Inflated expectations, inflated assets, inflated capital investments, bad policy. It’s all got to end pretty soon.
Artificial intelligence applications are everywhere. They unlock your phone, correct your spelling, redesign your slide presentations, recommend videos, scan and summarize reports and papers, converse with you. All these use cases are different, yet we continue to be misled by a vision of Big AI: one gigantic, controlling Artificial General Intelligence involving ever-larger concentrations of data, energy and compute.
The 7% decline in Oracle stock on October 7, due to the discovery that Oracle is losing money on rentals of small quantities of Nvidia’s chips, is a welcome market adjustment to the inflated expectations around artificial intelligence. But there is so much more to give us pause. OpenAI is projected to lose another $14 billion by 2026, with total losses from 2023 to 2028 potentially reaching $44 billion. Three of the biggest Big AI companies – Google, Anthropic and OpenAI – are seeing diminishing returns from their costly investments in newer, more powerful models.
If we look at the actual work AI applications are doing. we find a pretty mixed bag. Generative capabilities are thrust onto users incessantly by companies like Microsoft and Meta when they are neither wanted nor needed. The most useful AI applications currently are extensions of search capabilities; they are interfaces that accept natural language queries to find, compile, summarize and sometimes even make useful suggestions about the vast information and data resources humans have digitized and put on the internet over the last 30 years. Content recommendation algorithms and facial recognition technology are also producing useful results. Yet, we still don’t know how to detect and filter spam emails without a lot of false positives and negatives. The most intelligent application, the so-called frontier model LLM, can aid the step from search to usable output, but can also backfire.
An MIT NANDA study concluded that “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.” A Harvard Business Review article says “AI-generated ‘Workslop’ is destroying productivity.”
“As AI tools become more accessible, workers are increasingly able to quickly produce polished output: well-formatted slides, long, structured reports, seemingly articulate summaries of academic papers by non-experts and usable code. But while some employees are using this ability to polish good work, others use it to create content that is actually unhelpful, incomplete, or missing crucial context about the project at hand. The insidious effect of work-slop is that it shifts the burden of the work downstream, requiring the receiver to interpret, correct or redo the work.”
Universities are quite familiar with this pattern. Teachers have to invest considerable time and energy in detecting whether student outputs are genuine or AI work-slop.
Yet we persist in the belief that the chatbot is the killer application of AI. We don’t seem to realize that it is the business model of an online query system and thus will inevitably become an advertising business. It’s competing for our attention with hundreds of other online applications and if it succeeds in getting it, advertisements will follow, and with them, distortions in the results.
Our politicians – especially the populist-nationalists currently in control of the United States — would like us to believe that Big AI will lead to domestic control if not world domination. So the government and investors are pouring money into data centers and chatbots to make their datasets bigger and their computations faster. Despite the fact that AI evolved out of a globalized digital ecosystem and a transnational scientific and technical community, the US is trying to find ways to build fences around the components of AI production – chips, models, data – to exclude other nation-states. China, likewise, is trying to achieve self-sufficiency and correctly sees LLMs as a form of speech, and thus as something they need to control. We are effectively nationalizing these industries by making the government a shareholder/controller and implicating them with national security concerns.
The resultant fragmentation of the world market for digital services and infrastructure is going to be a huge blow to growth. Add to that the rise of tariff-based trade barriers in practically every other sector, and the bust will come. How quickly or slowly it comes depends on the impact of tariffs. I suspect it will be relatively gradual, grinding things down over a period of four or five years. But there will be a bust. whether it’s in slow motion or not remains to be seen. AI applications, like all commercial businesses, are going to have to meet a cost benefit test. Effective applications tailored to specific market needs will succeed, many others will not. Wild promises (or threats) that AI transforms everything will fall by the wayside as that happens. AI is just the latest phase of digital computing. AI applications will become just another software industry. Distributors, the retailers of those applications, will be just another platform – a Google, Amazon, Meta, Alibaba, Ten Cent, Huawei, Apple and hundreds of smaller others. Of course, many of the current distributors already are platforms. And so our tech policy mavens in Washington need to worry less about national security and our military power relative to China, and start worrying more about economic competitiveness and economic rationality.
If you look at the targets of this over-investment you get confirmation of my thesis that AI is just distributed computing, an evolving capability of a digital ecosystem, the essential components of which are digitized data, compute power, network connectivity, and software (not to mention electrical power, lots of it). Where is investment in AI going? It is going to data centers, to chip production and consumption, to cloud-based connectivity, to power sources as well as to AI application software developers. Often these components are integrated in cloud providers and platforms, which explains their leading role in Big AI.
It’s been 25 years since the dot com boom. Then, inflated dreams about the unlimited potential of the internet pushed the market to new highs and over-investment in bandwidth. The Internet was revolutionary and transformative, but it did not mean that pets.com was worth billions simply because it was on the Internet. The digital economy is currently fueled by a similar kind of over-investment in anything connected to AI.
Company PE ratio
- AMD 126
- Broadcom. 86
- Oracle. 66
- Nvidia. 54
We are overshooting the mark. We are at or near a peak. Many of the data centers and chip foundries we’re building now are going to fail or become a commoditized surplus capacity in a few years. Longer term this will be good for end users and consumers, as prices will fall, but some investors and operators will lose. This correction is likely to hit the semiconductor industry, the data centers, and the model developers hardest, with lesser ripple effects for network providers.
Public policy amplifies the bubble. AI Policy now is divided between those who think thot Big AI is going to take over the world and destroy us if not carefully “guardrailed” and regulated by the state; and aggressive mercantilists who think a few states are in a race to be the biggest in Big AI, because the winner of the race will dominate the world. They’re both wrong, but they both press us toward more concentration and centralization. The only thing that can cure us of this disease is the inevitable deflation of these expectations as AI hype meets business realities.
The United States has led the computing and software industry for the past 80 years; it has led Internet, platforms, cryptography and cloud computing for the past 40 years. It has led the telecommunications industry for the last 125 years. All these ICT industries thrived mainly because of open markets and competition, and the US led because it was the most liberal in all sectors. We did not succeed because of governmental golden shares, trade protection, export controls and tariffs. Government investments in universities and in science and technology research also played a role, but there was a clear division of function between the government support for public goods like basic science, and the commercial development of private goods like chips and Internet services on the other. Another reason the US led the world was because it actually had market access to the world and the world had market access to the US. Our firms were able to globalize, find markets in any location, draw talent, labor and services from any location. A downturn is imminent not only because of an inevitable correction to over-investment, but also because we are reversing the policy formula that generated leadership and success.
