- Wall Street Journal columnist Andy Kessler argues that cheaper AI models and the shift toward on-device computing could eventually undermine the economics of the data center boom.
Wall Street Journal columnist Andy Kessler argues that the artificial intelligence boom could eventually hit a turning point as cheaper models, more efficient chips and on-device computing challenge the economics of massive data centers.
In his October 11 opinion column, “How the AI Boom Ends”, Kessler predicts that demand for AI services will continue growing, but prices could fall sharply as the technology becomes more efficient. He warns that the resulting pressure on revenue and returns could leave some infrastructure investments struggling to justify their costs.
The argument comes as technology companies continue committing billions of dollars to AI chips, servers and data centers. The spending has fueled growth for semiconductor suppliers and cloud providers, but it has also increased scrutiny of whether the industry can generate enough revenue to support its expanding infrastructure.
Why Kessler Expects AI Prices to Fall
Kessler compares the current AI infrastructure model with earlier phases of computing, when expensive mainframes gradually gave way to smaller and cheaper systems. He argues that AI is entering a similar period of refinement, in which more efficient architectures could deliver comparable capabilities at a fraction of the cost.
One of his central points is that the cost of running AI models is likely to decline faster than many investors expect. Kessler cites venture capitalist Steve Jurvetson’s view that major technology shifts often occur when a new approach becomes roughly 10 times more efficient or cheaper.
The column points to the growing availability of lower-cost models from Chinese developers, including Z.ai and Alibaba, as an example of the pressure facing premium AI services. Kessler also highlights smaller language models that can run on personal computers and smartphones, reducing the need to send every request to a large, remote data center.
Advances in model compression and smartphone processors could accelerate that shift. Kessler cites work on compact AI models and newer neural processing units that could allow increasingly capable systems to operate locally, with costs largely tied to hardware and electricity rather than ongoing cloud-based inference.
The transition would not eliminate the need for large data centers. Kessler acknowledges that model training, scientific research and other computationally intensive tasks will continue to require substantial computing resources. His argument is that a growing share of everyday AI activity could migrate toward devices at the edge of the network.
That possibility creates an important question for investors in semiconductor companies benefiting from the AI earnings boom: whether demand growth will be sufficient to offset falling prices and changes in where AI workloads are processed.
Could Falling AI Costs Trigger a Data Center Spending Slowdown?
Kessler warns that lower prices do not automatically translate into stronger profits. If providers reduce prices substantially, they must attract a much larger volume of usage to maintain the same revenue, assuming other factors remain unchanged.
He illustrates the problem with the history of telecommunications infrastructure. During the dot-com era, companies invested heavily in networks to meet expected internet demand, but collapsing prices and excessive debt contributed to failures, including Global Crossing.
Kessler draws a parallel with the current AI infrastructure buildout. If model costs fall rapidly while data center operators are still carrying large financing commitments, some projects could become less economically attractive. He raises the possibility of construction delays, canceled projects and debt defaults if the economics deteriorate sharply.
The column also asks whether 2027 could mark a peak in data center investment, while acknowledging that the timing is uncertain. This is Kessler’s forecast, not an established industry consensus or a confirmed turning point.
The concern is relevant to the broader debate over how much major technology companies are spending on AI infrastructure. Continued spending can support chipmakers, networking suppliers and data center operators, but investors ultimately need evidence that the investment will produce adequate returns.
The risks also differ across the supply chain. Companies selling advanced chips may benefit from demand for training and inference even if individual AI services become cheaper. Cloud providers, meanwhile, could face pressure if prices decline faster than utilization rises or if customers shift more workloads to their own devices.
For investors, the key question is not simply whether AI adoption will continue. It is whether revenue, pricing and capital costs will evolve in a way that supports current valuations. Wealthier Today’s analysis of stocks positioned for the next AI spending cycle explores companies exposed to continued investment, while its coverage of bond yields and growth-stock valuations examines another factor influencing how markets price long-term growth.
Kessler’s broader conclusion is that technological progress can create winners while undermining the business models and investments built around an earlier generation of technology. AI demand may keep expanding even if the economics of data center infrastructure change dramatically.