Wednesday, October 4, 2023

"Fair Share" or Double Dipping?

European telco CEOs again pushed their demand for  "fair share" payments from a few hyperscale app providers in an open letter to EU officials. Some might argue the demand is anything but fair, since internet domains already compensate each other for asymmetrical traffic flows.


The argument is that mobile and fixed operators do not have the capital to build the next generation of connectivity networks. The letter says “at least €174bn of new investment will be needed by 2030 to deliver the connectivity targets.”


And the CEOs say they cannot do so. “The telecoms sector is currently not strong enough to meet that demand, with many operators barely earning their cost of capital,” the letter argues. 


Ignore for the moment “why” such claimed paltry returns exist, and whether it is not the CEOs of those firms whose job it is to boost returns, rather than others in the value chain who must do so. 


Ignore for the moment the notion that “those companies that benefit most from the infrastructure” are somehow obligated to reduce their profits to boost the profits of the ISP participants in the value chain. Some industries have low margins; others have high margins. 


The fact that all industries use seaports, airports, rail and road transport, electricity, wastewater systems, education and other infrastructure in and of itself tells us little about industry business models. 


Most people and businesses use the internet. That, in itself, does not drive business models for users. And most of the traffic asymmetry exists because an ISP’s own customers have asked for delivery of content from sources on other domains. 


Ignore for the moment the other calls for regulatory relief the industry has been asking for, such as more freedom for mergers that reduce the number of service providers operating in any market. 


Ignore for the moment the actions service providers might undertake on their own to improve their operating and capital investment models and practices; the industrial policy favoritism of domestic firms, compared to foreign firms in the value chain or the value of scale economies (most observers might say economies of scale are clear enough for app providers, but arguably not so important for internet access providers). 


Every retail business and consumer connected to the internet already pays the asked-for tariffs for such access, with the asked-for usage payments for usage. So some might argue ISPs want to be paid twice: once when traffic enters their domain and once when it leaves their domains. 


Internet service providers could do what all other businesses do when faced with cost increases: raise prices, change product attributes or make other changes so the business model remains intact. 


At the business-to-business level, internet domains already negotiate interconnection agreements for excess traffic “created” by either domain that is party to a bilateral agreement. So Netflix, Apple, Meta and Alphabet, for example, already pay fees to large ISPs for traffic asymmetries. 


Here are some examples of interconnection payments already made by hyperscale app providers to just a few ISPs. 


Hyperscale App Provider

ISP

Interconnection Payment

Netflix

Comcast

$1 billion

Netflix

Verizon

$750 million

Amazon Web Services

Comcast

$1.2 billion

Amazon Web Services

Verizon

$900 million

Microsoft Azure

Comcast

$1 billion

Microsoft Azure

Verizon

$750 million

Google Cloud

Comcast

$800 million

Google Cloud

Verizon

$600 million

Microsoft Azure

AT&T

$75 million per year

Alphabet

Charter

$100 million per year

Amazon

AT&T

$150 million per year

Microsoft

Charter

$75 million per year

Google

AT&T

$125 million per year

Meta

Charter

$50 million per year

Meta

AT&T

$75 million per year

Alphabet

China Telecom

$150 million per year

Amazon

NTT

$125 million per year

Microsoft

Deutsche Telekom

$100 million per year

Google

Telefónica

$75 million per year

Meta

Singtel

$50 million per year

Meta

Orange

$75 million per year




Monday, October 2, 2023

GPUs Fuel AI Processing Explosion

If you want to know why so many graphics processing units are being sold these days, and why sales are expected to increase, AI applications and use cases are a big part of the explanation. 


According to data published by the Allen Institute for Artificial Intelligence, computing intensity caused by use of artificial intelligence model training and inference operations will continue to grow at high rates, with computing cycle volume increasing by an order of magnitude over the next decade. 


In large part, those increases will be caused by AI adoption in most industries


Year

Computing cycle volume (exaflops)

2023

1.7

2024

2.4

2025

3.3

2026

4.6

2027

6.4

2028

8.9

2029

12.3

2030

16.7

2031

22.6

2032

30.4

2033

40.7

2034

54.3

2035

72.4

2036

96.6

2037

128.7

2038

171.2

2039

228.1

2040

304.2

Source: Allen Institute


Year

Projected increase in computing cycle volume due to AI training and inference operations (%)

2023

100%

2024

145%

2025

200%

2026

260%

2027

330%

2028

410%

2029

500%

2030

600%

2031

710%

2032

830%

2033

960%

2034

1100%

2035

1250%

2036

1410%

2037

1580%

2038

1760%

2039

1950%

2040

2150%

Source: Allen Institute


Sunday, October 1, 2023

AI is a General-Purpose Technology, Will Tend to Create Abundance in Place of Scarcity

The recently-settled strike of Hollywood writers, in substantial part related to the anticipated impact of artificial intelligence on writer employment and participation in residuals, is likely to be replicated--in concern about AI impact on employment and market value--in many other industries. 


Any examination of key technology developments over the past 30 years would suggest such an outcome, based on the impact lower marginal costs of personal computer, internet, mobile, and cloud computing industries have had on economic growth, productivity, the rise or fall of segments of the economy, jobs, value and income of workers in industries affected by the lower marginal cost of those technologies. 


Marginal cost declines have led to lower prices for products and services, which has increased consumer spending and boosted economic growth. But since one participant’s cost is another participant’s revenue, for every winner there is a loser. 


This has made these products and services more affordable for consumers, but reduced revenue and profits for suppliers. On the other hand, falling prices have the expected effect of higher usage. Lower costs of AI will encourage its wider use. And it would be hard to find examples of widespread new technology usage that did not affect supplier (and workers are suppliers of labor) prices. 


For example, the price of personal computers has fallen by more than 90 percent since the 1980s, and the price of internet access has fallen by more than 95 percent since the 1990s (price per unit), even using conservative assumptions about “average” or “typical” speed. Cost per Mbps of performance is therefore somewhat overstated by 50 percent or so in 2023, for example. 


Year

Typical speed (Mbps)

Absolute cost ($ per month)

Cost per Mbps ($ per Mbps)

Assumptions about median or average speed

1990

0.1

400

4000

Assumed that the typical speed in 1990 was 0.1 Mbps, based on the fact that the first commercial dial-up internet service was launched in 1990 and had a maximum speed of 0.144 Mbps.

1995

0.5

200

400

Assumed that the typical speed in 1995 was 0.5 Mbps, based on the fact that the first commercial DSL service was launched in 1995 and had a maximum speed of 1.5 Mbps.

2000

1

100

100

Assumed that the typical speed in 2000 was 1 Mbps, based on the fact that cable internet was widely available in 2000 and had a typical speed of 1 Mbps.

2005

5

60

12

Assumed that the typical speed in 2005 was 5 Mbps, based on the fact that broadband internet was widely available in 2005 and had a typical speed of 5 Mbps.

2010

10

50

5

Assumed that the typical speed in 2010 was 10 Mbps, based on the fact that fiber internet was beginning to be deployed in 2010 and had a typical speed of 10 Mbps.

2015

25

40

1.6

Assumed that the typical speed in 2015 was 25 Mbps, based on the fact that cable internet was widely available in 2015 and had a typical speed of 25 Mbps.

2020

50

60

1.2

Assumed that the typical speed in 2020 was 50 Mbps, based on the fact that fiber internet was more widely available in 2020 and had a typical speed of 50 Mbps.

2023

100

60

0.6

Assumed that the typical speed in 2023 is 100 Mbps, based on the fact that fiber internet is becoming more widely available and has a typical speed of 100 Mbps.


Marginal cost declines have also led to increased productivity, to the extent one believes the productivity of an office or knowledge worker actually can be measured. 


Disintermediation has been a hallmark of the internet era, spawning any number of “direct to consumer” distribution models” e-commerce; video and audio streaming being examples. 


Additionally, the rise of cloud computing has led to the substitution of remote computing “as a service” for ownership and operation of computing hardware and software. This has drastically reduced startup costs for software firms, for example. 


The point is that AI is going to affect the economy, assuming it is a general-purpose technology such as roads, electricity, computers, cloud computing, the internet or home broadband. 


And all those prior technologies have led to new industries being formed, work processes and demand being altered, value created or lessened in different parts of the value chain. Hollywood writers fear AI will replace some of what they do, leading to lower wages and residuals. 


Whatever the immediate perceived results of the settlement, the long-term impact is not likely eliminated. Work, life, the economy all are likely to be reshaped by AI, and not always in ways people might prefer.


Will Generative AI Follow Development Path of the Internet?

In many ways, the development of the internet provides a model for understanding how artificial intelligence will develop and create value. ...