Economy · Enterprise AI
Your AI Agents Cost More Than the Employees They Replaced
In April 2026, Uber's CTO disclosed something that should stop every enterprise AI buyer cold: the company had already burned through its entire 2026 AI coding tools budget — in four months. That wasn't mismanagement. Uber had actively incentivized adoption with internal leaderboards ranking teams by AI tool usage. They did everything the AI industry told them to do. And the math still broke. Uber is not an outlier. It's the canary.
The budget that vanished in four months
Uber's CTO Praveen Neppalli Naga told The Information in April that the company had exhausted its full-year AI budget on Claude Code alone. The COO is now openly questioning whether the spend is worth it. This isn't a startup burning VC money on vibes. This is Uber — a public company with $44 billion in annual revenue and some of the sharpest engineering leadership in tech.
What makes the Uber story instructive isn't the dollar amount. It's the mechanism. Uber gamified AI adoption — the more tokens your team burned, the higher you ranked. The implicit message from leadership was: use more AI. Engineers responded rationally. And the budget evaporated in a third of the planned timeline.
This pattern — incentivize adoption, watch costs detonate, quietly pull back — is now repeating across the industry at astonishing speed.
Microsoft pulls the ripcord
In May 2026, Microsoft began canceling most direct Claude Code licenses across its Experiences and Devices division, according to The Verge. Engineers were redirected to GitHub Copilot CLI with a June 30 deadline. This came roughly six months after Microsoft first opened Claude Code access to thousands of developers, project managers, and designers. The tool became popular. Too popular. The scale at which employees used it forced a reversal on a product its own engineers had come to depend on.
Microsoft isn't abandoning AI. It's abandoning unmetered AI. The distinction matters. Canceling Claude Code licenses doesn't affect Microsoft's $5 billion Foundry investment in Anthropic, or Anthropic's $30 billion Azure compute commitment. The infrastructure bet continues. The per-seat per-day usage bet is what broke.
Microsoft is far from alone. By June 2026, Amazon, Walmart, Cisco, Uber, and Meta had all introduced internal caps on AI tool spending, discouraged wasteful use, or pushed employees toward cheaper models, according to the Australian Financial Review. The phrase one executive used: "We created a monster."
The Nvidia VP who said the quiet part out loud
Bryan Catanzaro is Nvidia's vice president of applied deep learning. His team builds AI. His employer sells the chips that power it. And in an April 2026 interview with Axios, he said something extraordinary for someone in his position:
"For my team, the cost of compute is far beyond the costs of the employees."
This is Nvidia's own deep learning chief — not a skeptic, not a consultant selling AI-disillusionment reports — stating plainly that his AI compute bill exceeds his team's payroll. When the company selling the shovels tells you the gold is costing more than the miners, the economics have fundamentally shifted.
Catanzaro's comment landed in an environment already primed for cost anxiety. Big Tech had collectively committed roughly $700 billion to AI-related expenses in 2026 — a 69% jump from 2025. Goldman Sachs projected $527 billion in AI capex for the year. The buildout is staggering. The return, for most companies, is not.
The math nobody did: why agentic AI destroys budgets
The cost crisis has a specific technical cause that most procurement teams still don't understand: agentic AI doesn't use a little more compute than a chatbot. It uses catastrophically more.
Standard LLM interaction: one prompt, one response, maybe a few hundred tokens. Agentic coding task: the model reads your codebase, plans a solution, writes code, runs tests, reads errors, rewrites code, runs tests again — a loop that can call the model dozens or hundreds of times for a single task. A Microsoft Research study published in April 2026 found that agentic coding tasks consume roughly 1,000 times more tokens than standard code-chat interactions. The same task varied by up to 30x in total token consumption across runs. EY estimates a complex agentic interaction in 2026 costs approximately 30 times what a simple chatbot query cost in 2023.
This is not a linear scaling problem. It's exponential. And it means every productivity metric that assumes "more AI usage = more value" is structurally wrong.
The culture that emerged around this — "tokenmaxxing," as it's now dubbed — made things worse. Jensen Huang's comment that a $500,000 engineer should consume $250,000 in tokens annually became a permission structure for competitive overconsumption. Workers began racking up $150,000 in monthly token bills. Internal leaderboards at Meta (dubbed "Claudeonomics") and Uber rewarded teams by total AI tool usage. The incentives were perfectly designed to maximize cost, not output.
The ROI gap is real — and quantified
Morgan Stanley found that only 21% of S&P 500 companies can point to a concrete AI benefit. Gartner's 2026 data shows 80% of AI projects fail to deliver business value. RAND Corporation research revealed something more damning: companies that cut staff to fund AI saw identical financial returns to those that didn't. The efficiency dividend, on a net basis, hasn't materialized.
Salesforce is an instructive case. In January 2026, CEO Marc Benioff called Agentforce the company's most important product launch in years. Three months later, four senior executives tied to the initiative had left. Then came nearly 1,000 layoffs, with the Agentforce team among those affected.
The pattern is consistent: aggressive AI adoption → cost shock → quiet rollback. Not because the technology doesn't work — it does — but because the economics of deploying it at scale are punishing in ways the industry hasn't been honest about.
The token pricing trap
There's a familiar script in platform economics: subsidize adoption, lock in dependence, then raise prices. Cybernews called it out explicitly in May 2026: "Token costs are beginning to follow a familiar script — affordability at the start gives way to rising costs once dependence is locked in."
Even Gartner has warned that cheaper per-token costs won't translate to cheaper enterprise AI. The reason is the volume trap: as unit costs fall, agentic models consume exponentially more tokens per task, increased consumption outpaces falling unit costs, and AI providers have no incentive to fully pass through savings.
The economic viability ceiling is lower than most assume. Research from MIT and other institutions suggests automation is only economically viable in roughly 23% of job tasks at current token prices. For the other 77%, humans remain cheaper — even before factoring in the monitoring, error correction, and oversight that AI outputs still require.
What to actually do
I've been running AI coding agents daily for months. When I actually audited my token spend, I realized I was burning through the monthly equivalent of a junior developer's salary on tasks a human could handle in an afternoon. The agent was good. The math was bad.
Here's what the companies getting this right are doing differently:
1. Audit token spend per task, not per seat. A $200/month Copilot seat is predictable. A Claude Code agent that loops 40 times on a single bug is not. Track cost per completed task — not per license — and compare it against the fully-loaded cost of a human doing the same work.
2. Tier your models ruthlessly. Claude Opus or GPT-5.5 for architecture decisions and complex debugging. A cheaper model — DeepSeek, GLM, or even a fine-tuned small model — for boilerplate, tests, documentation, and code review. One enterprise I've spoken with reduced token costs by 60% just by routing 80% of their prompts to cheaper models.
3. Set the productivity bar at 2x. If an AI-augmented worker isn't at least twice as productive as a non-augmented one, you're losing money after token costs. This isn't a vibe check — it's a hard constraint. Measure it.
4. Beware of agentic AI for routine work. The 1,000x token multiplier means agentic workflows have the worst ROI right now. Use task-specific, single-pass models for classification, summarization, and extraction. Reserve agentic loops for work that genuinely needs multi-step reasoning — and cap the number of turns.
5. Lock in fixed-rate contracts. If you're dependent on API inference, negotiate volume pricing now. The cost curve is pointing in the wrong direction, and the providers have the leverage. Self-hosted open-weight models — even if they require more setup — give you a cost ceiling that APIs don't.
The companies that win won't use the most AI
The AI cost crisis isn't an argument against AI. It's an argument against using it indiscriminately. The most dangerous thing an enterprise can do right now is measure AI success by adoption metrics — tokens burned, seats deployed, leaderboard rankings. Those numbers go up. Costs go up faster.
The companies that navigate this won't be the ones that used the most AI. They'll be the ones that learned, quickly and sometimes painfully, when not to. The technology works. The economics, at scale and at current prices, often don't. Pretending otherwise is how you blow a year's budget in four months.