription rates due to scarcity of capacity they can devote to “hobbyist” users versus “professional” users.  This could eventually separate “hobbyist” users, especially students, into tiers of haves and have-nots.  Over the long term, the haves may be artificially advantaged in the race for education accolades, including college admissions and scholarships, by being able to incorporate generative AI into the work on a near-constant basis.  The have-nots, by contrast, may be judged as less competitive due to their work not having the artificial polish of AI assistance.

Labor Market Benefits due to Limited AI Expansion (Short Run)

One group that may benefit from the slower-than-expected expansion of AI are entry-level workers, who have recently struggled on the job market due to firms’ attempts to automate many entry-level jobs with AI.  Gen Z college graduates are suffering the worst, with white collar jobs featuring routine tasks like data entry and analysis becoming automated.  As AI expansion slows due to compute constraints, firms may decide to continue hiring entry-level graduates…at least in the short run.

Unfortunately for workers, AI services will almost inevitably grow cheaper in the long run, putting labor at a cost disadvantage versus capital.  Economically, firms will utilize whichever resource provides the most output per cost of input.  

Long Run Outlook: Compute Constraints Overcome

Government Subsidies to Ensure Equitable Access?

It is possible that, as AI use becomes the expectation in society, government funds will be used to ensure full consumer access.  This is similar to the vast push in public education to ensure one-to-one student access to a digital device during and after the 2020 Covid pandemic.  Will similar billions be spent in the future to ensure that all students in public schools be able to access AI at relatively equitable levels?  Similar to governments’ goals of expanding high-speed Internet access to the general public, a future goal may be expanding AI access.

The government subsidization of AI computing may help overcome today's constraints, though the cost may be borne by taxpayers.

The Productivity Effect: Workers Shift to New Fields

Despite fears of an “AI apocalypse” in unemployment, it is likely that many workers who are displaced in their original roles by AI will find new roles.  Going back to the Luddites of the 19th century, there have routinely been fears of mass unemployment caused by mechanization and automation.  Although there have certainly been painful unemployment shocks in specific industries, the growth of other industries have usually absorbed most of those unemployed workers quickly.

This re-adoption of workers is known as the Productivity Effect, and occurs when increased productivity in one area, such as automated mass production, leads to increased demand for workers in other roles, such as design, transportation, and sales.