Markets for "Good Enough" Home Broadband are Substantial

As far as home broadband platforms go, fixed wireless, though faster than digital subscriber line, is far less capable than fiber to home or hybrid fiber coax platforms. So why is fixed wireless growing as a percentage of U.S. home broadband accounts?

In 2023, for example, virtually 100 percent of U.S. net home broadband net account additions for the internet service providers with 90 percent of more of total market share.

Though we often seem to focus on the headline speeds and services, customers do not always want to buy those services.

Price-value relationships seem to matter most. In many parts of the United States, the competition is DSL and HFC. And there, fixed wireless is faster than DSL and more affordable than HFC. 

By most estimates, only about 30 percent of U.S. home locations have the ability to buy a fiber-to-home service. If so, then roughly 70 percent of the U.S. home broadband market is potentially amenable to fixed wireless purchases for those customers who want speeds faster than DSL but do not wish to pay the going rate for HFC service at speeds above 200 Mbps. 

In 2023, for example, virtually all net account additions in the U.S. market were supported by fixed wireless. Both FTTH and HFC platforms lost net accounts, according to Leichtman Research Group. 

Broadband Providers	Subscribers at end of 2023	Net Adds in 2023
Cable Companies
Comcast	32,253,000	-66,000
Charter	30,588,000	155,000
Altice	4,517,900	-114,100
Cable One	1,059,300	-1,100
Breezeline^	663,286	-29,184
Other major private companies,	7,020,000	-8,000
Total Top Cable	76,101,486	-63,384
Wireline Phone Companies
AT&T	15,288,000	-98,000
Verizon	7,650,000	166,000
Frontier^	2,943,000	75,000
Lumen	2,758,000	-279,000
Windstream	1,175,000	0
TDS	539,800	29,800
Consolidated	393,219	25,761
Total Top Wireline Phone	30,747,019	-80,439
Fixed Wireless Services
T-Mobile	4,776,000	2,130,000
Verizon^	3,067,000	1,536,000
Total Top Fixed Wireless	7,843,000	3,666,000
Total Top Broadband	114,691,505	3,522,177

              source: Leichtman Research Group

Over time, the percentage of customers who will be able to buy a fiber-to-home service will grow. Over time, the "typical" speeds customers require also will grow. So FTTH remains the platform of the future. 

It is just that, between now and then--and even if FTTH is eventually available to nearly the whole market, segments will continue to exist. 

The distribution of buyers might still resemble a "bell curve skewed to the right" in terms of segments. Most customers will still be in the center of the curve, with a bigger high-end tail than the low-end tail. 

Will AI Inevitably Lead to "More Leisure Time?"

It might seem obvious that artificial intelligence will “inevitably” lead to an increase in work productivity, which might or might not lead to a reduction in “work hours” or leisure time, based principally on AI’s ability to automate routine tasks. 

The historical record is likely clearer on productivity, but perhaps less clear for impact on leisure time, in some cases. Household appliances arguably did reduce the time spent on house cleaning chores, for example. 

It is arguably less clear that personal computers or the internet have had uniformly positive benefits for leisure time. Personal computers, smartphones and the internet  allow people to conduct work from home, “outside of work hours,” for example, so leisure time might be negatively affected. 

Technology	Impact on Leisure Time	Study/Source
Electricity	Increased leisure time through reduced household chores (washing, ironing)	Studies on the impact of electricity on household labor https://www.sciencedaily.com/terms/electricity.htm)
Washing Machine	Freed up time spent on laundry, potentially leading to more leisure time	Studies on the impact of the washing machine on women's time https://scholarworks.boisestate.edu/cgi/viewcontent.cgi?article=1175&context=mcnair_journal
Personal Computer	Debated impact. Increased efficiency in some tasks, but also created new demands (email, social media)	Studies on the impact of personal computers on work-life balance https://www.sciencedirect.com/science/article/abs/pii/S0148296321001570
Internal Combustion Engine (Cars, Trucks)	Increased productivity through faster transportation of goods and people, potentially creating more leisure time for individuals	“The Impact of the Automobile on American Life,” John B. Rae, 1971; “Trucks, Trains and the Engine of Prosperity: Productivity Gains in Freight Transportation, 1880-1950,” Kenneth A. Snowden, 2002; “Internal Combustion Engines and American Economic Growth, 1860-1920,” Peter L. Rousseau, 2002; “The Productivity Effects of Technological Change in Highway Transportation,” Lester Lave, 1966
Assembly Line (Manufacturing)	Drastically increased productivity in factories. Studies suggest a 400 percent increase in Model T production within a few years.	"Fordism: Historical and Theoretical Perspectives" by Jonathan H. Freeland (2002)
Steam power	Increased output significantly, particularly in textiles and iron production. Estimates suggest a 2-3% annual growth rate in manufacturing due to steam power.	"The Productivity Effects of Steam Power in British Manufacturing" by Nicholas Crafts (2004)

Modern IP Networks are Messy, Compared to PSTN

Every now and then, when teaching students about connectivity architecture and business trends, I have to describe the differences between public networks as they existed pre-internet and the way networks exist today, in the internet era. 

It always is messy, as we use both cabled internet data infrastructure, alongside with mobile, satellite and older voice networks, as well as various types of local access and indoor networks. Looking just at the global backbone networks will illustrate the differences. 

The public switched network was built to support voice calls, and was centralized and hierarchical, connecting voice switches (Class 4 for wide area transport; class 5 for local distribution and end user connections). Traffic origination was the class 5 switches. 

The packet switched network underpinning the internet is a distributed network featuring routers and servers connected over optical wide area transport networks supporting any sort of media type. 

Traffic origination is generally content or app servers interacting with end user locations. 

So one big difference is that the PSTN was built to connect the voice switches, while the global packet network is built to support servers and computing devices (data centers on one hand and personal computers or other devices on the other). 

Where the PSTN used core network active elements, the packet network is more organized around edge elements, with the wide area network basically consisting of routers that forward the traffic. 

So one difference between the PSTN and internet networks is that while traffic on the PSTN was mostly generated by telcos, a majority of traffic on the internet is generated by app providers and large data centers. By Cisco estimates, more than half of global wide area network traffic flows between data centers.  

Data Transport Provider	% of Global Data
Traditional Telcos	30%
Cloud Providers (AWS, Azure, Google Cloud, others)	40%
Content Delivery Networks (CDNs)	15%
Other (Satellite Providers, ISPs)	15%

The older PSTN was hierarchical, deterministic and well-structured. The internet-supporting packet networks are flatter, more distributed, less deterministic and heterogeneous. 

Fixed network access networks remain more recognizable, connecting end user customers with a traffic aggregation point or point of presence (generically, class 5 switch or equivalent for PSTN; ISP data center for packet and internet traffic). The difference for mobile and all other fixed networks is the use of radios for access, rather than cabling of some sort. 

It is the fixed network core and “long haul” networks that are different. The PSTN core network connects class 4 and then class 5 switches. The packet network connects points of presence and data centers. The modern mobile network is essentially a packet network with wireless radio tails. 

Satellite networks also have been different, as there is really no distinction between the core transport and access networks. Ground stations transmit and receive, with satellites essentially acting as signal relay and retransmission elements. 

But the physical architectures (network elements)  arguably pale in importance compared to the logical architectures (protocols and how data moves through the networks).

The AI Era of Computing Marks a Sharp Break From Prior Eras

With the caveat that we have not reached agreement on what to call the present and immediately past periods of computing (some prefer “mobile” for the present era while others prefer “personal computing” or “internet” appellations), artificial intelligence is a candidate to define the era that might follow. 

And though AI will include changes in processors and become part of most software experiences, AI is likely to mostly represent a change in functions.

Where in the past computing was mostly about executing instructions, AI is more about learning, predicting and problem solving (making inferences) based on knowledge of massive data sets. 

Era	Processor Hardware	Key Software	Platforms
Mainframe (1950s-1970s)	Central Processing Unit (CPU) - bulky, expensive, limited processing power	Punch cards, batch processing, COBOL programming	IBM mainframe systems (OS/360)
Personal Computer (1970s-1990s)	Microprocessor (CPU) - smaller, more affordable, increased processing power	Word processing, spreadsheets, databases, early games	DOS, Windows, Apple OS
Client-Server (1990s-2000s)	Improved CPUs, increased RAM	Email, web browsing, enterprise applications	Microsoft Windows Server, Novell Netware, Unix/Linux
Personal Computing (1980s-2000)	PCs, laptops	Productivity suites, web browsing, social media	Windows, macOS
Mobile (2000s-Present)	Mobile processors (low power, high efficiency), multi-core CPUs	Mobile apps, social media, streaming services	iOS, Android

Prior eras of computing were mostly about telling machines what we wanted them to do. The AI era will be the first where machines will try to figure out what else we might want to do. 

Prior eras of computing had a focus on tasks: Humans  provided the program, data, and commands, and the computer delivered the desired output. The machines lacked the ability to understand the context or intent behind our instructions and operated on a literal level.

The coming AI era will be different, as machines increasingly anticipate human wants or needs. 

AI will aspire to understand our needs and desires, even if we haven't explicitly stated them. 

In other words, the machines will attempt to predict what happens next, based on past behavior and context. AI also will continually learn from user interactions and data to improve its ability to anticipate wants.

In other words, AI will try to be proactive, suggesting solutions, completing tasks, or providing information before we even realize we need it. And that is a big change from all prior eras of computing.

What Value Will AI PCs Possibly Provide?

Nobody knows yet what we will eventually come to know as an “AI PC.” But it is almost certainly going to be more than new keys that offer shortcuts to one or more generative AI models and services, as currently is the developing practice. Almost everyone expects that AI PCs will include hardware allowing on-board processing or faster processing. 

Some point to neural processing units--specialized chips designed for efficient AI tasks--that offload AI workloads from the CPU and GPU, improving overall performance. Others expect significant increases in random access memory and faster storage solutions probably using solid state devices. 

Others might point to the use of AI assistants that anticipate user needs and proactively offer suggestions, such as automatically summarizing documents you're reading or translating content on the fly. The assistant would understand your workflow and automate repetitive tasks, providing a “context-aware” experience. 

The assistant could pre-fetch relevant data, suggest actions based on your activity, and personalize the user experience.

Most think AI PCs will conduct on-device processing for at least some operations, avoiding full reliance on remote processing. But one has to ask "So what?" as a rhetorical quesiton. What experiences and use cases will benefit from local processing?

Personalization, sure. Image processing or speech-to-text, sure. Still, what improvement might be possible, compared to using remote processing? The issue isn't "using AI," but rather where we do the processing.

To a greater or lesser extent, AI PCs might be able to learn and adapt over time, personalizing their functionalities to a user’s specific needs and preferences.

And some might believe AI PCs will be open enough to allow developers to create new AI applications and extend the capabilities of the PC. 

But it remains unclear at this point that the suggested use cases for AI necessarily require on-board processing. Much of the value of AI could still be provided by remote processing. 

For example, it is not clear that on-board processing has much value for use of AI assistants to brainstorm ideas, generate creative text formats, or streamline workflows for writers, designers, and students.

And though on-board processing is deemed an advantage for smartphones, AI-PC editing tools might not add much value, as such editing is likely not to require real-time adjustments. 

Automation of repetitive tasks such as data entry or file organization likewise might not provide much value over remote processing. 

Most personalized learning or education use cases would be using remote data bases in most, if not all, cases, so remote processing might also be adequate to support those activities, and not require on-board processing. 

Similarly, it might be questionable whether on-board processing is need to support many entertainment and gaming use cases, as access to remote databases would be required in any case. 

Remote AI is likely good enough for personalizing entertainment recommendations based on user preferences and watch history; dynamic, adaptive storylines for gaming or music composition.

AI assistants might arguably be more important for voice commands and speech-to-text or language translation “on the fly.” 

But other potential use cases, aside from facial recognition for security purposes, might work just as well using remote computation. Many security monitoring tasks, for example, could be handled that way. 

The point is that AI PCs are likely to be a niche product initially, partly because of high cost, and partly because there will not be many consumer applications that actually benefit. 

AI Won't Help Solve "Wicked Problems"

As effective as artificial intelligence can be at solving some problems, AI will be hard pressed to solve some sorts of problems, as much as some have hoped “experts” in the past could “solve” difficult social, economic or political problems involving conflicts between winners and losers. 

The term wicked problem has become a standard way for policy analysts to describe a social issue whose solution is inherently elusive. By definition, wicked problems are difficult to define and even harder to solve because aspects of a problem are interconnected with others that also must be solved. 

Also, causes and effects of the problem can be complex and not fully understood. Also, potential solutions have externalities, when the solution to one problem creates unintended consequences that become new problems themselves. These consequences can be positive or negative, but they are external to the original problem's intended solution.

Especially in any area of public policy, value judgments also must be made, as stakeholders have different interests that often conflict. As any proposed solution inevitably involves winners and losers, AI can only suggest solutions. Political, social and economic interests will affect whether any proposed solution can be implemented. 

Classic examples of wicked problems include climate change, substance abuse, international relations, health care systems, education systems, and economic performance. 

For wicked problems, data may be scarce, unreliable, or biased, while causal relationships might be quite difficult to pinpoint. 

Correlation, as the old adage goes, is not necessarily causation.