In some key ways, the artificial intelligence business is developing in quite a different manner than did the internet, at least at the level of foundational models that might be likened to operating systems.
Or, to use another analogy, the large language model business stack is built on the actual LLMs. And there is a possible divergence between the internet, built largely on open source or marketplace standards including TCP/IP and Ethernet, and generative AI, built more on the model of the personal computer operating system platforms.
The internet's development--at the key application level--was largely driven by startups and entrepreneurs. GenAI is largely driven by a relatively small number of large and established firms, even if startups abound at the app layer.
Building internet apps and services often required less initial investment compared to GenAI, such as the cost to build and train an application’s generative AI capabilities and inferences.
Application user experience and scalability arguably were more important than access to capital, in part because capital was so abundant at that time and also because the ability to scale (users) was seen as key.
AI models are dependent on vast access to data; the internet apps were not. So as mechanisms develop to codify “fair use” and licensed access, more capital is going to be required for access to quality data sources.
There are other angles as well. The early internet was powered by private data centers of modest size. But AI is powered by “cloud computing” mechanisms.
By most estimates, about 65 percent of the capacity in global data centers is owned by just three companies: Amazon, Google, and Microsoft. That matters for artificial intelligence provided “as a service,” as much of the digital infrastructure required to support AI will be provided by the handful of hyperscale “computing as a service” suppliers.
And some might note that one value of investing in an AI startup are the agreements to use a particular cloud computing provider.
Startups get investment, but also agree to use the investing cloud computing giant’s infrastructure.
Also, some note that Google, Microsoft and Amazon are actively investing in hundreds of AI start-ups, as well. In 2023, Google, Microsoft and Amazon invested as much as two-thirds of the $27 billion for AI startups, a report argues.
Ignoring for the moment the matters of governance or competition in markets, there are possible systemic dangers related to firm revenue and profits. In the internet bubble at the turn of the century, for example, many firms exaggerated their revenues or capital bases using various forms of financial excess.
Internet capacity providers engaged in a practice called "capacity swapping." They bought and sold unused bandwidth from each other, artificially inflating their reported capacity and network reach. This created the illusion of high demand and fueled investor confidence. But it was an illusion. Actual end user demand was not as high as it seemed.
Many internet app startups relied heavily on vendor financing. Vendors would extend credit to these companies in exchange for stock options. This allowed startups to show reduced costs while vendors could report higher sales.
Some companies also resorted to creative accounting practices to inflate their revenue figures or provide growth metrics. Companies might record barter agreements, where they traded services or advertising space instead of receiving cash, as actual revenue. This inflated their top line without reflecting any real cash flow.
Some companies recognized revenue from multi-year contracts upfront, treating the entire value of the contract as income in the current year. This practice distorted their current financial health and overstated immediate profitability.
Companies might capitalize expenses related to marketing, website development or customer acquisition as assets instead of showing them as expenses. This artificially inflated reported profits.
Some companies recorded revenue for services even if the customer hadn't paid yet, again inflating reported revenues.
The point is that the AI business is developing in quite a different way than the internet. At least until the spigots shut off, there was plenty of investment capital available during the internet bubble. I recall being quite shocked when told by a startup’s founders not to worry about some parts of a business plan I was working on, as there was “plenty of money.”
AI is different, at the model or platform level. It is extremely capital intensive at a time when capital arguably is not plentiful or affordable. So barriers to entry are quite significant for model builders. In that sense, GenAI more nearly resembles the PC operating system model than the internet.
No comments:
Post a Comment