Editor’s note: GeekWire publishes guest opinions to foster informed discussion and highlight a diversity of perspectives on issues shaping the tech and startup community. If you’re interested in submitting a guest column, email us at [email protected]. Submissions are reviewed by our editorial team for relevance and editorial standards.
Many economists argue that AI will automate white-collar work at an exponential pace, suppress wage income, collapse aggregate demand, and potentially devastate our economy. The logic is simple: if machines can think, why would anyone pay humans to do white collar work? Thought experiments like the widely circulated 2028 Global Intelligence Crisis offer a clean story arc: AI capabilities compound recursively, labor is displaced wholesale, and the economy spirals downwards. The displacement doomers have the rhetorical advantage of a terrifying punchline.
A recent note from the economists at Citadel Securities, The 2026 Global Intelligence Crisis, pushes back. Their core argument: just because AI can improve itself does not mean businesses will adopt it at the same accelerating pace. When technology makes workers more productive, that tends to increase the supply of goods and services, which is generally good for the economy. And there are real-world limits on how quickly companies can swap out human workers for AI systems.
The note’s most original insight is what I would call the compute-cost ceiling. If companies rush to automate everything at once, the demand for computing power surges, and the price of that computing power goes up. At some point, it becomes cheaper to pay a person to do the task than to pay for the AI to do it, and the automation stops. A natural economic brake, and one that the displacement doomers ignore entirely.
But hold on. There is a massive hole in this argument, and I’m about to drive a data center through it.
The compute-cost ceiling argument overlooks the fact that compute costs are dropping off a cliff. The famous Moore’s Law (the cost of compute halves every 18 months) has retired after a remarkable 50-year run, but there is an AI version where costs are falling by a factor of 10 each year. Specifically, LLM inference costs (on a per-token basis) have been dropping by roughly an order of magnitude YoY for the past two years. Andreessen Horowitz coined the term “LLMflation” to describe the trend, documenting a roughly 1,000x drop in cost over three years. In fairness, the cost per task is declining more slowly because frontier models are burning more reasoning tokens per query, but it is still moving down fast. A ceiling that drops by 10x every year is not a ceiling. It is a speed bump.
The Citadel note also invokes John Maynard Keynes’s infamous 1930 prediction of the 15-hour workweek, noting that Keynes was wrong because he underestimated human appetite for consumption. People just wanted more stuff. Fine, but this misses the distribution problem. The “people will just want more stuff” argument only works if enough people have the income to buy more stuff with. If AI-driven gains flow to the top 0.1%, and the rest of the population is out of luck, then the economy is in trouble. In other words, Musk and Bezos cannot really consume much more, and there are not enough of them to keep the economy humming.
So where does this leave us?
The Citadel note is right that institutional frictions slow down AI deployment, and that democratic societies will eventually respond with policy adjustments. These are real brakes, which buy time. But the forces pushing costs down are powerful: algorithmic efficiency, hardware improvements, quantization, distillation, and fierce price competition among inference providers. None of these are slowing down.
My conclusion is that the AI displacement doom loop is overstated and will take longer to arrive than many anticipate. But it is coming, and we are in for major economic and perhaps social upheaval. The doomers ignore the frictions. The optimists assume the frictions will hold forever, even as the cost curves erode them year after year. The reality will land somewhere in between, and it will be messy. Preparing for that is the real policy challenge, and so far, our policy makers seem to have their attention focused elsewhere.
