Costs of Creating Machine Learning Models is Up Sharply

With the caveat that we must be careful about making linear extrapolations into the future, training costs of state-of-the-art AI models have reached unprecedented levels, according to Stanford University’s Human-Centered Artificial Intelligence institute. 

Where OpenAI’s GPT-4 used an estimated $78 million worth of compute to train, Google’s Gemini Ultra cost $191 million for compute, HAI estimates. 

source: Stanford University Human-Centered AI report

HAI estimates suggest that model training costs have significantly increased. For example, in 2017,

the original Transformer model, which introduced the architecture that underpins virtually every modern

LLM, cost around $900 to train.

RoBERTa Large, released in 2019, which achieved state-of-the-art results on many canonical comprehension benchmarks like SQuAD and GLUE, cost around $160,000 to train.

Fast-forward to 2023, and training costs for OpenAI’s GPT-4 and Google’s Gemini Ultra are estimated to be around $78 million and $191 million, respectively, according to HAI. 

 

source: Stanford University Human-Centered AI report

AI Might Not be in an Investment Bubble, but Generative AI Might Be

At least as measured by new private investment in artificial intelligence, the danger of over-investment might be receding, as AI investment levels peaked in 2021 and appear to be declining, according to Stanford University’s Human-Centered AI institute. 

source: Stanford University Human-Centered AI institute 

On the other hand, investments clearly have shifted to generative AI. While overall AI private investment decreased in 2023, funding for generative AI sharply increased. In 2023, the sector attracted $25.2 billion, nearly nine times the investment of 2022 and about 30 times the amount from 2019. 

Generative AI accounted for over a quarter of all AI-related private investment in 2023.

source: Stanford University Human-Centered AI institute 

The most commonly adopted AI use case by function among surveyed businesses in 2023 was contact-center automation (26 percent), followed by personalization (23 percent), customer acquisition (22 percent), and AI-based enhancements of products (22 percent), according to research by McKinsey. 

Where Will AI Prove an Existential Threat to Whole Industries?

Right now, we all speculate about the potential changes artificial intelligence might bring, as well. Predictions range from the existential (CEO jobs will go away) to the mundane (how we write emails will be more automated).

But it is not the technology we need to watch and understand, as such. It is what the technology enables. We might agree that spreadsheets, word processing, presentation software, personal computers or smartphones have had an impact on work and learning, without also arguing those information technologies fundamentally changed the nature of work or most business models.

But sometimes, new IT has reshape whole industries.

If you worked in any managerial position in U.S. ad-supported media back in 1996, you are well aware of the huge shifts that has taken place in use of ad venues. Print media, linear video and radio have taken huge hits, while online digital venues have skyrocketed. 

Many--perhaps most--of the business issues facing managers of ad-supported assets flow directly from those shifts in activity and venues. 

Source: Gemini

Put simply, digital now claims up to 82 percent of all U.S. ad placements and revenue. Print has declined from 42 percent to less than three percent. Linear video dropped from 38 percent to 16 percent. Radio dipped from 10 percent to half a percent. 

Channel	1996 (Billions)	1996 (%)	2023 (Billions)	2023 (%)
Print (Newspapers, Magazines)	80.0	42.1%	10.0	2.7%
Linear Video (TV Broadcast, Cable)	72.0	37.9%	60.0	16.2%
Network Radio	10.0	5.3%	2.0	0.5%
Other (Radio Spots, Out-of-Home)	28.0	14.7%	18.0	4.9%
Digital Ads (Search, Social Media, Display)	-	-	300.0	81.7%

When channels shift that much, the appellation "disruptive" certainly applies. While the role of advertising in business did not change, almost everything else about how it is used, when and where, did change. 

We can't yet know how extensive the changes brought by AI will be, in each industry, job role and functions. But existential change is conceivable. Virtually all prior general-purpose technologies (such as electricity, the internet, mass propduction, the internal combusion engine) were existential for some industries and jobs. Some things simply went away. 

And virtually all enabled importnat new possibilities based on the extending of human muscle, senses or brain power. 

The internet had an existential impact on legacy ad-suported businesses. AI could have similar impact on at least some industries. 

AI Mice and Keyboards: Tension Between Curation and Openness Remains

Microsoft’s dedicated AI key on some keyboards--which opens up access to Microsoft’s Copilot--now is joined by Logitech’s Signature AI mouse, with a button to open up the Logi AI Prompt Builder software and ChatGPT. 

Both might be viewed as equivalent to shortcut keys that simplify access to a chosen AI engine or feature on a device, and likely will be pitched as an easier way for some users to use a specific AI engine, though in principle more-evolved versions by these or other suppliers might offer access to a user’s chosen engine or engines, or offer context-aware AI functions. 

The issue there is the eternal balance between the values of curation (walled gardens that simplify or unify experience, provide greater security and consistency of experience) and the values of openness (flexibility, power, choice at the cost of complexity, risk and security). 

Some might view such interfaces as gimmicks of a sort, and they also represent walled garden approaches to use of generative AI. 

Proponents might argue such buttons or keys provide value by making it easier for users to use one AI engine.  

Instead of navigating menus or opening separate applications, a single click on the AI button brings up Logitech's AI Prompt Builder software, for example. Users can write prompts, customize the desired response tone and complexity, and receive results directly within the Prompt Builder window.

Using the AI button, users also might be able to quickly access ChatGPT's “summarize” function for documents or emails, as well.

AI keys or buttons might be useful for beginners or those unfamiliar with navigating menus, some might argue.

Skeptics might argue this is reminiscent of the early days of the multimedia web, when AOL offered a walled garden internet experience deemed helpful for new users, but was less helpful--or limiting--for users who had some experience. The idea then was to simplify the user experience, and perhaps AI buttons and keys will do that for some new AI users. 

One might envision such keys or buttons launching built-in AI assistants enabling voice commands or  dictation.  In some cases functions might become context aware.

Ideally, users might be able to program the buttons or keys to perform their preferred AI action, such as providing image editing suggestions or content creation prompts.

So AI buttons and keys are an experiment in making AI features more accessible and user-friendly.

In that sense, they resemble voice interfaces for smart speakers, and speech-to-text functions, which aim to make interacting with technology more natural and efficient. But, as with earlier efforts to simplify access, users might quickly outgrow the interfaces, opting instead for the more-flexible and powerful use of menus or other open-ended interfaces. 

Using the AOL analogy, users rather quickly outgrew the walled garden interface and opted instead to rely on open, general-purpose browsers. 

And some innovations simply do not catch on. Smart speakers have failed to become a dominant interface, though voice-to-text functions on smartphones are routinely used. 

It remains to be seen whether walled garden keys and buttons actually provide the intended value, and if so, for how long before users become more AI-proficient and outgrow the interfaces. 

The Graphical User Interface (GUI) and touchscreens, on the other hand, provide classic examples of successful interfaces. And there always are exceptions. The Apple iOS is a walled garden that works. Windows Phone is an example of a failed walled garden. 

In other words, there always is a tension between curation of experience and choice, customization, broad access to features and an open approach. 

Some of us might be so convinced AI keys or buttons will be enduring interfaces. 

The Next U.S. Recession Will Test Resilience of Video, Communications Businesses

Whenever the next U.S. recession happens, we will see whether the many changes in the telecom, cable TV and video streaming markets will change the historic view of how telecom and video entertainment stocks behave during downturns. 

Traditionally, both telecom and cable TV equities have been viewed as resistant to customer defections in recessions as both are “essential” or “important” recurring services. 

But the markets and consumer tastes have been evolving: reliance on mobile phone services and abandonment of fixed network services; substitution or addition of video streaming services and reduced linear video subscription buying; increased importance of internet access and a decrease in importance of voice and linear video services. 

source: Broadband Search, Seeking Alpha 

All of which raises new questions, including the issue of whether streaming services will prove more resistant to customer churn during recessions, compared to linear video. 

Study Title/Author	Findings
"Do Consumers Cut the Cord in a Recession?" by John Beggs and Patrick/2010	Found a slight decrease in cable TV subscriptions, but not a significant decline.
"Telecom Stocks and Economic Downturns" by JPMorgan Chase (Investment Report) /2020	Indicated telecom stocks generally outperform the broader market during downturns.
"The Recession Resilience of Defensive Sectors" by Fidelity Investments (Market Commentary) /2023	Listed telecom as a sector with potential resilience, but noted the importance of specific company financials.
The Recession and Telecom, Deloitte (2009)	Revenue for telecom service providers remained relatively stable during the 2008 recession, but capital expenditures declined.
The U.S. Telecommunications Industry During Economic Downturns, The Brattle Group (2010)	While telecom revenue growth may slow during recessions, it generally holds up better than the broader economy.
Cord Cutting: What Do Past Recessions Tell Us? MoffettNathanson (2020)	Previous recessions saw limited cord-cutting, suggesting cable TV might retain some stability during downturns. However, the study acknowledges the changing media landscape.
Fama & French (1989)	Defensive sectors like telecom and utilities tend to outperform cyclical sectors.
Ang & Timmermann (1993)	Telecom and utilities exhibit lower volatility and higher risk-adjusted returns during recessions.
Blitz & Reichlin (2001)	Telecom and utility stocks are less affected by credit downgrades compared to cyclical sectors.

A recession might accelerate the secular trend of fixed network voice service abandonment, as consumers prefer mobile phone service. Likewise, a recession might also accelerate linear video abandonment rates, considering the relative expense, compared to streaming alternatives. 

To be sure, live sports will be a key issue for a portion of the buying public. Though most observers see a continuing shift of live sports to streaming services, that trend is not as developed, yet. So sports fans might still conclude they have no choice but to keep their linear video subscriptions. 

And that should continue to prop up demand during recessionary periods. 

On the other hand, perhaps a majority of consumers who are not sports fans can buy multiple streaming subscriptions at lower (or near equivalent) prices than they can buy a linear subscription, suggesting the possibility that streaming services could prove more attractive during a recession. 

Also, streaming arguably still is a growth business, while linear video is in decline. Any recession might accelerate such trends. 

source: Ryan Ang, Seeking Alpha 

The most recent recession, caused by the imposition of Covid shutdowns on the economy, might not provide much insight. With the “in person” economy largely shut down in many countries, demand for work from home or learn from home internet access was quite high. 

Take rates and usage of mobility services arguably rose for the same reason. And the value of streaming and even linear TV services arguably was boosted by the lack of other entertainment options. 

So the most-recent major downturns for which we arguably have data would be the 2008 global financial crisis and the 2000 to 2001 dotcom crash, when video streaming was not a mainstream business at scale. 

AI "Do Something, Now" Advice Will Mostly Lead to Irrelevant Outcomes

Whenever an important new technology arrives, it gets hyped, and the bigger the possible transformation, the greater the hype. And that is likely the case for artificial intelligence as well. “Act now, or fall behind,” the argument will be. 

The issue is that a truly-important technology is misunderstood, and therefore will tend to be deployed, early on, in suboptimal ways. Think of the way the early multimedia web featured firms posting their brochures online. The new is viewed through the lens of the old. 

That tends to be the case for all general-purpose technologies, from electricity to the internet and now possibly artificial intelligence. The impulse to “do something” will be overwhelming for many firms in many industries, even when the outcomes will turn out to be more or less negligible, as important as they might seem for any particular use case. 

Yes, we will be able to do existing tasks faster, with less labor, and maybe even “better” in some ways. But the real value will come later as we learn to remake whole processes. 

The initial use case for the internal combustion engine was to pump water out of mines. But the greatest value of the ICE was its impact on transportation and mobility. Though street lighting was among the first obvious uses of electricity, its greatest impact now is as a platform for all sorts of appliances, sensors and portable and mobile devices, communications and computing. 

Consider a study of 792 banking sector decision makers surveyed by Forrester in 50 countries and 12,500 consumers surveyed by IPSOS in 14 countries suggests there is a big gap between banks and their customer perceptions about “value.” 

In fact, 46 percent of consumer respondents said they were open to ditching their current provider for a new company if it offered the personalized services they want. In an age of hyper-connectivity, hyper-personalization is needed, the report argues. 

Nearly four out of ten customers now have an online bank, which 36 percent consult with at least once a day, the report published by Sopra Steria says. 

Mobile applications and websites are becoming the primary channels of exchange for 58 percent of respondents, and only 25 percent of customers say that they contact their advisor as their first means of communication.

That is part of the general “be digital” advice executives are given in virtually every industry. But it always is the outcomes that matter, despite the level of activity. 

Consider the value reaped by cloud computing. 

“A stunning 95 percent of European companies in our recent survey say they’re capturing value from cloud, and more than one in three say they intend to have more than half of their workloads on cloud,” say McKinsey consultants Bernardo Betley, associate partner; Hana Dib, associate partner; Bjørnar Jensen,  senior partner and Bernhard Mühlreiter, partner. That’s the good news. 

The bad news? “The vast majority of the value companies have captured, for example, remains in isolated pockets and at subscale,” the consultants also say. 

Some of that might be caused by the way European companies (the subjects of the study) have implemented cloud computing. 

“The focus of their cloud efforts, for example, has been disproportionately on improvements to IT, which generate lower rates of value than improvements to business operations,” McKinsey says. 

Somewhat oddly, “most companies (71 percent) measure it (benefits) in IT operational improvements, 66 percent in IT cost savings, and 63 percent in number of applications on cloud,” McKinsey says. “Only about one in three European companies, however, monitors non-IT outcomes, such as cost savings outside IT (37 percent) or new revenue generation (32 percent).”

That might seem odd for many observers, since those outcomes (cost savings or revenue enhancement) might seem the obvious way to measure outcomes. “Our research and experience are clear that about two-thirds of the potential value of cloud comes from revenue uplift and cost savings in business operations,” McKinsey says. 

Study Title/Author	Methodology	Findings
"The Business Value of Cloud Computing" by McKinsey & Company (Report) / 2019	Surveyed over 1,500 executives globally	Found that companies with a clear cloud strategy focused on business outcomes achieved 3x the return on investment (ROI) compared to those with a technology-centric approach.
"Cloud ROI: Why Businesses Need a Strategic Approach" by Forrester Research (Report) / 2021	Analyzed data from cloud adoption projects	Concluded that companies with a well-defined cloud strategy focused on business goals like agility and innovation saw a 20% increase in revenue growth compared to those with a tactical, workload-centric approach.
"Unlocking the Economic Advantage of Cloud Computing" by Accenture (Report) / 2022	Examined the impact of cloud on various industries	Identified that companies using cloud to drive innovation and customer experience saw a 15% increase in customer satisfaction and a 10% improvement in operational efficiency.
"Cloud Adoption and Firm Performance: A Meta-Analysis" by Florian Schreibert et al. (Journal Article) /2020	Analyzed 42 prior academic studies on cloud adoption	Concluded that while workload shifts can lead to some performance improvements, the most significant benefits are seen when cloud adoption focuses on strategic goals like cost reduction, efficiency, and innovation.
"The Cloud Dividend: How Businesses Benefit from Cloud Computing" by Accenture (Report) /2022	Analyzed financial data from over 10,000 companies	Identified a strong correlation between effective cloud adoption strategies focused on business outcomes and improved financial performance.

“Compared to U.S. companies, about five times more European companies are still pursuing an IT-led cloud migration, with significant emphasis on lifting and shifting existing workloads,” the consultants say. 

The point is that lots of advice about “doing something now” and “moving faster” with artificial intelligence is not going to produce measurable and useful outcomes. It will amount to mere activity. That always seems to be the case for important new technologies. 

That noted, another set of outcomes should be mentioned. Eventually, as firms and industries learn to use the new GPTs effectively, clear outcomes will be obtained. That is the good news. 

The bad news is that the value of the innovations will accrue to most firms in each industry, so sustainable long-term advantage will fail to be gotten. Important new technologies will rise to the level of “table stakes,” essentially forcing firms to invest to keep up with their competitors, and not to gain an advantage over them. 

In the meantime, lots of capital, information technology effort and time will essentially be wasted applying AI where it actually does not produce outcomes that matter. 

As during the turn of the century “dotcom” mania, every firm tried to be an “internet-first” firm. And many of those efforts simply fizzled. “We’re on the internet” or “we use the internet” failed to produce clear outcomes for most firms in most industries. 

Most firms selling in retail stores tried to add “internet” capabilities, while many new firms tried to operate “online only” in retail. Few actually produced material changes in market leadership, as did Amazon. 

Only a few later (after decades had passed) mastered the sharing economy by harnessing smartphones (with internet capability) to the process of public transportation (think Uber or Lyft), creating ridesharing, or peer-to-peer short-term accommodations (think Airbnb). 

Eventually, banks will learn to use AI in ways that materially affect customer experience, as they did with automated teller machines or online banking. 

But many early efforts will fail, in part because our present understanding of how to use AI in ways that matter is incomplete. We’re still at the stage of posting brochures online. We have yet to master the ability to “do things” online because we mostly will try to automate existing processes. 

Later, we’ll figure out how to remaster processes. But that will take some time.