big data technology.
The artificial intelligence industry will be large, but only a few companies offer models such as ChatGPT, Claude, and Google's Gemini family. However, many companies plan to provide tools for applying these models to specialized tasks for their business, government, and consumer customers. This is a very early judgment and may prove wrong, but it is based on solid economic theory and experience. What may change is the cost and opportunity of the services provided.
Some industries have a small number of companies, while others have many. When you think of manufacturers of large commercial aircraft, he is the only two companies that come to mind. What about restaurants? In my small suburban city, I see 24 cities affiliated with the local chamber of commerce. There are several hundred in the metropolitan area. To understand the market structure for AI, you need to understand why some industries have more business and some industries have less business.
Much of the market structure is driven by economies of scale. Last year, Airbus and Boeing produced 1,263 aircraft, or about 600 aircraft each. Suppose that production was distributed among 100 manufacturers instead of 2. That would be 12 planes each. Average production costs would have been much higher. Material costs may have been similar, but design costs would have been spread over far fewer planes. Specialized jigs and fixtures used in the production process are less common, increasing manual labor and rework.
Now consider a restaurant that serves perhaps 100 customers at a time. How much cheaper would it be if he seated 1,000 guests in one large restaurant? Meal preparation would likely be more automated, but servers and delivery companies would be in frequent conflict. Maintaining the quality of both the food and the experience can be more costly.
Diversity of demand also drives market structure. Many restaurants have a variety of specialty cuisines, including Mexican, Italian, and Thai. However, air travelers don't seem to be interested in seeing too many variations in airplane design.
Developing large language models to power ChatGPT and its peers is very expensive. Developing GPT-3 may have cost him $3 million, but GPT-4 cost him over $100 million. The model was significantly improved because it was trained on a larger dataset and required more processing resources. After a model is developed, the cost of running the model to answer queries increases with the size of the model.
AI's big value proposition comes from leveraging large-scale language models for specific applications. Large-scale language models are general-purpose and enable communication in everyday language. A company wants to answer customer inquiries about bill payments. Engineers want to consider design alternatives that fit certain parameters. Salespeople want to identify past customers who are likely to buy from them again. They all use the same general-purpose AI model and can “add” additional features, data sources, and practices.
Large language models can be fine-tuned for specific applications. In the most well-known example, GPT was tweaked to sound like a helpful assistant called ChatGPT. In some cases, a large language model is simply an input/output mechanism, translating common English (or another language) into instructions that can be input into a smaller, specialized artificial intelligence model. The results of that model are then fed back into a larger language model so that users can receive the results in everyday language. In other cases, large language models are connected to external information such as bill payment history or a product's service manual.
Each of these specialized applications can benefit from development by people with deep knowledge of the field. Mechanics know best what information is useful to other mechanics. There will be a large number of these specialized applications. It can be developed and implemented at relatively low cost. The large language model will resemble the Airbus and Boeing oligopoly, but the application division will look like sushi, hamburgers, pizza, etc.
There may be only two or three companies providing services within an application, such as customer return assistance. However, the number of professional services will be huge. Some companies build expertise in developing specialized applications and use that expertise to develop a variety of products. But small app developers will persist because there will always be entrepreneurs looking for opportunities that other companies have missed.
This view of the future market structure for artificial intelligence could go wrong in several ways, some of which are easily recognizable today. First, large language models can potentially develop better capabilities for handling specialized tasks. In 2021, I asked AI specialists about developing highly specialized applications for industries such as finance. He said it might take him a year, but in that case the application would be created based on his year-old model. Perhaps using a modern but less specialized AI model will yield better results. So I might take a year off in Tahiti and then use a generic model.
A second option for this market structure prediction comes from the decreasing cost of AI model development. By some estimates, training costs have been reduced by 80% over his two-and-a-half years, which equates to a reduction of about 50% per year. Other estimates decline by 20% to 70% each year. Many articles have pointed out the high price of Nvidia chips used for AI. These are expensive because the chips are very powerful and help reduce overall model training costs. So perhaps large-scale language models will become so cheap that many companies will start producing their own language models. Even if costs leveled off at, say, $100 million, that's well within the capital budgets of many companies.
Third, as in many other industries, regulation can change market structure. Countries may require or subsidize locally built AI models. If Canadian television stations have to air Canadian-produced content, that content should probably be generated using Canadian AI models. Some proposed regulatory schemes in the United States could lock in early incumbents and disadvantage start-ups.
Fourth, someone may develop a unique development method that is unknown to others. Recent research results in AI are often shared publicly or come up with the same solutions that a company has developed before, but new and amazing things are developed and kept secret. There are always possibilities.
Market concentration can also be influenced by other factors that don't seem to apply to AI as we know it, but that could change. Other factors include network effects and access to distribution channels.
Companies in the AI space should be as agnostic as possible about future market structures, but if they had to make a bet, it would most likely be that a few large companies develop large language models. The prospect is that many small and medium-sized companies will offer specialized applications.
