Is AI Eating the World?

I am generally optimistic about AI’s potential to impact the world. Undoubtedly, AI has been a significant driver of the trillions of dollars of market capitalization created by tech giants. AI also performs daily miracles for billions of consumers. I like looking at my phone to open it, and I love being able to look at a camera to enter a country in seconds.

For corporations too, AI can work wonders. For example, JPMC, currently the largest bank by market capitalization, is focused on more than 300 AI use cases to deliver $1 billion to $1.5 billion of impact. I have seen AI use cases that deliver 10x or even 100x return on investment in areas such as customer loyalty and automated customer care. Even my chatbot experiences have been getting better lately and are able to address more of what I need to get done. This all feels like progress.

On the debit side of AI’s balance sheet, it is fair to say we are in a part of the AI hype cycle, especially around generative AI. Although there are spectacular examples of AI impact and the average AI initiative ROI is improving, the typical AI initiative ROI at 5% to 6% is below the cost of capital for most organizations. Many data science projects, perhaps 80% to 90% of them, never make it to production, a stunning waste of scarce data scientist talent.

As laudable as the JPMC activities are, AI is only delivering a few percentage points of uplift on the bank’s bottom line. That is a typical uplift for non-tech companies, even those that are at the vanguard of applying AI to their business.

Opportunities to improve AI abound. However, AI has the potential to mimic human problem-solving capabilities beyond the scale and speed that humans can achieve alone. Given that human civilization has so many unsolved problems, I still want to shout, “bring it on.”

Finally, generative AI, like metaverse and blockchain before it, is probably on the hype part of the adoption cycle. That is not a surprise, as everything new tends to start overhyped.

Recently, getting my hands dirty testing open-source foundation models, I have been impressed with their ability to create usable solutions for tasks such as customer sentiment analysis within days or even hours with a small team, rather than the typical weeks and months with larger teams of data scientists and engineers. This impact still needs to be proven reliably at scale, but it does bode well for reducing the time and cost of AI initiatives and hopefully improving ROI.

Who is making money, and who might make money?

Three questions are top of mind:

1. Will AI “eat the world,” as Andreessen famously said software will, and did?
2. Which business models are attractive now and in the future?
3. What would my investment thesis be if I had real money to invest?

1. Will AI eat the world?

Value in the technology stack has generally shifted from infrastructure to applications.

That is for several reasons. Applications are what customers directly experience, and therefore the software layer enables a company to differentiate more. Software can be difficult to integrate and learn how to use. Once it is installed, especially for enterprise customers, it can be difficult to shift, and lock-in creates stickiness and pricing power.

Direct customer touchpoints also provide more timely and granular insights that drive better product differentiation, especially for software-as-a-service applications. Finally, the scale economies of software are phenomenal. Marginal cost of production is zero, so acquiring new customers enables strong margin expansion and impressive cash generation.

I am sure this is not an exhaustive list, but all of these are strong contributors to the phenomenal software flywheel that has powered many a stellar company valuation.

Does this still hold true? What about data and AI? Cloud providers? Integrated hardware and software players like Apple?

Some things have changed to challenge the supremacy of software.

For one, the cloud hyperscalers seem to have created a natural monopoly around their cloud services. They have also blended infrastructure and applications, so that when you buy a cloud service you are not just buying commodity storage and compute. You are also buying important composable application components that can be seamlessly woven into your application stack. They have leveraged software to make their hardware more powerful, and hardware to make their software more powerful.

Apple also has a successful integrated hardware and software proposition that gives it tight control over the user experience and ecosystem. It was once said that Apple captured more than 80% of the available smartphone profit pool from its successful integrated business model.

Data and AI are more complex. I deliberately separate data from AI because the distinction is important.

My intuition has always been that data is a better source of proprietary competitive advantage than an algorithm. Algorithms are usually only effectively created by having access to relevant data in the first place, and they can be easily copied. Data can be a natural monopoly too. Many billions of dollars of value have been created by aggregating data to create unique and valuable data sources. IMS Health, now IQVIA, is a great example of a company that succeeded by aggregating pharmacy prescription data and selling access and insights to the life sciences industry.

Can you name any companies that created a sustainable advantage from an algorithm alone?

Google maybe comes close. But perhaps Google’s sustainable innovation is more around the paid search business model that monetized its search advantage, and the data Google has on websites and other searchable information. Google’s advantage seems less rooted in an algorithm and more rooted in data and business model innovation.

At one level, AI is just a glorified algorithm. It can easily be copied and needs access to the right data to be effective. AI is also rarely consumed as a standalone product. Most AI is integrated into software or into the cloud offerings of hyperscalers. Enterprise AI also requires tight integration into enterprise data and applications, and regular updating based on the latest internal and external data trends. All of this suggests that AI has a certain level of software-like stickiness.

On the other hand, AI seems to require more adaptation to the enterprise than software. That means its marginal cost of production is higher, and AI products may be less valuable, less differentiated, and more likely to act as building blocks of internal proprietary solutions.

Take the current hot topic of large language models. OpenAI’s GPT models are the gold standard taking the world by storm. They also require huge data sets and compute resources to train, perhaps costing tens or hundreds of millions of dollars. That may be beyond the reach of most corporations.

The models are expensive to use, not just in API costs but also due to high inference compute intensity. Thankfully, there are many free open-source alternatives. Go to a site like Hugging Face and you will find hundreds of free models that can perform miracles for you.

Probably every sizable corporation will have its own LLM, trained or at least tuned on a curated set of relevant internal and external data. They will need to source and integrate multiple components to make this happen.

Will we see AI equivalents of large enterprise applications such as SAP, Oracle, and Salesforce perform this role? It is dangerous to make a prediction, but difficult to imagine that we will.

For a start, all those large enterprise application vendors are busy incorporating AI into their products to make them more attractive and to reduce the need, and threat, of a mega AI enterprise application. Secondly, the value of your own LLM is in the co-specialization outlined above. That may not be a product you can actually buy.

To answer my first question, AI will not eat the world in the same way as software.

AI will, hopefully, unlock trillions of dollars of value, as many pundits forecast. But its flywheel is not as powerful as software’s. It will be constrained by the supremacy of data. In fact, AI will make data even more valuable.

AI will also probably, and hopefully, be more democratized than software. That is a bold statement, but the economic forces at play seem to call for less concentration, assuming that data itself does not become the choke point that creates monopoly-like rent extraction.

Please stay tuned to my blog for answers to the other two questions.