Token Era Forces Mastery of AI Economics

In the mid-1980s, a young engineer named Daryush Ashjari started his IT career in the thick of the mainframe era, when a single metric governed enterprise computing: MIPS, or millions of instructions per second. It was the currency of the data center. Expensive, finite and vendor-locked.

Four decades later, Ashjari sees history repeating. The currency has a new name. The economics feel familiar.

"History is clearly repeating itself," said Ashjari, chief technology officer for Asia Pacific and Japan at Nutanix. "We've simply traded instructions for tokens."

Deloitte describes AI tokens as the unit of consumption that links system performance to financial outcomes, and argues that self-hosted infrastructure can deliver better unit economics for predictable, high-volume workloads. In July, Diginomic reported on KPMG’s Global AI Pulse Research that found cost visibility is becoming a leadership priority as token economics and agent-driven workflows introduce variables that traditional IT budgets cannot manage.

Ashjari has weathered many major waves of innovation. In the early days of any technology shift, the industry obsesses over raw scale and adoption velocity. How many can we use? How fast can we deploy? Then the market matures, and the question changes. How efficiently can we consume them? That inflection point has arrived for AI, and it is forcing enterprises to confront a reckoning over token economics, infrastructure readiness and data accuracy.

The stakes are quantified in the 2026 Enterprise Cloud Index (ECI) from Nutanix. The survey showed 85% of IT leaders see AI accelerating their modernization plans. Yet 82% admit their current on-premises infrastructure is not fully ready to support intensive AI workloads. The gap between ambition and readiness is where costs spiral, as teams resort to inefficient, siloed deployments just to get AI running.

The tension is sharpest in regulated industries. The same ECI found that 80% of organizations now prioritize data sovereignty when making infrastructure decisions. They want to run AI locally. Their data centers are not ready.

"The readiness gap is where token costs spiral out of control," Ashjari said. "Teams resort to inefficient, siloed deployments to get AI running."

What started as a scramble to appease one workload has grown into an increasingly sought-after alternative to fragmented infrastructure. The data-center equivalent of a Swiss Army knife in a world of single-blade tools.

The MIPS Lesson, Rewritten for Tokens

Ashjari draws two lessons from the mainframe era that apply directly to the token economy. The first is that efficiency is the only sustainable strategy. Just as efficient COBOL code saved MIPS a generation ago, efficient model orchestration saves tokens today. The second is that infrastructure matters more than the model. You can have the best application, but if the underlying platform is fragmented and under-utilized, the cost will eventually kill the project.

A third lesson has emerged that the MIPS era only hinted at: data accuracy as the new quality control. In the mainframe days, "garbage in, garbage out" led to wasted processing time. In the token era, inaccurate data leads to hallucinations that are not just costly in wasted tokens but dangerous for business integrity.

"A token spent on a hallucination is a token wasted," Ashjari said. "Unlike the MIPS era, where an error might just crash a job, an error in the token era can propagate misinformation at scale."

For regulated industries, Ashjari advocates taking a sovereign AI approach. Banks, for example, can run AI agents in their own data centers to satisfy data sovereignty requirements while avoiding the unpredictable egress and token costs of public cloud APIs.

The Great Replatforming

Ashjari calls the current moment the "Great Replatforming." Gartner forecasts worldwide AI spending will total $2.59 trillion in 2026, a 47% increase year over year, with AI infrastructure accounting for more than 45% of the market. As enterprises move into the token era, Ashjari said, they must not repeat past mistakes by creating new silos.

The challenges extend beyond cost. Organizations are fighting complexity and risk on four fronts: the data-readiness hurdle, where unstructured "dark or dirty data" lacks the lineage required for model training or retrieval-augmented generation; the infrastructure-readiness gap; organizational silos and Shadow AI; and a skills shortage that has outpaced the talent market.

"Teams are struggling to find people who understand both the old world of stable infrastructure and the new world of agentic AI orchestration," Ashjari said.