AI Deflation Is the Thesis, Not the Threat

Recent media commentary asks whether cheaper, smaller AI models could cool data center demand. Deflation is disruptive, but it is the disruption we model and invest around, and we see no sign of a tipping point.

Falling Costs Are What We Look For

ARK’s research blends top-down and bottom-up work, and it often starts with a cost-decline curve. The question is which law to trust. Many investors default to Moore’s Law. We prefer Wright’s Law, and the irony is that it is the older of the two. After nearly 90 years, it remains an excellent forecasting tool.

Cost Curves Have Built Industries Before

AI Is the Newest and Steepest Curve

ARK research has found AI training costs falling roughly 75% a year and inference costs roughly 90%, and well over 95% by some estimates.¹ In ARK’s analysis, about half of the decline comes from hardware and half from software efficiency. When AI costs fall, it is not a surprise and not a coincidence.² It is the pattern we underwrite.

Cost declines have not just been passed through as lower prices. They have been reinvested in capability, and each step below consumed far more compute than the last.

Case Study: Openai’s Reported Navier-Stokes Run (Sept. 2026)

We’ve Seen This Before in AI: The DeepSeek Moment

Markets have spent the time since digesting AI deflation, and the evidence has favored demand over fear. The current debate about smaller models is the same question in a new package, not new news.

Lower Costs Unlock New Convergences

Less expensive AI does more than make today’s use cases cheaper. As costs fall, progress in AI spills into other platforms and opens markets that were not economical before.

On-Device AI Complements the Cloud

Small models that run on a phone or laptop keep improving, but on-device AI adds to data center compute rather than replacing it. Siri already works this way, handling simple requests locally and complex ones in the cloud. OpenAI and Anthropic both released cloud-based environments for Codex and Claude Code, tools that traditionally run on a developer’s device, because developers wanted mobility and no dependence on a charged, connected laptop.

The Bottom Line

When intelligence gets less expensive, we find more to do with it. We see no sign that cycle is ending, and our research process is built to capture it.