Economics of Digital Technology and AI

• Lecture 4 discusses the economics of digital technology and AI

Digital Economics

• Rule-of-thumb question: Which costs decrease with new technology adoption? For example, ICT decreased computation costs.
• This cost reduction affects market power shifts.
• Digital tech lowers search costs, increasing product variety but also information overload. Strategic behaviors such as obfuscation and dark patterns increase as well.
• Replication costs reduce and lead to > emergence of megabundles which capture value.
• Transportation costs reduced due to logistics improvements and unbundling of the location of value generation/use.
• Tracking costs reduced through rise of 'attention markets' and personalized pricing/advertisement. Intermediate Data markets lead to revival of price discrimination strategies increase.
• Verification costs reduced with introduction of feedback mechanisms at scale.
• Experimentation costs reduced by digital twins, A/B tests, lower entry costs to digital entrepreneurship, and layering of services.

Towards AI Economics

• To understand how AI impacts the economy in general, ask: Which type of cost does AI reduce?
• With data and computing power, AI systems are efficient 'prediction machines'.
• AI lowers the cost of prediction, leading to efficient decision making.
• Prediction tasks can be increasingly executed by AI, if cheaper/more effective.
• Activities that are `prediction-intense' are more likely to be impacted by AI, for example Logistics, Adtech, Fintech, HR, and scientific discovery.

Information Goods

• Software is a key product of the digital economy.
• Software is a type of information good, characterized by: increasing returns to scale.
• There is a high fixed cost of production, low marginal cost of reproduction resulting in non-rival goods.
• Economies of scope include cross-domain data feedback loops, more data, and more applications.
• Consumption externalities i.e. network effects, the users/consumers' benefit grows as the user base grows.
• Digital markets are prone to tipping, monopolization, and concentration.
  • Examples include fixed costs, which act as a barrier to entry and foster economies of scale
  • Network effects, which act as positive feedback for incumbents (collective switching costs & lock-ins)
  • Incumbents exerting dominance through acquisition, exploitation, exclusion, vertical integration, and lateral expansion
  • Big tech companies act as 'essential facilities'/infrastructures and commoditize highly regulated industries (e.g. health)

Price Discrimination

• Price discrimination is defined as "selling the same product to different buyers at different prices".
• First-degree (perfect) price discrimination is personalized pricing.
• Third-degree price discrimination is group pricing (e.g. age, location, occupation).
• Second-degree price discrimination is versioning through of a menu of "packages“ to match consumers’ willingness to pay. This includes:
  • Freemiums
  • Damaged goods (effort to make a good worse)
  • Bundling (two or more products in a single package)
• The digital economy has enabled personalized advertising.

Platforms

• Platforms are a set of stable components that support variety and evolvability in a system by constraining the linkages among the other components; architecture.
• A platform architecture partitions a system into stable core components and variable peripheral components (complements). Useful structure to cope with complexity and uncertainty.
• Key importance of interfaces between core and complements.
• Platform architecture features in product design, product families, strategy, and marketplaces.
• Marketplaces that intermediate among actors and these can be online or offline (e.g., shopping mall or business fair)

Platforms

• Economis platforms are the wunderkind of the digital economy successful.
• Value is generated by key 'ingredients' of digitalisation: economies of scope, increasing returns to scale, and network externalities.
• Low-cost structures given the nature of architecture as the value is in the intermediation.
• Most industries have been affected by platformisation which affect top worldwide companies.
• Different types of platforms exist.
• Multi-sided platforms interconnect distinct groups of users and help internalize cross-side externalities.
  • They tackle `chicken-and-egg problem' and ‘ignition' strategies, by attracting marquee users and limiting compatibility/multi-homing.
  • There is caution when platform integrates a side, as this could bring risk of self-preferencing.
• Peer-to-peer markets allows sellers entry into a market by permitting supply ‘on-demand'.
• Platforms can be monetised by ad-funded means such as Google or devise funded through apple.

AI Concepts

• Al systems are complex products and some solutions can tip to become dominant.
• Once a model attains a headstarts, it retains incumbency advantages even with open source.
• Al-based solutions can be easily packaged and repackaged, for example LLMs such as GPT4, GPT40; o1; o3-mini Claude Haiku/Sonnet/Opus;
  • Llama 7/13/70B; Mixtral 7/8x7B; Gemini 2.0, mini, experimental; this can lead to product explosion to increase revenues
• Existing Al products utilise leverage interfaces to achieve success in user applications such as ChatGPT/Claude.
• LLMs may become the substrate through which applications can develop platformisation strategies similar to the OpenAl GPT Store.

AI Platformisation

• AI systems are progressing towards platformisation through automation and intelligent design.
• There are different business models of AI implementation depending on the level of access granted.
• There is a 'gradient' with the release of AI.

Al production

• AI can be produced through full-stack producers to builders of applications on off-the-shelf models.
• The industry experimentation varies depending on different business model variants.
• AI producers range from full-stack to builders of AI applications on off-the-shelf models. Incumbents prevail and some entrants succeed due to a first-mover advantage.
• There is emerging intermediaries within model marketplaces such as HuggingFace as their existence lowers entry cost for user deployment/ standardise comparisons
• Increasing cost of computing to produce an AI product as debate ensues on wether upscaling is beneficial or not.

Model release and capabilities timeline

• There is a race at the frontier among oligopolistic actors in AI systems.
• Benchmark scores determine position in landscape, scores can be averaged out.

Profitability

• The AI industry is in a state of so-called Bertrand (price) competition.
• The revenue source could switch to ad-based as how Google search profited.

Cost Performance

• The cheapest LLM has been shown to have with a decrease in cost since 2022.

AI Problems

• Lots of investments in AI are hype-driven.

• The business model has investments without guarantees.

• Current returns also rely on the systems left unchecked through lawsuits.

• AI companies have yet to find the killer app/business model that generates reliable long-term profits.

• Al systems and LLMs in particular have already a lot of use cases

• Industries grounded on content generation have little share of the total economy, and won't be fully disrupted.

• There is a 600B $ revenue gap from infrastructural capital investments.

• Is the Al business model sustainable in the long term, and is there an expectation of a future crisis?