I was on a call when an engineering director at Uber muttered, “We already blew the 2026 AI budget.” You could hear the optimism leak out of the room. The tools were humming, but the product roadmap wasn’t catching fire.
I’m going to walk you through what that silence means for big bets on AI and for your instincts about automation. You know the headlines: Claude Code commits, token consumption, and a promise of exponential productivity. I’ve watched the promises hit balance sheets and, often, not deliver the promised lift.
Praveen Neppalli Naga’s team exceeded a year of AI spend in four months — now leaders are trimming hiring
Observation: Uber’s CTO said the company had already outspent its full 2026 AI budget inside the first four months of the year.
That confession matters because the follow-up was equally stark: executives chose to raise AI spending further and slow hiring to cover the cost. You can see the math in the meeting notes — more agent experiments, more tokens, fewer new heads. The AI rollout is a high-revving sports car idling in first gear.
Practical consequences show up where you expect them: teams claiming that “10% of code changes came from agents” (Claude Code among them) but struggle to point to the new features users actually notice. When engineering leaders are asked to draw a line from token consumption to shipped functionality, the line is often a hairline at best.
Are AI investments improving productivity?
Short answer: sometimes, but not uniformly. You’ve probably tried a tool that shaved time off a task and felt the ROI in your calendar. At scale, those micro-wins need to aggregate into features that keep customers and grow revenue. That aggregation isn’t automatic.
Executives assumed the speed and round-the-clock availability of models would translate directly into economic gains. Instead, you get new operational costs — token bills, platform integrations, and monitoring — that land on the ledger immediately while product returns arrive slowly, if at all.
A senior engineering leader saw Claude Code drive commits but couldn’t show shipped customer value
Observation: A senior engineer reported that agentic tools were producing a sizable fraction of commits, yet those commits didn’t clearly map to customer-facing functionality.
I’ve spoken with teams chasing internal leaderboards for AI usage. Nvidia CEO Jensen Huang publicly argued that a $500,000 (€460,000) engineer should consume “at least $250,000 (€230,000) worth of tokens” yearly — a provocative benchmark that turns token spend into a performance metric. The problem is the token bill doesn’t vanish into thin air; it’s paid by engineering time, product delays, or headcount reductions.
Token consumption has turned into a bonfire executives keep feeding, convinced it will thaw their margins. Meanwhile, product managers ask for evidence: which projects moved off the cutting-room floor because of these tools? If you can’t answer that, the tradeoffs look less like strategy and more like hope.
Who pays for AI token consumption?
It’s not a mysterious entity — it comes out of company budgets and, increasingly, from employee headcount and time. Some firms nudge usage with internal leaderboards; others reallocate hiring to fund the run-rate. The effect is the same: resources shift from ship-to-user work toward building and operating AI infrastructure.
Uber feared commerce agents would disintermediate its marketplace — that didn’t materialize (yet)
Observation: A year ago, boardroom scenarios at Uber warned that agentic commerce would redirect transactions away from platform apps.
Those fears were sensible. If a chatbot orders your lunch and pays with an embedded wallet, who needs the app? But Macdonald says that conversation hasn’t played out in practice: commerce through large models hasn’t yet disaggregated platforms at scale. You can follow the chatter — Perplexity, Amazon, Google, Microsoft experiments — and see pilots, but not the massive rerouting of commerce many feared.
That doesn’t mean the threat is gone. It means the timetable and the mechanism are uncertain. If agentic commerce does arrive, the companies with the right hooks into payment, logistics, and user trust will win. If it doesn’t, those same companies will face questions about why they funded vast infrastructure bets that didn’t produce immediate returns.
Will agentic AI replace platforms like Uber and DoorDash?
Not overnight. The technical pieces exist, but the commercial plumbing — trust, payments, logistics — is sticky. Platforms still control last-mile execution and user trust, and those are hard to displace. For now, the risk is headline-driven strategy that reallocates resources before product economics prove out.
You and I can draw a few blunt lessons: measure shipped value, not just internal usage; make token consumption visible to product KPIs; and treat early agent wins as experiments, not replacements. Companies that bank on token-fueled productivity without a clear path to customer-facing outcomes are making a bet on an uncertain timeline.
There’s one persistent, uncomfortable question that executives dodge: if the promises of immediate productivity gains are wrong, who absorbs the loss — employees, investors, or customers — and what does that mean for the next wave of AI spending?