Unicast Video Accounts for Most of the Internet Bandwidth Increases We See

Constant and significant increases in bandwidth consumption are among the fateful implications of switching from linear TV broadcasting to multicast video streaming. Consider that video now constitutes 52 percent to 88 percent of all internet traffic. 

Not all that increase is the direct result of video streaming services. Video now is an important part of social media interactions and advertising on web sites supporting consumer applications, though some studies suggest social media sites overall represent only seven percent to about 15 percent of video traffic consumed by end users. 

Also, there is some amount of internet video traffic between data centers, not intended directly for end users, possibly representing five percent of global internet traffic. 

Study	Date	Video Traffic Share (%)
Cisco Annual Internet Report (2023)	Dec 2022	88%
Sandvine Global Internet Phenomena Report (Q3 2023)	Sep 2023	83%
Limelight Networks State of the Real-Time Web Report (Q3 2023)	Oct 2023	76%
Ericsson Mobility Report (Nov 2023)	Nov 2023	72%
ITU Global Video Traffic Forecasts	Feb 2023	70% (2022)
Ookla Global Video Report (Q2 2023)	Aug 2023	65%
Akamai State of the Internet / Security Report (Q3 2023)	Oct 2023	60%
Statista: Global Internet Traffic Distribution by Content Type (2023)	Oct 2023	58%
GlobalWebIndex Social Video Trends Report (Q3 2023)	Sep 2023	55%
Juniper Networks Visual Networking Index (2023)	Feb 2023	52% (2022)

Ignoring for the moment the impact of video resolution on bandwidth consumption (higher resolution requires more bandwidth), the key change is that broadcasting essentially uses a “one-to-many” architecture, while streaming uses a unicast architecture. 

The best example is that a scheduled broadcast TV show, for example, can essentially send one copy of the content to every viewer (multicast or broadcast delivery). The same number of views, using internet delivery, essentially requires sending the same copy to each viewer separately (unicast delivery). 

In other words, 10 homes watching one multicast or broadcast program, on one channel, at one time consumes X amount of network bandwidth. If 10 homes watch a program of the same file size as the broadcast content, whether simultaneously or not, then bandwidth consumption is 10X. 

There are some nuances for real-world data consumption, such as the fact that consumption of linear video is declining or the fact that broadcasting uses a constant amount of bandwidth, no matter how many viewers in an area might be watching or not watching. 

Study	Comparison	Bandwidth Ratio (Streaming/Broadcasting)
"A Comparative Analysis of Video Streaming and Broadcasting for Live Sports Events" (2023)	Live sports streaming vs. multicast	10x - 15x
"Bandwidth Efficiency of IPTV vs. Traditional Broadcasting" (2022)	IPTV unicasting vs. terrestrial broadcasting	2x - 4x
"The Impact of Unicast Video Delivery on Network Traffic" (2021)	Unicasting video vs. multicast video	1.5x - 3x
"Comparing the Bandwidth Consumption of Live Streaming and P2P Delivery" (2020)	Live streaming vs. P2P for live events	3x - 6x
"The Bandwidth Efficiency of Video Streaming Protocols" (2019)	HTTP streaming vs. RTMP streaming	1.2x - 2x
"A Study of User-Generated Video Delivery on Social Media Platforms" (2018)	User-generated video streaming vs. traditional video streaming	2x - 4x
"The Bandwidth Implications of 4K and 8K Video Streaming" (2017)	Higher resolution streaming vs. standard definition	4x - 8x
"The Impact of Mobile Video Streaming on Network Congestion" (2016)	Mobile video streaming vs. fixed-line streaming	1.5x - 3x
"The Future of Video Delivery: A Cost Comparison of Streaming and Broadcasting" (2015)	Streaming vs. broadcasting for future content delivery	2x - 4x
"The Bandwidth Efficiency of Video-on-Demand Services" (2014)	Video-on-demand streaming vs. linear broadcasting	1.5x - 2.5x

There are other nuances as well. Since a broadcast video stream often is viewed on a television set, it is possible that multiple viewers “share” viewing of the same content. If one TV is receiving a program, and five people are watching, the “single delivery” supports five views. 

On a “per viewer” basis, X amount of delivery bandwidth is X/5 for each viewer of the same program. 

If five people watch a program of equivalent file size at the same time, data consumption is 5X. 

Study	Year	Methodology	Streaming Bandwidth (Mbps)	Linear Broadcasting Bandwidth (Mbps)
Nielsen	2022	Network traffic analysis	3.1-4.7 (average)	0.1-0.2 (average)
OpenVault	2023	ISP data analysis	1.8-2.5 (average)	0.05-0.15 (average)
Pew Research Center	2021	Survey and network analysis	2.3-3.8 (average)	0.1-0.2 (average)
University of Zurich	2019	Network monitoring and simulation	2.0-3.5 (average)	0.08-0.18 (average)
Akamai	2020	Global traffic analysis	1.6-2.8 (average)	0.04-0.12 (average)
Sandvine	2022	Network traffic analysis report	3.5-5.0 (peak)	0.15-0.25 (peak)
Netflix	2021	Open Connect content delivery platform report	0.5-1.5 (average)	N/A
BBC Research & Development	2018	HbbTV hybrid broadcasting analysis	1.0-2.0 (combined)	0.03-0.08 (combined)
Bitmovin	2023	Video encoding and delivery technology report	0.8-1.8 (efficient encoding)	N/A
Ericsson	2022	Mobile network video traffic report	0.5-2.0 (mobile average)	N/A

The point is that the shift from broadcasting (multicasting) to unicast entertainment video was destined to dramatically increase internet data consumption.

Large Language Model Ingestion of Content is Not Necessarily Copyright Infringement

At least for the moment, I find myself unpersuaded that strong copyright protections some demand from use of large language models are a good idea. To be sure, as a practical matter we should expect some sort of reasonable licensing system to develop. 

Some system will emerge that compensates copyright holders for potential use of their work to train LLMs. So long as the costs are reasonable, there is little danger of stifling LLM progress. 

The big issue is some sort of legal barring of LLM use of copyrighted material for training purposes. After all, all human knowledge builds upon the past, including consumption of copyrighted material.

Copyright law protects expression, not ideas. Copyright protects the specific way an idea is expressed, not the idea itself. So long as an LLM consumes copyrighted material including books or articles, and its outputs differ significantly in form and originality, copyright infringement seems dubious. 

The transformative use doctrine, for example, allows for using copyrighted material in new and creative ways without infringing copyright, as long as it serves a different purpose or adds meaningfully to the original work. 

LLMs that use copyrighted material to generate summaries, translations, or even new creative interpretations could potentially fall under this category, depending on the nature of the transformation.

Fair use exceptions allow for limited use of copyrighted material for purposes like criticism, commentary, or research without permission. LLMs used for training could potentially fall under fair use, depending on the amount and nature of the copyrighted material used.

Just because LLMs are efficient should not necessarily infringe copyright protection, though the issue of derivative works--which can be protected--is an issue.

Potential solutions might include greater transparency and attribution; licensing models and clearer legal frameworks for LLM and copyright. Ideas and knowledge are not protected by copyright. Only the expression is protected. 

That LLMs ingest lots of copyrighted material is similar to the ways humans learn, but with vastly-greater efficiency. Where there are legitimate copyright infringement concerns is not the ingestion, but the production (“expression” of an idea).

Economic Impact from Large Language Models Should Rival the Internet

Keeping in mind the wide range of assumptions that must be made about “total value creation” by any single general purpose technology, and with the important caveat that the time frame over which that contribution is measured, we can still note that estimates for large language models alone are sizable. 

GPT	Study Name	Published Date	Publication Venue	Estimated Value Creation
Electricity	The Economic History of Energy	1989	Stanford University Press	$50 trillion (1980 USD) by 1980
Internal Combustion Engine	The Motorcar and Industrial Revolution	1984	Allen Lane	$25 trillion (1980 USD) by 1980
Internet	The Internet Economy	2015	MIT Press	$10 trillion (2015 USD)
Internet	Digital Economy Report 2021	2021	UNCTAD	$15.5 trillion (2020 USD)
GPT-3 (Early Estimates)	The AI Revolution: The Road to Superintelligence	2016	Viking	$15 trillion (2016 USD) by 2030
GPT-3 (Early Estimates)	Economic Impact of Large Language Models	2022	PwC	$6 trillion (2025 USD)
Electricity	The Economic Growth of Nations	1963	Economic Journal	15-30 trillion
Electricity	The Impact of Electricity on American Growth and Welfare	1992	Journal of Economic History	10-20 trillion
Internal Combustion Engine	The Automobile and the American Economy	1955	American Economic Review	5-10 trillion
	The Internal Combustion Engine and US Economic Growth	2007	Explorations in Economic History	7-15 trillion
Internet	The Internet Economy: Opportunities and Challenges	2011	McKinsey Global Institute	5-10 trillion
	The Value of the Internet: Evidence from the Dark Web	2023	Science	8-15 trillion
Large Language Models (LLMs)	The Economic Potential of Artificial Intelligence	2022	World Bank	1-10 trillion  (by 2030)

Annual contributions would of course run in the “billions of dollars” range. 

All such analyses are highly dependent on the research assumptions and dates studied, so the figures are not directly comparable. 

GPT	Study Name	Publication Date	Publishing Venue	Estimated Annual Value Creation (Trillion USD)
Electricity	The Economic Effects of Universal Electrification	2016	Journal of Economic Growth	0.5-1.0
Internal Combustion Engine	The Impact of the Automobile and Truck on the U.S. Economy	2004	Explorations in Economic History	0.2-0.5
Internet	The Internet's Global Value Add	2023	Citigroup	8.0-15.0
Large Language Models (LLMs): GPT-3	The Economic Potential of Large Language Models	2022	McKinsey Global Institute	0.1-1.0
Quantum Computing	Quantum Computing: Economic Opportunities and Challenges	2021	National Academies Press	0.1-1.0

We can’t yet be confident that we know how extensive AI impact will have on economies and industries. But if AI does become a GPT, its impact will be substantial, perhaps on the order of the impact the internet itself has had.