The Impact of AI, Part I

Below is our first installment in a multi-part series detailing our perspectives on the artificial intelligence (AI) boom. We are personal and professional enthusiasts of AI, especially as it impacts the sports and asset management industries. Recent AI innovations likely represent a new computing paradigm that will change how we engage with our devices; how brands engage with consumers; how workflows are managed; how businesses are staffed; and where economic value accrues over the next few decades.

In Part I, we first sort through what’s real and not about the current AI boom. Next, we review how we think live entertainment is positioned in an economy dealing with AI-driven tech deflation. Finally, we explain what this all means for sports and media investors in what we call the ‘Global Content Glut’.

It is worth reviewing what is genuinely new about the present moment for AI research to isolate exactly why it is exciting, using as little speculation about future developments as possible.

The catalyst behind the substantial increase in attention on AI was the November 2022 release of ChatGPT by OpenAI. ChatGPT, and its cousins from competing providers, are the most human-like chatbots produced at scale with reasonable levels of compute and capital investment. Underlying ChatGPT is a “large language model” (or LLM) that allows it to produce plausible responses to general user queries. Several ingredients required to produce ChatGPT have been in development since the 2000s; however, the main breakthrough was the introduction of transformer models (the ‘T’ in GPT) in 2017. Combined with reductions in compute costs and the increasing power of AI-specific chips, transformer models like those underlying ChatGPT can be calibrating using significantly less and significantly messier data than in prior AI paradigms. Our favorite illustration of how these tools have captured imaginations is Fig. 1, which shows test scores achieved by GPT-4 (ChatGPT’s latest LLM) across a range of academic and professional disciplines. Applications like ChatGPT fall under the header “Generative AI”.

Our favorite description of how an LLM “thinks” is to imagine it is a box with (literally) billions of small knobs. Into the box, you feed many examples of human-generated text—effectively, the entire internet. A GPT model is just the correct, final calibration of those billions of knobs to their most effective positions for replicating human-like responses to the next input. Given the complexity of human language, you need >billion knobs (or ‘parameters’). The highest parameter count confirmed today is 540 billion in Google’s PaLM and Minerva. GPT-5 should have 2-5 trillion parameters. Proliferation of parameters has been enabled by greater computing power and the proliferation of available public and private data through the internet and connected devices. Prior to the advent of GPU-based training in 2010, training compute doubled roughly every two years, in line with Moore’s Law. Since the Deep Learning era, doubling time has been greater than Moore: <10 months (Fig. 2).

However, these models do have limits. First, these are not human agents. Unlike humans, they do not have internal representations of their external environment, do not have private internal-only thoughts or ruminations, and cannot make correct evaluations or predictions—e.g., of moral correctness, or appropriateness, or truth/falsity—except by luck, i.e., they do not build theories of how things should work, test those theories against intuition or evidence, etc. A tongue-in-cheek, but not inaccurate, description of LLMs is that they are highly sophisticated bullshit generators, in that they do not know or care if they produce truth or falsehood. These models just optimize for plausibility.^[1]

But it turns out a bullshit generator that can score a 1410 on its SATs, has memorized the entire internet, and doesn’t need a wage, food or sleep can still be extremely useful for business and scientific applications. While this current AI paradigm may not fully capture all nuances of human creativity, strategic thinking, prediction, or normative judgment, it is likely to augment or automate much of knowledge-based work and exhibits a limited form of creativity, so much so that some researchers believe this technology is a potential path to fully human reasoning capability. To demonstrate this, we asked GPT-4 to invent a sport that would appeal to Gen Z; its answer was an interesting form of tech-enabled Quidditch called DroneBall (Sidebar). Perhaps too cute, but hard to argue this isn’t creative.

In our view, Generative AI represents an early form of non-human intelligence—operating by different principles but with capabilities that mirror and often surpass a reasonably intelligent adult at tasks like classification, summarization, document production (‘substantiation’), search, and synthesis (i.e., inventing new concepts that combine components from existing concepts, like DroneBall).

The other major limitation at present is expense. There are two sources of incremental cost: the training and inference steps. Training requires state-of-the-art, expensive AI-specific chips, scarce scientific talent, and time. Researchers estimate that GPT-3 – now two years old – took approximately 34 days to train on 1,024 NVIDIA A100 GPU chips. Using AWS prices, this would cost over $4 million.^[2] A smaller model, Meta’s LLaMA, used 2,048 GPUs and took 21 days. If an AI company retrains their LLM every quarter, using the most sophisticated models as the benchmark, it would cost ~$40 million per year.^[3] Besides the fixed cost of chips, time, and talent, the inference step—e.g., a query to ChatGPT to plan your European vacation—is also expensive for companies utilizing fully training models. The inference step is currently about 1,000 times more compute-intensive than a Google search. Compute costs are growing so rapidly that they have become U.S. government national security concern. The good news is that competition for new chips is coming—a market leader with ~75% market share (NVIDIA) in a market growing at a 20%+ CAGR implicitly has a target on its back. But this will take time and further innovation. Internet applications took about a decade to convert into fully scaled businesses; now, after two decades, several of these companies have grown to economic dominance. AI applications could scale faster than internet applications did, but we can’t be sure.

Here are several applications of the current suite of GenAI models that are set to arrive from either Big Tech or other application software businesses – these will likely begin to be released over the coming months with consumer adoption taking two+ years in our view (Table 1).

We will spend time on how GenAI should specifically impact the sports business. To preview, we believe there are major efficiency gains possible for sales and service, digital media programming and marketing, and back-office functions. But we want to start with the implications for the macroeconomy and the media ecosystem in which sports live. Specifically, we believe that GenAI applications to content production should result in accelerating fragmentation—considerably more than experienced since the streaming era. Fragmentation in the AI-era will favor premium, legacy brands with authenticity anchored in physical experiences that cannot be easily replicated. As such, GenAI should have implications for the next era of value creation in sports—shifting from traditional telecast rights as the main driver to a robust content bundle anchored by the live experience that must aim to be an oasis of authenticity and meaning in a world of abundant AI-generated content we call the ‘Global Content Glut’.

Prior to 1870, economic value creation was predominantly extractive, overall productivity growth was low, and innovation was concentrated in the slow mechanization of agriculture. After 1870 until today—what economist Brad DeLong calls the “Long 20th Century”—there was a long process of globalization, urbanization, and mass mechanization of production, powered by the steam engine, fossil fuels, and capital investment from large private firms. During this period, the “labor intensity” of agriculture—the amount of worker hours you need per unit of agricultural output—had already declined. By the start of the Long 20th Century, labor had begun to shift to its next most productive use, i.e., where labor intensity remained high: the manufacturing sector. Urban manufacturing drove both investment and employment from 1870 until the mid-20th century, when the cumulative impact of decades of manufacturing automation and further globalization made mass production significantly cheaper. Now, both investment and employment are congregated in services sectors with lingering labor intensity, especially business services (think software developers and lawyers), healthcare, and education.

We believe that GenAI will be the catalyst for this last source of lingering labor intensity to lessen, specifically among knowledge work professions.

One interesting lesson of the Long 20th Century concerns the complicated effects of increased efficiency. As described in an intro economics textbook, more efficiency means satisfying more consumer needs than we were previously able to afford. Sounds great. But for the investor or business owner, economy-wide efficiency gains can be painful. Businesses earn profits due to some durable advantage or moat that cannot be easily competed away. For example, a profitable laundromat may generate value through a favorable lease relative to its location, due to few competing options in a densely populated neighborhood. But new technologies that are rapidly deflationary—i.e., that generate new avenues for competition at lower costs for consumers—result in capital loss, as prior investments behind existing moats are rendered moot due to new workarounds enabled by technology. Imagine online-only “virtual” laundromats that offer pick-up and drop-off and subscription-based payments; this would be a form of tech-enabled competition that would threaten our laundromat’s moat. This is what happened to the small proprietor in the wake of Amazon and Walmart, both creatures of internet- and computing-enabled innovation.

The interesting question for sports owners is: if Walmart is so threatening to the small proprietor, how did some small proprietors survive? We do not all just shop at Walmart and Amazon.^[4] Which parts of the economy have been immune to prior waves of automation or losses in competitive position, and which parts are likely to remain immune to this phase of AI-driven automation?

The above story, though over-simplified in several respects, frames how the economy has slowly chipped away at inefficiency. We’ve discussed two persistent veins of inefficiency, at a high level: labor intensity and capital immobility. Labor intensity is the amount of labor required to produce a unit of some good. This is improved by technology. Capital immobility refers to constraints—usually imposed by governments or geography—that keep capital from being deployed where it is most productive. This is improved by globalization.

Starting in the 18th century, we saw agricultural automation bring workers off the farm and into cities. This drove down the real price of agricultural commodities—our first ‘tech deflation’^[5] (Fig. 3). During the Long 20th Century, we saw manufacturing automation and globalization reduce the value of both factory labor and manufactured goods—our second tech deflation (Fig. 4). During the 21st century, we expect increasing deflationary pressure on sectors previously immune to these effects, driven by the twin effects of (i) Generative AI driving augmenting and eventually automating segments of the knowledge workforce; (ii) knowledge worker “globalization” due to work-from-anywhere and increasing connectivity. We are in early days, and these trends take multiple decades to play out, though we would flag that the speed of AI innovation has surprised some experts.

To get a sense of which parts of the economy have been immune to tech deflation, see Fig. 5, which includes the manufactured goods categories that deflated (vehicles, clothing, furniture, etc.) from Fig. 4, as well as a handful of services and experiences sectors that did not. These include—in order of least to most inflationary—gambling (3.4% inflation, or 0.3% real price growth), housing (0.5%), financial services (0.6%), experiences and live entertainment (0.7%), car repair (0.8%), insurance (1.4%), healthcare (1.7%), professional services (1.8%), and education (2.3%).

The Baumol effect—named after economist William J. Baumol—explains why this has occurred: all these sectors enjoyed intrinsic properties or circumstantial economic advantages that rendered them immune from the combination of productivity growth and competition that would otherwise have driven a tech deflation. The Baumol effect states that, as productivity grows in the economy overall, the price of everything that cannot be produced more efficiently gets bid up.^[6]

Classical music performances are intrinsically prone to Baumol inflation, due to the need for difficult-to-train human performers; pieces that cannot be altered dramatically; the appointment-viewing nature of the programming; etc. In contrast, professional services, healthcare, education, and financial services have experienced Baumol inflation only circumstantially. With GenAI, there is increased risk of a third tech-driven deflation in these sectors. To make this tangible, in each of these sectors there is typically a friction introduced by the need for professionals with costly, specialized training in a primarily knowledge-based field: think lawyers, doctors, accountants, sales professionals in some industries, some consultants, etc. GenAI can already augment—and in some cases imperfectly replicate—these professional services roles. Over the next five years, the increasingly cheap / ubiquitous compute available to train these models should make GPT-5+ even better at knowledge-based exams. Businesses focused on these sectors must be focused on retooling their tech stack and work force over the next decade to protect margins and remain competitive.^[7]

However, some sectors experience Baumol inflation structurally. We have seen how live music performances—especially at the highest, most premium end—count as examples. Similarly, any firm in the business of furnishing a product or service that is unique, authentic, original, or one-of-a-kind with difficult to replicate inputs (a “Unique Real Asset”, as described by our Insights piece on the topic ^[8]) or in the business of staging experiences—as opposed to simply providing services—should be more protected.

However, we do not believe that this feature means all the media & entertainment business will remain intact. In fact, our view is the opposite: some sub-sectors of the media & entertainment business are especially prone to disruption from GenAI. However, we believe that sports—and the premium end of live entertainment generally—is uniquely immune from GenAI tech deflation. In our final section, we apply the above framework to media & entertainment broadly to come to a thesis about how the sector will evolve post-AI.

We think we are entering the late innings for the legacy media model, and the winners will be the largest distributors (Big Tech), premium content, and premium live experiences. All other content niches—including some emerging sports leagues—will face increasing pressure to differentiate.

Here is our thesis for sports, step-by-step:

Besides sports leagues, another clear winner of AI-driven tech deflation and the Global Content Glut is Big Tech—possibly with a few new members, depending on how the ‘AI wars’ shake out. This follows from point #2 almost immediately: Big Tech represents the largest and most sophisticated ad platforms that will benefit from ad inflation. They could use the windfall to acquire legacy media companies, but this is unlikely for regulatory reasons. Instead, we think several media companies will have to bow out of the streaming wars over the next decade, as Big Tech becomes even more enrichened.

This thesis primarily concerns sports as a media property in a world of accelerating AI-driven content innovation. GenAI is not a predictive tool; it is fundamentally a tool for creation. GenAI is a universal compiler for content—both across mediums (text-to-image, image-to-text, text-to-video, etc.) and across languages. However, no matter how great, this compiler cannot translate authentic, physical experiences into a digital substitute of similar value. Hence, we believe the in-venue, live entertainment part of the business represents an underappreciated protective value layer for sports leagues and franchises.

In addition, athletic performance has qualities shared by art or music, namely, the fact that the product itself is human generated. (Performance-enhancing drug regulation is so important because “artificially altered” performance, even slightly, undercuts a basic assumption underlying why customers value the product.) Replicating the product of elite sports leagues more cheaply would require efficiencies in the production pipeline of great (human!) athletes above what the pro leagues can achieve today, which is already considerable. GenAI does not touch this.

Finally, on the media and sponsorship side, we would note that the Global Content Glut will also be a Global Novelty Glut. As GenAI becomes cheaper to use, we think novelty, hyper-stimulation, and hyper-personalization will also get cheaper and eventually commoditized. What happens to media consumer habits in this world? We speculate that viewership and attention could shift to content that is not novel, i.e., content that is analogous to “mom’s home cooking”. However, even if that is wrong, our thesis stands that GenAI will drive ad inflation and fragmentation, and so long as sports continues to retain audience, it is a natural winner from GenAI. This has implications for leagues, namely that retaining audience and relevance to the widest pool of consumers will become an increasingly ‘hard problem’ to solve with the highest long-term reward.

Generative AI (GenAI) is a genuine innovation. While it will likely progress to full consumer adoption more slowly than many think and will start with human augmentation before it moves to full automation, there will be real economic effects, especially for businesses that provide repeatable, knowledge-based services produced and delivered primarily by trained, human specialists. These specialists serve to intelligently process and communicate the contents of what are, in effect, large unstructured datasets—e.g., code samples or best practices, legal term sheets, standard salesforce efficiency or go-to-market playbooks, audit reporting and regulatory requirements, optimal responses to a concerned customer looking for a refund, and even optimal responses to a mental health patient in a psychotherapy setting. GenAI does impact these activities, and standard economics would suggest that over time we will see deflation in these AI-sensitive sectors, offset by Baumol-driven inflation in whatever sectors remains immune to GenAI. Sports, live entertainment, and other premium consumption niches are intrinsically Baumol-robust sectors as live entertainment properties with difficult-to-replicate inputs. Furthermore, the content business is tip-of-the-spear for GenAI impact, but it only will serve to commoditize attention further and drive resources to the Big Tech platforms and whatever content properties can retain authenticity and support ads. As the locus of economic value shifts from delivering services to staging experiences, we are excited to own premium sports assets at scale.

[1] This point underlies much of the concern with this technology as it pertains to disinformation risk.

[2] Deepak Narayaran et al, “Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM,” pre-print, August 2021, retrieved at: https://doi.org/10.48550/arXiv.2104.04473. See also Gregor von Dulong, “GPT-4 Will be 500x Smarter Than People Think,” Medium: https://medium.com/codex/gpt-4-will-be-500x-smaller-than-people-think-here-is-why-3556816f8ff2.

[3] Clement Delangue, CEO of Hugging Face, stated to CNBC in March 2023: “We are actually doing a training right now for the version two of Bloom and it’s gonna cost no more than $10 million to retrain.”

[4] Amazon captured about 4% of U.S. retail sales in 2021. While this is impressive share for a single company, Amazon looms much larger in the imagination than it does in the economy! Source: Amazon financial reports; U.S. Census Bureau Annual Retail Trade Survey; Statistics Canada.

[5] We use the term “tech deflation” to refer to goods and services deflation driven by the twin forces—i.e., both technology and globalization, the latter of which is also enabled by technology. As we’ll see with 21st century trends, globalization will be increasingly tech driven.

[6] The easiest model to understand this is a simple two-sector model in a barter economy. There is a goods sector that produces widgets more productively every year. Widgets can be exchanged for a ticket to a live performance of Beethoven’s Piano Sonata No. 12. For constant supply of capital and labor in the goods sector, productivity growth means more widgets every year—that’s just the definition (producing more with the same ingredients, e.g., through technological ingenuity). However, Beethoven’s Piano Sonata cannot be produced more efficiently. We can stage more performances, but it is just hard to produce world-class musicians that people are willing to pay to see (i.e., humans); and Beethoven’s Piano Sonatas cannot be sped up or altered. With the number of widgets available to use on live performances structurally going up, for a given level of unit demand for performances, the price of each performance must go up. This is the Baumol effect. Above-inflation price growth in inefficient sectors is a natural result of productivity growth elsewhere in the economy. For a detailed review of this phenomenon, see Eric Helland & Alex Tabarrok, “Why Are the Prices So Damn High?: Health, Education, and the Baumol Effect,” Mercatus Center at George Washington (2019), and William J. Baumol, “Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis,” The American Economic Review, Vol. 57, No. 3 (June 1967), pp. 415-426.

[7] The first law firm that decides to build a GenAI model that (a) does much of the low-level work of junior staffers, like document summarization and comment implementation, or (b) uses their proprietary database of (say) private equity-backed LBO credit agreements to automate first draft “market-standard” term sheets to facilitate deals would likely be able to run leaner, charge clients significantly less, and capture market share fast.

[8] Arctos Insights, “A New Look at Resilience: Sports Assets & Inflation,” July 2022.

[9] This is a play on Ben Bernanke’s famed “Global Savings Glut” concept to explain persistently low long-term interest rates. See Bernanke (2005): https://www.federalreserve.gov/boarddocs/speeches/2005/200503102/.

©Arctos Partners, LP, 2024. All rights reserved.

This material is provided by Arctos Partners, LP solely for informational purposes and is provided as of the date indicated above. Arctos Partners is not providing or undertaking to provide any financial, economic, legal, accounting, tax or other advice or recommendation in or by virtue of this material. The information, statements, comments, views, and opinions provided in this material are general in nature and (i) are not intended to be and should not be construed as the provision of investment advice by Arctos Partners, (ii) do not constitute and should not be construed as an offer to buy or sell any securities or to make or consider any investment or course of action, and (iii) may not be current. Arctos Sports Partners does not make any representation or warranty as to the accuracy or completeness of any of the information, statements, comments, views or opinions contained in this material, and any liability therefor (including in respect of direct, indirect or consequential loss or damage of any kind whatsoever) is expressly disclaimed. Arctos Sports Partners does not undertake any obligation whatsoever to provide any form of update, amendment, change or correction to any of the information, statements, comments, views, or opinions set forth in this material.