Thursday, August 15, 2024

How Many Firms Will See Payback on Generative AI, and How Soon?

Though some pioneers claim they already are seeing revenue gains from generative artificial intelligence, we are probably justified in some skepticism about those outcomes.


A Gartner survey of 822 business leaders, conducted between September and November 2023, suggests that various generative AI projects cost between $5 million to $20 million. But that might not be the biggest impact, as costs for inference operations (asking questions, getting answers) could run between $8,000 to $21,000 per user. 


For a 1,000-user firm, that might suggest $8 million to $21 million annually in inference operations. 


source: Gartner 


And there is a bit of a contradiction in the reported results. Gartner notes that GenAI requires a higher tolerance for indirect, future financial investment criteria versus immediate return on investment (ROI). 


That noted, survey respondents have reported 15.8 percent revenue increases, 15.2 percent cost savings and 22.6 percent productivity improvement, on average.


One suspects we should take those quantifiable results with a bit of skepticism, as most of the returns from GenAI are indirect and hard to measure. 

There’s a reason increasing use of generative and other forms of artificial intelligence is linked to data center capacity: model training is getting more compute intensive. So large language model training costs are growing. 


And model creation and training might not be the biggest cost. 


 

source: Epoch AI


Some of us would not be at all surprised if disappointment with GenAI outcomes becomes more pronounced as projects seem not to provide the anticipated financial outcomes, in the near term. 


To the extent AI is the next general-purpose technology, as was the internet, we could ask the same questions about near term return from internet investments. 


How many firms will see near-term and quantifiable revenue upside from their capital investments and operating expenses directly related to GenAI? 


Outside of graphics processing unit suppliers; cloud "AI as a service" providers and big system integrators such as Accenture--who should be able to point to quantifiable revenue gains--not many end user firms will be so lucky.


We are likely years away from a substantial number of firms being able to say they can quantify revenue gains from using GenAI.




Tuesday, August 13, 2024

"You are the Product" Works Because Users Get Value

It is commonplace these days for observers to note that if a user is not paying for a product, then the user is the product. Sometimes that is viewed as a bad thing, but there is a bargain being struck here: users get value in exchange for being subjected to advertising.


And that is a time-tested value proposition. Users and customers get free or reduced-cost products they value for less money than would otherwise be the case. 



Product

Form of Subsidy

Value for Users

Free Online Games

In-app purchases, advertisements

Free gameplay, access to new levels/characters

Free Mobile Apps

In-app purchases, advertisements

Free core functionality, additional features for purchase

Free Video Streaming Services

Subscription model, advertisements

Free access to content, ad-free viewing option

News Websites

Advertisements, paywalls

Free access to news content, in-depth articles for subscribers

Social Media Platforms

Advertisements, premium subscriptions

Free connection with others, enhanced features for paid users

Free Email Services

Advertisements, paid storage upgrades

Basic email functionality, additional storage and features for a fee

Open-Source Software

Donations, corporate sponsorships

Free access to software, potential for customization and improvement


In principle, other forms of subsidy, discounts or premium value also are common. 


Product

Value for Users

Wholesale Clubs (Costco, Sam's Club)

Bulk discounts on groceries and household goods

Streaming Services (Netflix Premium, Spotify Premium)

Ad-free viewing/listening, access to exclusive content

Gyms and Fitness Centers (Monthly Memberships)

Access to workout facilities, classes, and potentially personal training

Subscription Boxes (Beauty, Food, etc.)

Curated selection of products delivered regularly, often at a discount

Cloud Storage Services (Dropbox Plus, iCloud+)

Increased storage capacity for documents, photos, and other files

Software Subscriptions (Adobe Creative Suite, Microsoft 365)

Access to the latest software updates and features, often with cloud storage included


The point is that as much as some decry the use of advertising, sponsorships and memberships, these are simply ways of creating value for buyers, offering lower prices and discounts, free access or perceived higher value. 


Monday, August 12, 2024

AI to Boost ASIC and FGPA Sales

As artificial intelligence is driving demand for data center capacity, connectivity and graphics processor units, so AI should increase demand for Application-Specific Integrated Circuits (ASICs) and Field-Programmable Gate Arrays (FPGAs). 


Perhaps the only question is what the rate of growth might be. 


According to MarketsandMarkets, the global AI chip market is expected to grow from $7.6 billion in 2020 to $57.8 billion by 2026, at a compound annual growth rate (CAGR) of 40 percent. 


Study Title

Publication Date

Publisher

ASIC Market Growth (CAGR)

Artificial Intelligence Chip Market - Trends, Forecast and Competitive Analysis

May 2024

Mordor Intelligence

32.50%

The Future of AI Processors: A Comparative Analysis of ASICs, FPGAs, and Neuromorphic Computing

March 2024

McKinsey & Company Report

35.8% (High Performance)

AI Hardware Market: 2024-2029

July 2024

Grand View Research

31.20%

The Rise of Domain-Specific Architectures for AI Workloads

June 2024

Gartner Research Report

38.1% (Forecast to 2027)

How AI is Transforming the Semiconductor Industry

February 2024

Deloitte Insights Report

34.70%

Data Center Capacity to Double Over the Next Four Years

Amazon, Microsoft and Google now represent 60 percent of all hyperscale data center capacity, while doubling their capacity over the last four years, says Synergy Research. The firm expects total hyperscale data center capacity will double again in the next four years.


Synergy Research sees non-hyperscale colocation data centers accounting for 22 percent of worldwide capacity, with on-premise data centers rounding out 37 percent of the total. 


source: Synergy Research


In turn, that capacity drives demand for connectivity as well, as a huge portion of global long-haul capacity is used to connect data centers. Estimates peg the percentage of “data center to data center” capacity between 30 percent to 50 percent of total. Add in connections to internet points of presence and as much as 75 percent of total global backbone networking capacity connects data centers with other data centers and internet points of presence. 


Category

Global Capacity

Data Center Interconnections

30-50%

Internet Points of Presence (PoPs)

30-40%

Enterprise Networks

10-15%

Other

5-10%


The advent of artificial intelligence is likely to drive capacity trends. 


The AI "data center to data center" interconnection services market is expected to grow from $4.2 billion in 2023 to $9.1 billion by 2027, a CAGR of 16.8 percent, according to analysts at Gartner.


Gartner also forecasts the overall "data center to data center" capacity market is forecast to grow from $48 billion in 2023 to $72 billion by 2027, a CAGR of 10.6 percent.


According to IDC, the AI-driven "data center to data center" interconnection services will account for 22 percent of the total "data center to data center" capacity market by 2025, up from 15 percent in 2023.


The global "data center to data center" capacity market is projected to reach $64 billion by 2025, growing at a CAGR of 12.7 percent from 2023, IDC also believes.


Forrester researchers believe the market for AI-enabled "data center to data center" interconnection services will grow at a CAGR of 19 percent from 2023 to 2027, reaching $8.5 billion by 2027.


Forrester also estimates the overall "data center to data center" capacity market will grow at a CAGR of 11 percent from 2023 to 2027, reaching $68 billion by 2027.


Markets and Markets predicts the AI "data center to data center" interconnection services market will grow from $4.6 billion in 2023 to $10.2 billion by 2027, at a CAGR of 17.3 percent .The global "data center to data center" capacity market is forecast to grow from $51 billion in 2023 to $78 billion by 2027, at a CAGR of 11.2 percent, the firm says. 


The point is that if data center capacity keeps doubling every four years, then  “data center to data center” connections are going to grow as well. The issue is “how much” growth will be needed and how long the trend might last.

Saturday, August 10, 2024

Is the Print Disruption Relevant for Video Entertainment?

It might be fair to say that video entertainment distributors and content providers face business model challenges that rival those faced by publishers decades ago, when digital media disrupted physical media. 


For example, total revenue for newspaper publishers dropped significantly from $46.2 billion in 2002 to $22.1 billion in 2020, a 52 percent decline, according to the U.S. Census Bureau. Estimated revenue for periodical publishing, which includes magazines, fell by 40.5 percent over the same period.


U.S. Census Bureau 


Print advertising revenue saw a significant decline, from $73.2 billion in 2000 to $6 billion in 2023. 


Similar damage was seen in the video on demand business, where video tape and disc rental revenue decreased by 88.5 percent.


The point is that advertising revenue began a long shift to digital and online media after 1996, now perhaps representing 75 percent of all U.S. advertising. 

source: Omdia 


How suppliers of linear video products will fare is the issue. Warner Brothers Discovery just booked a $9 billion impairment charge, while Paramount  took a charge of about $6 billion, for example, reflecting a devaluation of linear video asset value.


And while it is easy to note that video streaming is the successor product to linear TV, the business models remain challenged, for distributors and content suppliers. 


Distributor revenue is an issue, as linear TV could count on both advertising and subscription revenues, while video streaming services mostly rely on subscriptions, though ad support is growing.


The other revenue issues are lower average revenue per user or per account, compared to linear TV and high subscriber churn. 


Also, content costs have been higher than for linear services, as original content uniqueness has been more important than for most linear channels. Customer acquisition costs and marketing expenses also are higher in a direct-to-consumer environment. 


Video content providers also face significant business model challenges as the industry shifts away from traditional linear distribution, in large part because carriage fees paid to content owners by distributors are not available, while the core revenue model shifts from reliance on distributors to a “direct-to-consumer” model. 


The linear model provided large potential audiences with a wholesale sales model, where the customer was the video distributor.  The DTC model has to be created from scratch, and requires actual retail sales. 


Also, up to this point revenue has been based on subscriber fees, and most streaming services have struggled to gain scale.  


Linear video also featured a clear windowing strategy and revenue opportunities (when content was made available theatrically, then to broadcast, followed by cable and then syndication). Steaming is much more complex, with less certain revenue upside. 


In principle, linear programming formats were geared to large audiences while much streaming content is aimed at niches and segments. That arguably represents higher risk and higher unit costs. 


Using the prior example of what happened with print content as online content grew, we might well expect profit margins for linear media to decline as content consumption shifts to streaming formats and linear scale is lost. 


We should also see a shift of market share from legacy providers to upstarts, with perhaps the greatest pressure on niche or specialty formats such as industry specific sources. At some point, as legacy industries consolidate and shrink, there is less need for specialized media or conferences, for example. 


The demise of virtually all the former telecom industry; personal computer and enterprise computing  events and business media provide an example. 


Content producer shares also could shift. We already have seen new “studios” including Netflix, Amazon Prime, Apple TV and others arise, for example. 


To the extent they were able to survive, print media had to innovate revenue models, moving from print advertising and subscriptions to digital ads and paywalls, though the scale of the new businesses was less than the former business models. 


Similarly, video streaming services are exploring tiered subscriptions, ad-supported models, and content bundling, though we might speculate that the new business could well be smaller than the former industry. Some degree of fragmentation should be expected, where the legacy providers have less scale than before. 


All that could lead to a “professional” video entertainment industry that is smaller than it has been in the past, as hard as that might be to imagine. The model there is the growth of social media and user-generated content revenues compared to “professional” content provider revenues. 


For example, it is possible that digital platforms--Netflix, Prime Video and others--replace the traditional “TV broadcasters” or “cable TV networks” as the dominant “channels.” 


It also is conceivable that the content producers (studios) emerge as dominant in their “direct to consumer” roles, disintermediating the present distributors (TV broadcast networks, cable TV networks). Or, legacy networks and distributors might find some way to continue leading the market, as unlikely as some believe that is to happen. 


In the medium term, we might see a pattern similar to the former “print” industry, where a few legacy providers continue to have significant share, but where upstart new providers also have arisen. The legacy video networks might continue to represent venues for highly-viewed events, although much of the rest of viewing is fragmented among many suppliers.  


If the earlier transformation of “print” content is relevant for video entertainment, legacy businesses might never again be as large or important as they once were.


Friday, August 9, 2024

Synthetic Data Might be Quite Useful for Domain-Specific or Privacy-Critical Use Cases

There might be upsides and downsides as generative artificial intelligence systems--after crawling the whole internet--likely start to learn from each other. To be sure, some new data stores conceivably can be crawled, but that process will increasingly be expensive and involve much smaller, more-specialized sets of data, such as some proprietary enterprise content. 


But all that will be incremental. What is likely to happen is that models start to learn from each other, using “synthetic data” that is artificially generated mimicking real-world data in its statistical properties and structure, but without actual real-world data points. 


That could have both good and bad implications. Perhaps synthetic data can help compensate for scenarios where training data is under-represented or unavailable. That can help improve model performance and robustness. 


Model Type

Benefits from Synthetic Data

Domain-Specific Models (e.g., medical, legal, financial)

Access to large, private, and high-quality datasets is crucial for performance. Synthetic data can bridge this gap.

Models for Low-Resource Languages

Synthetic data can augment limited real-world data, improving model performance for languages with fewer available resources.

Models Requiring Diverse and Sensitive Data

Generating synthetic data can protect privacy while providing exposure to diverse scenarios, reducing biases.

Models for Data Augmentation

Synthetic data can expand training datasets, improving model robustness and generalization.


Since synthetic data doesn't contain real individuals' information, it can be used to train language models on sensitive topics without risking privacy violations.


Carefully generated synthetic data can be used to balance datasets and reduce biases present in real-world data, potentially leading to fairer language models. 


In domains where real data is scarce or expensive to obtain, synthetic data might provide a viable alternative for training language models. Cost effectiveness is a possible advantage as well. 


Also, models could be pre-trained on large synthetic datasets before fine-tuning on smaller real-world datasets, potentially improving performance in data-limited domains. Likewise, synthetic data could be generated to support training for languages with limited real-world data available.


On the other hand, there are potential downsides. When AI systems learn from each other, there's a risk of amplifying existing biases present in the original training data. As models build upon each other's outputs, subtle biases can become more pronounced over time.


With AI systems learning from each other, there's a danger of converging on similar outputs and losing diversity of perspectives.


Of course, it might not always be the case that synthetic data accurately represents real-world scenarios. The same danger exists in terms of models learning incorrect information from other models.


But there are many use cases where synthetic data is necessary or useful, including domain-specific models or privacy-sensitive models.


Why Marginal Cost of Content Creation is Generative AI's Superpower

Virtually every observer might agree that artificial intelligence will automate laborious tasks and therefore increase process efficiency. A...