By April 2026, the executive suite at Uber hit a wall that wasn’t on any of their Q1 projections. Their entire annual AI budget, a figure meant to last until the following January, had vanished in just four months.
The culprit was agentic coding spend.
Uber’s case is the cleanest example of what happens when frontier tooling escapes the lab and lands in a 5,000-engineer organization without hard inference controls. Roughly 84% of Uber’s 5,000 engineers started using agentic coding tools, and individual token costs climbed as high as $2,000 per month for heavy users. That mix: broad adoption, recursive tool use, and no meaningful spend friction: was enough to exhaust the company’s entire 2026 AI budget by April. At that point, this stops being “developer productivity software” and starts behaving like an unmanaged compensation line.
I have been calling this the arrival of the Silicon Salary. For the last three years, the narrative was that AI would commoditize labor. In 2026, the script has flipped: high-end inference has become so expensive that, for 77% of tasks, the person sitting in the chair is actually the budget-friendly option.
Kill the “All-You-Can-Eat” Fantasy
For a brief window, we lived in a subsidized reality. Big Tech was willing to bleed margins to win the land grab, offering “Unlimited Copilot” seats for $19 a month. Those days are officially dead.
The shift at GitHub from “Premium Requests” to AI Credits is one example of Inference Inflation. In the old model, a “request” was a flat unit of currency. It didn’t matter if you asked for a typo fix or a 10,000-line refactor: the price was the same. Under the new 2026 credit system, you pay for the weight of the thought.
If your agent spends an hour “reasoning” through a complex dependency tree, and you’re paying for the compute. For companies that moved too fast into managing AI agent sprawl, this has created a “bill shock” that is forcing a radical rethink of human labor.
The 77% Cost Reality
The most significant piece of evidence in this shift comes from the MIT CSAIL FutureTech study. Researchers moved past the “can AI do this?” question and asked “should we pay for it?”
Their findings were a bucket of cold water for the “automate everything” crowd: 77% of vision-based tasks are still cheaper to perform with humans.
Take their now-famous “Bakery Example.” Technically, an AI system can inspect every loaf of bread for quality. But the fixed costs of development, deployment, and the specialized hardware: plus the ongoing “Inference Salary” of the model: dwarf the cost of a human baker. The study found that while AI might save $14,000 in labor, it could cost over $160,000 to implement and $120,000 a year to maintain.
In 2026, we are seeing this play out in the knowledge economy. When you factor in the Agentic Surcharge: the cost of an AI agent “thinking” recursively: you find that many middle-management and engineering tasks are now 2x more expensive to perform with AI than with the existing human team.
The Agentic Surcharge
Why is this happening now? Because we moved from “Chat” to “Agents.”
A chatbot is a one-and-done transaction. An agent is a loop. When an engineer uses a tool like Claude Code or GitHub’s new agentic architecture, that tool might make 50 calls to a frontier model to solve a single ticket.
If that model is a high-reasoning variant like Opus or GPT-5, each “step” in the agent’s logic adds up. We are seeing cases where a “fully autonomous” bug fix costs $150 in tokens. A human engineer, even at a high hourly rate, can often identify the same bug in ten minutes for a fraction of that “Inference Salary.”
The Pullback Stories
The clearest pullback signal right now is Microsoft’s decision to wind down Claude Code access in its Experiences and Devices group by June 2026. By internal accounts, Claude Code was incredibly effective, which is exactly why usage exploded. In some teams, adoption reportedly hit 84%. That is what happens when a tool genuinely changes developer flow: people do not need to be pushed into it. They pull it into every workflow they can.
Then the invisible bill shows up. Claude Code is not a static chatbot. It is an agent. It runs loops, recursive searches, and multi-step reasoning chains behind the scenes. That means the real unit of cost is not “one prompt.” It is accumulated inference. In practice, that pushed spend into the thousands of dollars per engineer per month for heavier usage. So does this mean Microsoft is pulling back from frontier models or AI usage internally? No. The pivot was more pragmatic: steer developers back toward GitHub Copilot CLI, tighten quotas, and put a more governable interface in front of demand. The lesson here is that if unmanaged Silicon Salaries had to be reigned in inside Microsoft, they likely will need to be more tightly governed and controlled in your company too.
Uber is yet another example of what budget blowouts look like when adoption and spend controls get detached from each other. By April 2026, Uber’s entire annual AI budget was gone, only four months into the year.
Leadership introduced Claude Code in December 2025 and then incentivized adoption through internal leaderboards ranking engineers by usage. By March, 84% of engineers were agentic coding users. Average monthly spend hit $150–$250 per engineer; heavy users reached $2,000. Token-based billing on recursive, multi-step agent workflows doesn’t behave like a software license. It behaves like an unmanaged headcount line.
The bitter irony: the tools delivered. Teams loved them! Nearly 70% of committed code came from AI systems, and 11% of live backend code updates were written entirely by AI agents.  CTO Praveen Neppalli Naga wasn’t managing a failed rollout. His own summary: “The budget I thought I would need is blown away already.” The productivity gains were real. The spend controls weren’t. That gap is the actual lesson.
Falling Behind or Exploding Costs
This is the real paradox now. If you pull back too hard on frontier tools, your best people fall behind peers at companies that are learning faster. AI skill formation matters. Tool fluency matters. In many teams, the engineers closest to these systems are already developing new instincts about architecture, testing, debugging, and product velocity. If you ban the tools outright, you create an identity threat for top talent and push experimentation underground.
But the other side is just as real. If you keep offering effectively unlimited access to the most expensive models, your margins evaporate. Finance loses forecasting discipline. Teams optimize for convenience instead of economic fit. And every workflow quietly drifts toward the most powerful model whether it needs it or not. That is how you end up with Uber-style budget blowouts.
So no, the answer is not to ban frontier tools. The answer is to end the “All-You-Can-Eat” era. Access needs tiers. Reasoning depth needs thresholds. Autonomy needs approval paths. And business units need visibility into what kinds of work are burning premium inference versus where a smaller model, or an actual person, is the smarter economic choice.
How LBZ Advisory Helps with Inference Governance
If your Board is asking why the AI “productivity gains” haven’t shown up on the bottom line, it’s likely because you’ve traded a predictable payroll for a volatile token-maxxing trap.
At LBZ Advisory, we help leadership teams build Inference Governance frameworks so they can keep the best tools in market without accepting Uber-style budget blowouts. In practice, that usually means three things:
- Set the inference budget at the workflow level: Don’t govern by vendor contract alone. Govern by task class, recursion depth, exception path, and business value. A bug triage flow, a contract review flow, and a board memo flow should not have the same spend permissions.
- Match model class to decision risk: We see teams using frontier reasoning models for routine internal work because nobody wants to be accused of “using the cheap model.” That’s organizational psychology, not strategy. We create routing rules so small models handle default volume and frontier models are reserved for moments where the extra reasoning actually changes the outcome.
- Put humans where margin and judgment matter most: The winning design is rarely “AI only” or “human only.” It’s a governed handoff model. Humans handle ambiguity, exception resolution, and accountability. Models handle acceleration. That keeps capability high without letting inference costs run wild.
The Return of the Specialist
The irony of 2026 is that the very tools meant to replace us have reminded us how “cheap” human intuition really is. We don’t require 1.2 gigawatts of power to remember a conversation from last week. We don’t charge per word. And we don’t hallucinate more when the prompt is slightly wordy.
As the “All-You-Can-Eat” era of AI fades, the CFO’s best friend is the skilled professional who knows how to get the job done without burning $2,000 in tokens before lunch.
We’ve spent three years obsessed with the capability of AI. 2026 is the year we need to get serious about the economics and the control plane around them. The “Silicon Salary” is real, and it’s rising. The hard part for boards and CEOs will be designing an operating model where teams can use the best frontier tools, keep talent sharp, and still protect margin. That is the job of Inference Governance. Done well, it lets a company avoid the false choice between falling behind and going broke.
Related reading
Board and C-Suite AI Advisory
Liat works directly with boards and executive teams on AI strategy, governance, and investment decisions.
