For much of the past two years, AI was sold as an inevitability: exponential gains, limitless demand, and profits that would eventually justify the infrastructure build-out. That story is starting to crack.
Not because AI stopped working, but because it started showing up on balance sheets.
Across Big Tech and venture-backed startups, usage continues to rise. Microsoft stated more than 85% of Fortune 500 companies now use at least one of its AI products. OpenAI reported more than 900 million weekly active users in December 2025. The shift is in how executives, investors, and finance teams talk about that growth.
The language of breakthroughs has been replaced by the language of unit economics.
Where the Costs Accumulated
Running LLMs remains expensive. Training costs are largely sunk, but inference has proven harder to compress than expected.
Executives at Google Cloud and Microsoft have stated in recent earnings calls that AI workloads currently deliver lower gross margins than traditional cloud services. Analysts from Bernstein and Morgan Stanley point to energy consumption and inference as the fastest-growing expenses, particularly as token usage rises at scale.
Microsoft’s AI CEO has said that keeping pace with frontier AI will require “hundreds of billions” of dollars in investment over the next decade.
The Shift From Scale to Returns
In 2023 and early 2024, AI spending was framed as a land grab. By late 2025, the conversation had shifted. CFOs are now asking which workloads generate revenue or measurable savings, and which merely demonstrate technical capability.
Companies that rushed into broad AI pilots are consolidating vendors, capping usage, or pulling workloads in-house where possible. Several firms that announced sweeping rollouts have quietly narrowed scope to specific use cases including customer support triage, code review – areas where cost savings can be clearly measured.
This shift doesn’t show up as a collapse in demand. It shows up as slower revenue expansion and tighter contracts.
Capital Spending Wasn’t the Mistake
Billions have been poured into data centres to support AI workloads, and much of that investment is still generating demand. Chipmakers such as Micron have reported strong sales driven by AI data centre build-outs. What proved wrong was the belief that revenue would scale faster than operating costs.
AI infrastructure behaves less like software and more like industrial equipment. It requires constant power, cooling, and regular hardware refresh cycles. Investors have begun to price this reality in.
Companies such as Oracle have seen their shares come under pressure after projecting AI-related capital expenditure running into the tens of billions of dollars.
Why Agents Are Gaining Ground
That cost pressure explains why AI agents and orchestration layers have moved from niche tooling to strategic priorities. The shift is not about more impressive demos. It is about economics.
By reducing repeated model calls, caching decisions, and pushing logic outside the model, companies can materially lower per-task costs. These systems often look unremarkable in presentations and far more compelling in operating margins.
Why Narrow Data Wins
The idea that the internet’s text corpus has been exhausted is broadly accepted inside the industry. What’s less acknowledged is that most commercial value doesn’t come from general knowledge.
It comes from narrow, high-quality data: internal documents, transaction histories, industry-specific workflows. Progress is slowing at the frontier, but accelerating in constrained environments where accuracy matters more than novelty.
That is not the end of AI development. It is the end of spectacle.
A Familiar Ending
Cloud computing went through a similar cycle a decade ago. Early hype promised limitless margins. Reality delivered steady growth, intense competition, and constant pressure to optimize costs. Cloud did not fail. It became routine.
AI appears to be heading in the same direction. The hype premium is fading. The technology remains. What follows is the harder work of making it pay.
