Uber burned through its entire annual AI coding tools budget in four months. Not over a year of runaway adoption — four months. That number should be getting far more attention than it is.
Companies that told their developers to use AI as much as possible are now introducing spending tiers and usage caps, and slowing new commitments. Enterprises are explicitly switching models mid-stream. One AI startup, Lindy, moved 100% of its API traffic from Claude to DeepSeek in a single decision driven by cost.
CNBC framed it plainly this week: users are shifting from "tokenmaxxing" to efficiency. The era of throwing money at AI to see what sticks is ending. The era of justifying what you spent has arrived.
Your CFO wants receipts. Most engineering orgs don't have them.
What "Efficiency Mode" Actually Means
This isn't about companies losing faith in AI. It's about the bill coming due.
The pattern is consistent: enterprises deployed AI broadly in 2024 and 2025, often without tight attribution between spend and output. Now the budgets are hitting real constraints, the easy wins have been captured, and leadership is asking the question that was always coming. What did we actually get for this? Nobody loves that conversation.
If your answer involves vague improvements in developer experience, qualitative feedback from your team, or the general sense that things are "faster," you are not going to like Q3 planning.
The companies navigating this well are doing one specific thing differently. They're treating AI spend as a capital allocation question, not an IT line item. That means they can tell you, with specificity: this workflow costs X per month in API calls, it produces Y output, we've validated the quality against this standard, and the ROI math looks like Z. That's a defensible position. "We gave everyone Copilot and it seemed helpful" is not.
Uber's president and COO Andrew Macdonald put the test plainly:
"If you're not actually able to draw a direct line to how [many] useful features and functionality you're shipping to your users, that trade becomes harder to justify."
— Andrew Macdonald, President and COO, Uber, Fortune
That's the standard. Most companies can't meet it yet. Nicholas Arcolano, head of research at Jellyfish, said the quiet part out loud:
"Whether extreme spend pays off comes down to the ultimate business value of shipped code (e.g. revenue), which most companies still can't measure."
— Nicholas Arcolano, head of research at Jellyfish, TechCrunch
The Model Routing Problem Is a People Problem
Here's what the smartest technical leaders figured out before the CFO pressure arrived: not every task should run on a frontier model.
Complex reasoning, architecture review, novel code generation, high-stakes analysis — yes, frontier. Summarization, classification, routing, extraction, first-draft generation — no. A cheaper tier handles that work just fine, often indistinguishably so, at a fraction of the cost. Think Claude Haiku, GPT-4o mini, an open model like Llama.
The technical architecture for model routing isn't hard. You can build it. The problem is that building it correctly requires someone who understands both the business workflow and the model landscape well enough to make those routing decisions systematically, audit them over time, and update them when the model options change.
That's not a job description that existed two years ago. It barely exists on job boards now. Search the title and you'll get a grab bag of ML reposts and DevOps roles wearing a new hat. But it's the role that pays for itself fastest in an efficiency-obsessed environment.
What you're looking for is someone who sits at the intersection of systems architecture, AI tooling, and business process — not a pure ML engineer, not a solution architect, not a product manager. The closest existing title is probably "AI Platform Engineer," but even that misses the business judgment component.
The Two-Track Labor Market Is Here
PwC published their 2026 AI Jobs Barometer this month, and there's one finding that stands out.
They split AI's labor market impact into two tracks. Jobs being "professionalised" — where AI automates the routine work so human judgment takes center stage — are seeing twice the growth in available jobs as roles being "democratised" by AI. Professionalised roles are also seeing 42% faster salary growth.
Engineering falls in the professionalised bucket. And separate data backs it up: SignalFire's 2025 State of Talent Report found that at the major tech companies and startups it tracks, engineers now represent 55% of new tech hires — up from 46% in 2019. The fear that AI would eliminate engineering jobs hasn't materialized. What's actually happening is closer to the opposite: the value of senior technical judgment went up because AI made the execution layer cheaper.
The roles under pressure are the ones that mainly execute against a defined process. The thing is, if most of what someone does can be described in a prompt, their value went up. That's true at every level of the org.
The practical implication for staffing is this: in an efficiency-focused environment, you don't need fewer engineers. You need different engineers — people who can define the constraints, validate the outputs, and make the judgment calls the models can't.
What to Do Now
If you haven't already mapped your AI spend by workflow, start there. Not in aggregate — by workflow. What does each deployment cost, what does it produce, and what would it cost to produce that output without AI? You need that math before someone above you asks for it under deadline pressure.
Then audit your model usage. Are you running frontier models on tasks that don't need them? A surprisingly large fraction of enterprise AI spend goes to processing work that a smaller, cheaper, faster model handles identically. That's not a product problem — it's a decisions problem, and fixing it requires someone who can own those decisions.
Finally, re-examine your hiring profile. The efficiency reckoning rewards technical leaders who can operate with financial accountability — people who know what their infrastructure costs, can optimize it, and can speak to ROI in CFO language. If that's not a lens you're hiring against, it's time to update the scorecard.
The AI spending boom isn't over. But the free-money phase is. The organizations that come out of this reckoning ahead will be the ones that treated the accountability question as an engineering problem, not an obstacle. Start keeping receipts.
VC5 Consulting specializes in placing senior engineering and technical leadership talent at companies navigating AI-driven transformation. If you're building the team that can actually show AI ROI — not just deploy the tools — let's talk.