The AI Spending Debate Has Two Camps. The CTOs Building At Scale Belong To Neither.

Enterprise AI advantage now hinges on whether a system stays solvent at production volume. That problem only appears after the demo works and the token bill arrives. The most capable models sit one API call away for anyone willing to pay, so the differentiator has moved from access to economics, from which model a team can reach to whether it can run that model millions of times a month without breaking the business. The public argument over how much to spend has been loudest at the extremes, but for the CTOs building AI-powered products, survival depends on a subtler discipline: routing each task to the cheapest model that will do the job, hedging against providers that reprice mid-quarter, and knowing when to leave AI out of the loop entirely.

Eric Belair is Chief Technology Officer at Polaris I/O, a decision-intelligence platform that converts unstructured global-event data into prioritized signals for enterprise account teams. Belair architects the agentic backend that lets Polaris process vast volumes of continuous, unstructured information without the economics collapsing under it.

"Everyone has access to the same models now. The differentiator is economics. Can you call an AI agent ten million times a month and still have a business?" says Belair, locating the real contest in the cost of running a model at scale rather than the ability to reach one. The framing cuts against a debate that has played out at the extremes, and understanding why it matters means looking at how far apart those extremes have drifted.

The two camps, and the ground between them

On one side sits the maximalist case, argued from the biggest stages in tech. Nvidia CEO Jensen Huang told the All-In Podcast at GTC 2026 that a $500,000 engineer who consumes only $5,000 in tokens over a year should set off alarm bells, and that he would be deeply alarmed if that same engineer burned through less than half their salary in compute. Meta CTO Andrew Bosworth pushed the same direction, saying his best engineer spends the equivalent of his salary in tokens while producing far more, and framing it as easy money with no ceiling.

On the other side sits the data. Ramp's June 2026 AI Index, drawn from transaction records across more than 70,000 U.S. businesses, found the top 1% of companies spending $7,500 monthly per employee on AI, while the median firm spends $11.38, roughly the cost of a single ChatGPT or Claude subscription seat. That 680-fold gap means almost nobody is yet spending on AI what they spend on an engineer's salary, even as the trajectory climbs. For the teams building at production volume, the argument between the two camps skips the question that decides who survives, which is whether the math holds when the system runs continuously across an entire customer base.

The cost discipline that matters

Belair's team runs agentic backend systems, which are networks of AI agents performing different tasks across customer accounts, from surfacing signals about leadership changes and supply-chain disruptions to ranking the relevance of those signals for different teams. The scale creates a constraint that most AI demos never encounter, since the system processes vast amounts of unstructured information continuously, around the clock.

"If you're taking a naive approach to the problem, that's going to end up bankrupting you," Belair says. "We've really spent a lot of time making sure we're getting the right answer cheaply, and getting it right the first time rather than optimizing after the fact."

Large engineering organizations are discovering the same constraint from the opposite direction. Uber burned its entire 2026 AI budget by April after Claude Code spread across 5,000 engineers faster than finance expected, then capped spending at $1,500 per employee per coding tool. That tension, a tool too valuable to drop and too expensive to leave uncapped, is what Belair designs around. Each agent call earns its keep, with analytics matching the model to the task and its cost, since some problems do not need a frontier model and a few do not need AI at all.

"There are times where you have to know when not to use AI," he says. "Sometimes a deterministic system is the right answer. We have to maintain discipline about when to lean into it and when we don't need to."

The vendor risk that keeps CTOs up at night

The cost equation gets more volatile once you factor in what Belair calls "rug pulls," the sudden changes in pricing, usage limits, or subscription models that arrive from the major AI providers. "We constantly have to work around those limitations and be aware that at any point in time we could get another rug pull where maybe the token cost skyrockets for whatever reason," he says.

The anxiety is grounded in recent events rather than speculation. Anthropic announced that programmatic Claude usage would move to a separate metered credit billed at full API rates, then reversed course and cancelled the change on the day it was due to take effect. The same season, Anthropic shipped Claude Opus 4.7 with a new tokenizer that, by its own documentation, can map the same input text to up to 35% more tokens, so a request could cost more even though the rate card stayed flat. OpenAI, for its part, rolled out enterprise spend controls and monthly credit limits, moving generative AI toward a metered utility model that finance teams have to forecast rather than a flat site license they can budget once. For teams running production systems, every one of those changes forces a recalculation of unit economics.

Belair's hedge is architectural. He designs model-agnostic systems with hot-swap capability, complemented by selective use of open-source, self-hosted models for certain workloads. The pipeline can run a frontier model on a given agent today and switch it tomorrow without a rebuild. "We never want to paint ourselves into a corner," he says. "Even if we determine that a specific model is the best for a particular agent, we're not married to it."

The feedback loop gap

The most forward-looking part of the conversation is where Belair sees the industry falling short. Most feedback loops, he says, are still human-gated. "It might just be updating a skill markdown file as a pull request where a human has to go in and approve it, or collecting data that a data scientist has to analyze."

The tooling is moving, LangChain's Better-Harness among it, but it stops short of what Belair is after. He wants agents that rank feedback from subject-matter experts, weight it by expertise, and improve on domain-validated signals without a person in the loop, and he says plainly that nobody is there yet, his own team included. "Eventually someone will come up with the way to help your agents improve themselves. Maybe it's from LangChain, maybe someone else, maybe we come up with our own way."

The build-versus-buy question that won't sit still

As the major platforms bundle more orchestration, observability, and agentic tooling into their offerings, the case for building your own gets harder to make every quarter, which is the same tension driving other operators to argue that companies must own the workflows that create their value rather than rent them from a generalized model. Belair's team builds anyway, for now, using LangGraph for orchestration and managing its own agent architecture and context engineering.

"There are definitely going to be organizations that build their own fleet of agents," he says. "And there are companies like Vercel that just let you deploy agents on their infrastructure. We're a nimble team. We don't have as many resources devoted to that type of task as some of these bigger companies. So it's always something we're keeping our ears open to."

What this means for senior decision-makers

Belair's experience suggests every AI infrastructure investment ultimately comes down to three operational questions. The first is what a system costs to run at production volume rather than at demo scale. Ramp's data shows a 680-fold spending gap between the most AI-aggressive firms and everyone else, and the blended cost of AI fell 67% year over year, from $18.40 to $6.07 per million tokens, even as total spend climbed because agentic systems burn many times more tokens per developer than the workloads the budgets were built around.

The next is what happens when the model provider changes terms. The Anthropic tokenizer shift and OpenAI's move to metered enterprise billing are only the most recent reminders that a provider can reprice the API a whole product runs on, which is what makes architectural flexibility non-negotiable.

The last is whether the system improves on its own. If the answer runs through a human reviewing every change, it is a human-gated process wearing an agentic label. The gap between human-reviewed systems and ones that improve safely on their own may become the next real competitive advantage. For now, the industry's marketing still runs ahead of its engineering.