AI is more than infrastructure.

When cloud computing arrived, the story was infrastructure—Amazon Web Services (AWS) turning services like EC2 and S3 into clean, metered utility computing. AWS succeeded not because it invented new application models, but because it removed constraints that allowed existing ones—especially SaaS—to scale more efficiently and globally.

Companies building on AWS were not primarily judged on infrastructure consumption as a success metric. Nobody walked into a board meeting saying they are succeeding because their AWS bill went up. While cloud costs mattered operationally, success was ultimately defined by the value delivered to end users through the applications built on top of that infrastructure.

The AI wave is likely to follow a similar pattern. The defining measure of success will not be token volume or infrastructure spend, but whether AI enables domain-specific applications that solve real problems and deliver measurable value in specific industries and workflows.​

The cloud era proved where value accrues.

​Frontier model providers today resemble cloud hyperscalers, capturing significant value as foundational infrastructure in the AI stack. But token consumption is not a meaningful success metric for the consumers of frontier models. Treating it as such risks repeating a familiar mistake: optimizing for infrastructure usage rather than the value created on top of it. More tokens in do not necessarily translate to more value out.​

A misleading narrative is emerging—particularly among some vendors and investors—that equates token usage with enterprise success. I believe this is wrong. Spend is not the metric that matters. Outcomes are.

In this context, success does not mean consuming the most intelligence, but building the most effective applications: domain-specific, purpose-built systems measured by real-world results rather than raw usage volume.

Tokens are the new power grid—an enabling layer, not the measure of success. Advantage will come from what is built on top of it. Companies that deeply understand their domains and apply AI to concrete problems will define this wave. Others risk optimizing for activity rather than impact.​

AI is repeating the same measurement mistake.

​Security operations have lived through a version of this dynamic for over a decade. Early SIEM models often encouraged a “collect everything” approach, where ingestion volume became an implicit proxy for coverage and value. The logic was simple: more logs should mean better visibility. In practice, costs increased faster than operational clarity.

Over time, security leaders pushed back—not because telemetry is unimportant, but because volume alone was never a reliable measure of security effectiveness.

What ultimately matters are domain-specific outcomes: detection quality, triage accuracy, investigation speed and the ability to scale operations without linearly scaling analyst headcount. These are the signals of a mature security operations function—not the amount of noise it processes.

This pattern is now resurfacing in AI. Treating token spend as a proxy for success is familiar logic in a new disguise.​

Security teams learned this lesson.

​The killer app isn’t a marketing strategy; it’s an architecture. AI cannot simply be layered onto existing processes and expected to deliver transformation. Without structured data, clearly defined workflows and systems designed for machine reasoning, higher inference spend does not create value—it often amplifies inefficiency.

The companies that are likely to define this wave are asking harder questions: What problems do we uniquely understand? Where is our domain expertise materially stronger than general-purpose models? What outcomes become possible now that were previously out of reach?

Do we have the local knowledge—encoded as context, workflows and structured representations of decision-making—that enables high-efficacy, defensible and verifiable automation in real-world environments?

In practice, this looks like large language models combined with smaller, domain-specific systems—whether those are fine-tuned models, structured context graphs or workflow and decision traces—that together form domain-trained platforms. These killer apps will be the systems that I believe will succeed in embedding intelligence directly into how work is actually done.​

The defining metric of the AI era is impact.

​For AI companies and enterprise adopters alike, the shift is clear: stop measuring inputs and start measuring outcomes. Are customers realizing meaningful value from applications built on frontier models—value that goes beyond raw model outputs? Is the “last-mile gap,” shaped by domain knowledge and workflow integration, clearly translating into higher effectiveness?

Token consumption is a relevant revenue signal for frontier model providers. For everyone else, it is largely a distraction from what actually matters.

The key question is not how many tokens were consumed, but what new capabilities were unlocked for customers. That is the product. That is the business. The AI era will not be defined by compute consumption, but by domain precision—and by the companies that can translate that precision into real-world capability.​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?