Near-Zero Marginal Cost for AI-Enabled Knowledge Goods?

Mustafa Suleyman, DeepMind cofounder and now Microsoft AI's CEO argues that, because of artificial intelligence, "the economics of information are about to radically change because we're going to reduce the cost of production of knowledge to zero marginal cost."

At least two observations are possible. The perhaps-negative view is that such thinking tends to happen with technology bubbles. The perhaps-positive interpretation is that a major disruption of information-related businesses and industries--perhaps on a greater scale than the internet--is possible. 

Most digital products can have marginal costs close to zero, which is quite different from physical goods. 

source: IP Cariier 

It might be fair to note that what Suleyman refers to is marginal cost, not sunk cost. In other words, the cost of information infrastructure is one matter. The cost of producing the next unit can be marginally close to zero.

Think about communications infrastructure and platforms, where the sunk cost of networks is quite high and very capital intensive, while the cost of producing the next unit is almost immeasurably low. All that has huge costs for many content creators, distributors and firms in many industries. 

Product/Service	Description	Near-Zero Marginal Cost Explanation
E-books	Digital books	Once created, distributing additional copies has negligible cost
Software	Applications, operating systems	Copying and distributing software digitally has minimal incremental cost
Streaming Media	Music, movies, TV shows	Serving content to additional users has minimal bandwidth costs
Digital Information	News articles, blogs	Sharing information online has negligible distribution costs
Online Courses	MOOCs, video tutorials	Adding more students to an online course has minimal additional cost
Cloud Storage	File hosting services	Incremental storage has very low cost due to economies of scale
Social Media	Platforms like Facebook, Twitter	Adding new users has minimal cost once infrastructure is in place
Digital Advertising	Online ads	Displaying ads to additional viewers has negligible cost
Open Source Software	Linux, Drupal	Community-developed software has near-zero distribution cost
3D Printed Objects	Custom products	Once design is created, additional prints have low material costs
Renewable Energy	Solar, wind power	Generating additional electricity has very low marginal cost
Ridesharing	Services like Uber	Adding passengers to existing routes has minimal additional cost
Home Sharing	Platforms like Airbnb	Renting out unused space has low incremental cost for hosts

“Near-zero pricing” (or the perhaps-better known expression of “marginal cost pricing”) is a business principle that underpins and complicates business strategy in a wide range of industries, ranging from internet apps to computing; retailing to media; communications and consumer electronics.

Marginal cost is a universally accepted pricing principle, representing the incremental cost to produce one more unit. The key idea is that it is profitable to keep producing additional units right up to the point where marginal cost and marginal revenue hit zero. At that point, one stops producing, as losses will occur.

But physical goods and digital goods have different marginal cost curves. For a communications service provider, at some point there is so much demand that a network has to be upgraded. That adds capital investment cost, so the marginal cost actually has to rise.

Digital products are different. Once the original is created, the marginal cost can actually remain infinitesimal, even with vastly-greater usage. That also implies that retail price can be very close to zero, and still yield a profit.

In fact, some believe zero marginal cost might be among the most-important business drivers in the early 21st century, though the idea remains controversial.

A company that is looking to maximize its profits will produce “up to the point where marginal cost equals marginal revenue.” In a business with economies of scale, increasing scale tends to reduce marginal costs. Digital businesses, in particular, have marginal costs quite close to zero.

source: Praxtime

In other words, the incremental cost of adding one more Gmail user or one more Facebook user are infinitesimally small.

But marginal costs also are immeasurably small even in some industries with high capital intensity. What, for example, is the incremental cost to supply one more megabyte of internet access capacity; one more minute of voice usage; one more text message, on a network that already is built and operating?

To be sure, additional sales help most businesses, digital or physical. But profit margins for digital goods--based in large part on near-zero marginal costs--often exceed those of physical goods. 

source: Barbara Hoisl

But the danger of pricing at marginal cost (increasingly a price very nearly zero) is that “where there are economies of scale, prices set at marginal cost will fail to cover total costs.”

Think of the “sunk cost” of building a mobile or fixed network. Retail pricing has to be set at a level that allows recovery of that initial network cost, plus profit. So overall pricing cannot be set at the marginal cost of the last units, but at a rate including recovery of sunk costs.

Add to that the possibility that product prices for the end user also include revenue generated by third party partners (advertisers, retailers on a platform) and end user consumption can actually be subsidized.

The point is that even if the incremental cost of supplying one more megabyte of data consumption, one more minute of a voice call or one additional text message is quite close to zero, a service provider cannot price at marginal cost, forever.

That accounts for the business advantage many app, content and services providers hold over a facilities-based connectivity provider selling apps and services. An over-the-top app provider does not have to recover a physical network’s sunk costs.

If Suleyman is correct--and many will disagree--we could see dramatic new disruptions of existing information-based industries and activities as well as the potential creation of entirely-new industries. 

The near-zero marginal cost of digital goods has led to the emergence of various business models. to Freemium models, advertising-supported content, and almost anything that can be bought “as a service” provide examples. 

Digital platforms and marketplaces that leverage create massive scale and network effects that create the platform for revenue and monetization. Using past history, when low marginal cost created cloud computing, software as a service, social media, video and audio streaming, digital versions of physical products (e-books) emerged, AI is likely to produce new products, platforms and industries. 

As was the case for the internet impact on digital goods in general, AI has the potential to alter any number of functional costs. How much of that impact will be incremental, and how much exponential, remains to be determined. 

Aspect	Physical Goods	Digital Goods
Production cost	Significant material and labor costs for each unit	Near-zero cost for additional units after initial creation
Distribution cost	Shipping, handling, and storage expenses	Minimal costs for digital distribution (e.g., bandwidth)
Inventory management	Requires physical storage and logistics	No physical inventory needed
Scalability	Limited by production capacity and resources	Highly scalable with minimal additional costs
Customization cost	Often expensive to customize individual units	Can be customized at little to no additional cost
Geographical limitations	Subject to shipping costs and trade barriers	Can be instantly delivered worldwide
Depreciation	Physical wear and tear over time	No physical degradation (though may become obsolete)
Replication cost	Significant cost to produce exact copies	Virtually costless to create perfect copies

No "One Size Fits All" for Generative AI

There is no “one size fits all” generative artificial intelligence strategy. Instead, successful innovations will build on existing supplier strengths, competencies and business models. 

Microsoft's core business arguably includes productivity software suites like Office 365. Copilot helps by suggesting completions, refactoring code and identifying potential errors when customers use the tools. 

Apple’s business model, on the other hand, is driven by smartphones. So Apple is using AI for smartphone functions such as facial recognition (Face ID), voice assistant (Siri), and image optimization that are relevant features for people using iPhones. 

Google's core competency is information retrieval through search, which creates the advertising revenue streams that dominate Alphabet’s business models. So Google will use AI for natural language processing and machine learning to understand search queries better, rank results more effectively, and personalize search experiences.

Meta makes its money from social media platforms. So Meta is likely to prioritize use of AI for content moderation, personalized recommendations, and targeted advertising.

That focus on enhancing current products and value means Apple is focused on “on-device” processing to a greater extent than Microsoft or Google (with the exception of Google Pixel devices). Google search and Copilot can generally rely on remote processing. 

Amazon’s revenue reliance on e-commerce means AI will be deployed to improve product recommendations. 

But Microsoft, Google and Amazon will rely on AI to support cloud computing as a service operations (AWS, Google Cloud, Azure), in large part to support model hosting and training “as a service.” 

Generative AI, for example, can be used by many firms, for many purposes. By firms other than Google, Apple, Microsoft and Meta. But the actual degree of value will be tested over time. 

Though GenAI could well be something closer to a game changer for suppliers of GenAI models and inferences, most firms and industries trying to use those models and inferences face more-difficult challenges. 

As helpful as generative AI can be, it is not so clear that most implementations by most end user entities are going to produce measurable financial outcomes, at first. And where that happens, it might well be the case that cost reductions are the metrics, rather than revenue enhancements. 

source: Global X

Is Generative AI a "Game Changer?"

Generative artificial intelligence might sometimes be called a “game changer” for industries, which might refer to transformative impact and disruption.

That can happen in a number of ways, but measurably when a firm or industry is able to use an innovation to make significant shifts in market share, often by creating new markets or substantially altering existing ones. Think about search, social media or e-commerce in that vein. 

A game changer can produce exponential revenue growth by creating new markets. Think Uber, Lyft, Airbnb. In other cases, revenue growth can come from changing value chains, such as any “direct to consumer” shift, where distributors are taken out of the distribution chain. 

Game changers tend to have industry-wide impact, forcing all  competitors to adapt or risk becoming obsolete.

Game changers also tend to recreate value chains and sources of value, usually manifested in revenue, profit margin or market share statistics, often by enabling new business models. Ad-supported technology products are a prime example. 

Industry	Innovation/Technology	Impact
Media	Printing Press	Revolutionized information dissemination and literacy
Media	Radio	Enabled real-time mass communication
Media	Television	Combined audio and visual for widespread entertainment and news
Media	Internet	Democratized content creation and distribution
Media	Streaming Services	Transformed content consumption patterns
E-commerce	World Wide Web	Enabled online retail and digital transactions
E-commerce	Secure Online Payments	Facilitated trust in online purchases
E-commerce	Mobile Commerce	Allowed shopping from anywhere, anytime
E-commerce	AI-Powered Personalization	Enhanced customer experience and targeting
E-commerce	Cloud Computing	Enabled scalable and flexible online operations
Manufacturing	Assembly Line	Dramatically increased production efficiency
Manufacturing	Robotics	Automated repetitive tasks and increased precision
Manufacturing	Computer-Aided Design (CAD)	Improved product design and prototyping
Manufacturing	3D Printing	Enabled rapid prototyping and customized production
Manufacturing	Internet of Things (IoT)	Enhanced monitoring and optimization of production processes

Scalability--innovations that are broadly applicable--also is a hallmark of game-changing developments. The new technologies or practices should apply across many industries and functions. Innovations that affect a single industry are not generally “game changers.”

In that regard, machine learning or autonomous capabilities are more likely to be game changers than generative AI, even if GenAI has accelerated a shift in thinking about AI in general, in other forms.

"Free Speech" Protections Now Rely on Voluntary Private Actor Content Curation

Traditional protection of citizen "free speech" rights has aimed squarely at preventing government suppression. In the internet era, that might not be the biggest issue. Instead, it is suppression or curation by private entities that have a right to do so, which is different from saying they should do so.

In fact, many government actions seem quite trivial compared to the content moderation practiced by private gatekeepers.

The Federal Communications Commission wants U.S. telcos to block delivery of messages that use artificial intelligence to create fake political messages. Irrespective of the merits of that position, it does once again show how technology is making a shambles of the traditional regulatory models used to govern political discourse (supposedly protected “free speech”) in traditional print media; similarly protected content transmitted over telephone and data communications networks; slightly more regulated “broadcast TV” speech and unregulated internet media. 

Perhaps oddly, the FCC position shows an encroaching amount of content regulation over an industry and business that traditionally has been required to allow freedom of speech. The original intent of common carrier regulation, after all, has been to enable universal access with no restrictions on “what can be said” over a communications network. 

Generally speaking, content restrictions have been forbidden on communications networks; data networks (including internet media) and print media. There have been some restrictions or rules on broadcast TV and radio. 

But some creeping content regulation seems to be increasing. Despite what we might think about AI-generated content that could be considered “fake,” rules forbidding it are content regulations, no different in kind from rules mandating childrens’ programming, “public interest” programming, “obscene content” or “harassing” phone calls. 

The Communications Act of 1934 established the FCC to regulate both common carriers (like telephone companies) and content carriers (like broadcast media). However, technological advancements have blurred the lines between these categories, making consistent regulation challenging.

Traditionally, common carriers like telephone companies were expected to be neutral conduits of information, not regulators of content. However, the FCC has been given broad power to regulate broadcast media, creating a tension between these two approaches.

The FCC's content regulation, particularly regarding indecency, has raised significant First Amendment concerns. This creates tension with the principle of common carriers as neutral platforms for speech. The same might be said of bans on AI-generated content, as much as we might tend to agree that “fake content” is a problem. 

The ongoing debates over net neutrality and the regulation of social media platforms further highlight the changing context of content freedom and regulation. 

We might not, in the past, have seen communications networks as platforms supporting “freedom of speech” in the same vein as newspapers and magazines. But the whole concept of “common carriage” supports content freedom: networks cannot bar speech or the expression of ideas. 

Oddly enough, since “freedom of speech” principles only prohibit government entities (notably the federal government) from restricting free speech, more issues are arising from private entity restrictions on free speech (social media, search and other content moderation and curation policies that suppress some views and promote others). 

One might argue that the crucial freedom of speech issues now reside in the private sector, not government. Content curation by private firms seemingly has emerged as a much-bigger challenge to free expression of political ideas than government restrictions. 

As an adage suggests, “freedom of speech exists for those who own printing presses.” In other words, even if individuals have the right to their views, gatekeepers that can amplify ideas also exist, and did so in the pre-internet eras as well. 

These platforms employ content moderation policies and algorithmic curation that determine what content is amplified, suppressed, or removed entirely. This role is analogous to editorial decisions made by traditional media gatekeepers, but on a much larger scale and often with less transparency. 

The point is that protecting “freedom of speech” traditionally has meant restricting the power of governments to censor speech. In the internet era, most of the major issues happen with private actors against which citizens have no actual legal protections. 

As in earlier eras, it is the willingness of content curators to operate in a fair and unbiased manner makes the difference. There is a difference between the ethos of “presenting both sides” of an issue (actually there typically are many sides) and the different methods of “being transparent, rather than presenting a balanced view.”

How Much Will ACP Demise Affect Home Broadband?

It is not unusual to see estimates of U.S. households that will “lose their home broadband service” because a particular subsidy program is out of funding at about six million homes, or roughly 25 percent of households that had used the ACP subsidy, as estimated by researchers at Maravedis, for example. 

According to the Federal Communications Commission, some 21 million ACP accounts were in service in December 2023. 

The issue then is how many of those households will find some other way to obtain home broadband, as was the case before the Affordable Connectivity Program was funded. In other words, users of ACP subsidies often had already been buying home broadband service, or had chosen to use their smartphones, before the ACP was available. 

So the end of the ACP subsidy does not automatically mean all or most of those who used the program will simply stop buying home broadband services. By way of reference, it has been estimated that “low-income households” with less than $30,000 a year in income had home broadband buy rates in the 70-percent range before the ACP was funded. 

source: Pew Research

Also, according to Pew Research estimates, about 15 percent of U.S. adults say they rely on their smartphones to access the internet, though such practices are about 27 percent of adults with incomes less than $30,000 annually  and about 19 percent of adults with incomes less than $50,000 but at least $30,000 annually. 

 

source: Pew Research

By implication, perhaps 46 percent of U.S. adults with incomes less than $50,000 annually rely on their smartphones for internet access instead of buying home broadband. 

Keep in mind, though, that home broadband is purchased “by the household” while “smartphone internet access” is purchased by the person. 

So 86 percent of U.S. adults with incomes of $30,000 or less are estimated to use internet access services by some means, while 91 percent of U.S. adults with incomes of $30,000 to $50,000 have internet access. 

In total, between 93 percent and 95 percent of U.S. homes purchase home broadband services, according to Pew Research estimates.  

If one assumes that it was primarily the adults with annual incomes of less than $30,000 annually that were the bulk of the ACP subsidy recipients, and if ACP accounts numbered about 23 million households, with an average household density of about 2.5 persons, that suggests as many as 57.5 million U.S. adults might have been using the ACP (some of the household members were younger than 18). 

The Federal Communications Commission survey of ACP recipients in December 2023 said ACP recipients numbered 21 million. The FCC survey found that mobile-only service was used by 25 percent of survey respondents; 30 percent had both mobile and home broadband access and 23 percent already had home broadband service. About 22 percent reported they had no internet access prior to the ACP. 

The FCC survey also found that 11 percent of the “had no internet access service” respondents also said they had no need to access the internet. And more than 80 percent of those reporting no pre-ACP internet access said they used access services in some other way (library, school, business or friend’s account). 

Using the FCC’s 21 million ACP account figure, and subtracting the 11 percent of respondents who said they had no need to use the internet, perhaps 2.5 million of the ACP accounts were homes that wanted to use the internet and did not have a home broadband service prior to the ACP.

On the other hand, more than 50 percent of respondents who already had mobile internet access before the ACP was available said they used the subsidy to reduce the cost of mobile service they already had, and did not add a home broadband service. But a significant percentage (about 38 percent) said they did add home broadband using the ACP benefit. 

About 84 percent to 85 percent of respondents who already had some form of internet access used the ACP to reduce the cost of their existing plans. 

None of that is to minimize the remaining portion of the U.S. adult population that does not buy internet access or home broadband. But consumer choices also play a role. 

The whole point of public, school and business access is to allow people to use the internet without buying a personal subscription. Some consumers make rational choices not to purchase home broadband, instead using their mobile connections or free access of some sort.

Not every statistical artifact is the result of invidious action or circumstances.