Who captures the most value in the AI explosion?

Applications vs. Knowledge vs. Models vs. Infrastructure

McKinsey’s report on artificial intelligence (https://www.mckinsey.com/industries/public-and-social-sector/our-insights/the-potential-value-of-ai-and-how-governments-could-look-to-capture-it), “The potential value of AI—and how governments could look to capture it,” makes an oft-cited bold claim about the expected returns to the global economy generated by artificial intelligence over the decade. The writers claim that by 2030, we can expect an overall increase in productivity by $13 trillion, which we can attribute to artificial intelligence and its transformative impact on industries, consumption, governments, and unit productivity.

Let’s assume that this estimate is directionally true, and allow for an order of magnitude of error in the estimate. The contemporaneous trends in venture capital seem to indicate that investors believe that this estimate (with the aforementioned caveats) is highly plausible. With Open AI, Character AI, Anthropic, Jasper AI, and other artificial intelligence ventures recently receiving massive valuations and venture capital investments, capital managers, entrepreneurs, and existing enterprises are levying large amounts of resources towards actualizing this future-state of global productivity.

However, among the general population of accredited and retail investors, I believe that there is some confusion as to who and which organizations will be able to capture and deliver this $13 trillion of added economic value to the global economy. This confusion has to do with distinguishing the different parts of the artificial intelligence value chain; following from the misunderstanding around the AI value chain, investors are then unable to accurately predict which companies will be the most successful at capturing the returns to the productivity gains.

When a company claims to be in the artificial intelligence space, what exactly do they mean by this statement? In the aforementioned McKinsey report, the analysts outline fifteen distinct domains of use cases for artificial intelligence, which range in scope from personalization insights and digital marketing to fraud & debt analytics and analytics-driven accounting & IT. However, these are simply applications of artificial intelligence – implementations of predictive technology that can make existing job functions more productive. While the productivity gain from these applications would be transformative, they are only a piece of the AI value chain.

So to answer the titular question of this article, let’s examine all parts of the AI value chain.

To build an artificial intelligence solution, we need (1) hardware to reliably and reasonably run artificial intelligence operations, (2) an artificial intelligence model – a language, image recognition, audio recognition, etc. model, (3) a place to host, deploy, secure, and manage this model, (4) a way to train, monitor, and improve this model to its specific domain, and (5) a workflow built around this model to make it useful. This fifth piece is where the McKinsey report spends most of its analysis, although the prior three points are equally ancillary to the expected value of AI in the global economy. Without dedicated models, hardware, hosting, or training services, there is no artificial intelligence, and certainly no applications of artificial intelligence in the world.

With this in mind, we can simplify the world of the artificial intelligence value chain into five pieces:

1. Hardware Infrastructure - the specialized hardware utilized to run artificial intelligence models efficiently. Example players in this segment: Nvidia

2. Cloud Infrastructure – the hosting, management, security, and deployment of artificial intelligence models. Example software in this segment: Amazon Web Services, AWS Sagemaker

3. Models – the underlying software that performs a predictive task (the base artificial intelligence piece). Example software in this segment: OpenAI’s GPT, Google’s BERT

4. Knowledge – the labeling, fine-tuning, monitoring, and validation of models that allow artificial intelligence models to accurately reflect a human’s intelligence for a given domain. Example software in this segment: Scale AI, Snorkel AI

5. Applications – the implementation of artificial intelligence model(s) in a larger workflow; the UI/UX, functionalities, and capabilities that make AI useful. The fifteen domains mentioned in the McKinsey report are, in my view, “Applications.” Example software in this segment: Jasper AI, Primer AI’s Command

This is a simplified model, and doesn’t take into account non-software or non-business related pieces of artificial intelligence that undoubtedly contribute to the value chain: governance, public policy, ethics, etc. are all critical components of ensuring the use, maintenance, and deployment of artificial intelligence is beneficial, equitable, and sustainable for the global community. While necessary to delivering the full productive value of AI to the global economy, they are non-technology components However, for the purposes of evaluating the private-sector investment potential of artificial intelligence, we can classify AI products & solutions with one or more of the above tags.

With this model in mind, when investors talk about companies and products in the artificial intelligence space, there are four different classes of solutions that to which we could potentially be referring; this isn’t to say that a company or a product couldn’t span multiple parts of the value chain, and businesses in the artificial intelligence space tend to do so. However, general discourse around AI tends to lump all solutions into the same category of software, and thus ignore the idiosyncrasies between solutions that allow us to make artificial intelligence actually useful to individuals and organizations.

So with different contributing factors to the AI value chain, the question of who captures the most value from supporting the AI-enabled productivity transformation remains unanswered; it would appear to be even more complicated, now that there are three other classes of players in this space.

Strong cases can be made for any tranche to be the dominant field of value capture in SaaS:

-       Companies focused on building AI applications (e.g., the ones McKinsey outlined) have direct customer relationships outside of the technology industry, and are most closely tied to real world productivity…

-       …however, for any of these AI applications to work, they need knowledge/labeling capabilities in order to make sure their models are performing with a level of accuracy that could justify replacing humans with machines…

-       …but regardless of the utilities for fine-tuning these models, the research organizations that built the original models can charge premiums for their usage, even if all they did was develop base models…

-       …however, the hosting services for these models collect premiums every time anything to do with a model is used, which would mean the infrastructure service providers would be billing regardless of the use case, model performance, or model itself

-       …yet, ultimately artificial intelligence operations rely on having computer processing that can reasonably return results and perform inference for an end-user; this efficient processing is enabled by GPUs, which allow for parallel computations

This complicates the artificial intelligence value chain substantially, since companies across the chain provide equally critical components to the artificial intelligence-enabled transformation.

From my perspective, I don’t believe a single tranche of the value chain is more predisposed to capturing more of the value chain than any other. Infrastructure companies don’t have an inherent competitive advantage over models, just because they host artificial intelligence models, and knowledge companies don’t have an inherent competitive advantage of applications, just because applications require domain-specific fine-tuning of base models in order to be accurate. Companies such as OpenAI don’t have an inherent advantage over labeling companies such as Scale or Snorkel because, as models become more sophisticated, their use cases and applications (i.e., the way in which they replicate human intelligence), will become increasingly complex and require different kinds of human input to learn from; as such, labeling may change shape, but the knowledge applications will still be critical in delivering value from artificial intelligence.

However, the artificial intelligence companies that will be the most successful are the ones that can successfully diagonalize across the tranches of the value chain. Diagonalizing in this case refers to two vectors of growth: (1) a company’s ability to incrementally, vertically assume different parts of the value chain internally, and therefore reduce the cost of services paid out to other parts of the value chain; (2) a company’s ability to incrementally, horizontally box out the competition within each tranche they can claim to operate in. If we consider the value chain to be “vertical” (infrastructure to applications) and if we consider the space of companies within each tranche to be “horizontal” (the companies within each of the four categories), diagonalization in this case would be the judicious growth of product offerings and internal tools & infrastructure that allow a company to oscillate between growing the capture of “horizontal” value of their tranche and investing in capabilities to reduce dues paid out to their other counterparts along the “vertical” value chain axis.

While a company may start out in one tranche, the more that they can provide evidence of reducing their long-term costs to be “owed” to the other tranches while simultaneously growing within their tranche itself, the likelier they are to secure a larger share of the long-term productivity gains that come from artificial intelligence. To concretize this, consider the following example:

A company begins by building an artificial intelligence application for a single industry use case; as they expand that product’s roadmap into adjacent solution spaces, the company simultaneously invests in building an internal labeling and monitoring tool as to reduce its dependency on external services; eventually, the company then converts these internal tools into another product, and parlays the revenue from that product into investment in its own model hosting and management infrastructure to avoid dependency on a cloud provider

While we are making some assumptions about the company’s internal management and product organization (that they are competent at execution), since this hypothetical is only concerned about the upper-tail of the returns on artificial intelligence’s value, we can constrain the analysis to reflecting on how a diagonalized company captures more value relative to a pure horizontal or vertical company. A purely horizontally-growing company, one that stays within its tranche, will always owe costs to businesses in the other tranches, due to the nature of artificial intelligence – as such, the other tranches will continue to collect premiums from this business. A purely vertically-growing company, one that invests upfront in building all of its AI capabilities in-house, is incredibly risky as the costs required to stand-up infrastructure, an artificial intelligence research team, and labeling and monitoring services & utilities, all to support a single use case/application (as the company is not growing beyond a single product case, since it’s vertical), are likely to be disproportionate relative to the business model’s obtainable market; as such, a company employing a purely vertical strategy to eliminate long-term costs owed to other tranches is a highly risky venture.

As such, it’s my belief that artificial intelligence technologies that incrementally invest in internal capabilities that reduce external dependencies alongside growth opportunities for their business model will be able to capture more value than companies that hyper-focus on one tranche of the value chain (or on one specific use case). Companies that diagonalize effectively are not the only companies that will succeed (i.e., generate a premium of the risk-free rate of return on invested capital to investors) in the artificial intelligence space; but they will be able to capture more of the long-term value that artificial intelligence intends to bring to the global economy, as their strategy revolves around simultaneously de-risking the return on invested capital while reducing long-term dependencies and premiums owed to other businesses in the artificial intelligence value chain.