Starbucks Kills NomadGo; Hyperscaler AI ROI Turns Negative
Starbucks shut down its NomadGo AI inventory system nationwide in May 2026—nine months after a full North American rollout—as AI critics and financial analysts converge on the same thesis: enterprise AI is failing at deployment scale. Three independently reported signals form a coherent counter-narrative to the capex supercycle.
What the Source Actually Says
NomadGo was deployed to every North American Starbucks location in September 2025, positioned as a $2.5 billion cost-savings initiative through automated inventory counting. The system instead hallucinated objects, misidentified labels, and failed to count accurately—creating more work for frontline staff and triggering shortages. Management praised the system as recently as February 2026 while employees reported mounting failures. By May 2026 the company shut it down and reinstated manual counting.
Gary Marcus, amplifying a report via @HedgieMarkets, argues Starbucks is not an isolated case. The same failure script—clean-environment demos, messy-input production failures, vendors paid regardless—is playing out across Picnic ($53M raised, Domino's partner, shut down), Zume Pizza ($500M burned), Waymo (pausing eight cities), Microsoft (reportedly ending internal Claude Code use), and Uber (exhausting its 2026 AI budget in four months). Marcus: "The vendor gets paid through the failure cycle. The buyer eats the cost and quietly retires the product."
The macro picture adds urgency. Financial Times projections, flagged by Marcus, show AI ROI under best-case assumptions has turned negative at four of five hyperscalers: Microsoft −9%, Google −15%, Meta −28%, Oracle −35%, with Amazon barely positive. A parallel signal comes from enterprise buyers: analyst @hkarthik reports that "tokens got burned for millions of dollars without any real significant ROI"—and that the largest OpenAI and Anthropic accounts are now being assigned field engineers to prevent churn as first-year contracts lapse.
Strategic Take
The deployment gap is AI's structural problem: demos run on clean inputs; production fails on messy ones. For builders and advisors, this is the moment to contractualize AI deployments around outcome metrics, not capability claims. Buyers who tied vendor payment to measurable ROI are ahead; those who paid on deployment are now absorbing the Starbucks lesson.
