AI cost reality bites: Uber, Starbucks, and the enterprise ROI reckoning

Uber burned through its entire 2026 AI budget within four months, spending it entirely on Anthropic's Claude Code. That single fact, surfaced by Quartz in late May, set off what the company's president and COO Andrew Macdonald described as "company-wide conversations" about whether the cost of AI consumption can be squared against what it displaces, including headcount.

Around the same time, Starbucks quietly pulled an AI-powered computer-vision inventory management system it had deployed less than nine months earlier. Restaurant Dive reported that employees described the system as unreliable, citing persistent miscounts and mislabeled products. The tool had been positioned as a cornerstone of CEO Brian Niccol's strategy to fix the chain's product availability problems. It did not survive contact with daily operations.

The ROI gap is now a board-level conversation

Macdonald's candor is notable precisely because of his seniority. Speaking about the Claude Code spend, he told Quartz that the link between higher token consumption and measurable customer experience improvements is simply "not there yet." He framed the core tension plainly: enterprises will need to start treating token consumption as a line item to weigh against headcount, not a separate innovation budget insulated from scrutiny.

That tension is not unique to Uber. Techerati reported on a Gartner forecast projecting that by 2028, the cost of AI coding assistants will become a significant budget concern for enterprise technology leaders broadly. The analyst firm flagged that as adoption of generative AI coding tools scales across development teams, the cumulative token and licensing costs will demand the same financial governance applied to any other enterprise software category.

The Gartner projection matters for CIOs and IT procurement teams evaluating multi-year software agreements today. A tool that looks affordable at pilot scale can look very different when it is running continuously across dozens or hundreds of developers, each generating high token volumes.

Starbucks' nine-month lesson in parallel testing

The Starbucks case carries a distinct operational warning. The inventory system, announced in September 2025 as part of a broader push to improve in-store accuracy, was integrated into live operations rather than run alongside existing manual processes long enough to validate reliability. According to Restaurant Dive, employees flagged inaccuracies repeatedly, and the company eventually pulled promotional materials tied to the system before discontinuing it entirely.

Computer-vision inventory tools are not inherently unreliable. But the Starbucks outcome illustrates what happens when a high-visibility operational system skips the parallel-run phase that would normally expose edge cases before they affect store-level decisions. For operations leaders in retail, food service, and any sector where inventory accuracy directly affects customer experience, the case is a concrete data point, not an abstraction.

A pattern forming across enterprise AI

Forbes contributor Gene Marks, writing about the Uber and Starbucks developments, noted that both cases reflect a wider pattern: large enterprises are discovering that deploying AI at scale introduces cost structures and reliability variables that were not visible during the evaluation phase. The observation is backed by the specifics. Uber's four-month budget burnout was not caused by reckless spending; it was caused by legitimate developer adoption of a tool the company had approved and deployed.

For enterprise buyers, the implication is structural. AI tool evaluations need to model consumption at full team scale, not pilot scale. Operational AI deployments need a defined parallel-testing window before live cutover. And ROI definitions need to be agreed upon before contracts are signed, not after budgets are exhausted.

OpenAI, meanwhile, is pressing forward with its own infrastructure ambitions. Techerati reported that the company has unveiled a custom inference chip called Jalapeño, developed with Broadcom, which it expects to deploy with Microsoft and other infrastructure partners through 2026. For enterprise operators, that development is relevant context: the underlying cost of AI inference is something vendors are actively working to reduce. Whether those economics flow to customers through lower pricing or stay as margin will be a critical variable in 2027 and 2028 renewal cycles.

What this means for your team

• Model token consumption at full deployment scale before signing AI coding assistant or operational AI contracts. Pilot-scale economics are not predictive of production costs.
• Require a parallel-run phase of at least 90 days for any AI system touching inventory, fulfillment, or customer-facing accuracy before live cutover.
• Build explicit ROI thresholds and a defined review timeline into AI vendor agreements. If the link between usage and business outcome cannot be measured, the spend cannot be defended.
• Watch inference cost trends closely. Custom silicon investments by major AI vendors may shift pricing in 2027-2028; factor that uncertainty into multi-year budget models.