Uber used up its 2026 artificial intelligence budget in four months, which led it to restrict access and give developers priority over AI services for customers on its app, Goldman Sachs said in a report, according to The Economic Times on 19 October 2026. The bank calls Uber’s response “strict gatekeeping on LLM usage”.
The report, which ET says is titled ‘Connecting the Dots’, collects examples of companies limiting what they spend on AI tokens. “Many enterprises are currently capping AI spend, given elevated AI costs,” it notes, with examples from technology, retail, telecom and financial services.
Business Pill · COST PER TASK
A short explainer on cost per task: paying for the job, not the word. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Goldman’s list shows AI tokens moving from an open tap to a managed budget: Uber rationed access after using its 2026 AI budget in four months, and Amazon and Walmart moved to team allowances.
According to ET, the report describes three ways companies are controlling AI bills: setting usage limits, removing tools that increase consumption, and moving to smaller AI models.
Amazon shut down an internal tool that encouraged token consumption and moved to allowances for individual teams, the report says, while Walmart moved away from “unlimited token usage” and introduced team allocations.
Meta shut down an internal leaderboard that tracked AI usage because it encouraged “tokenmaxxing”, according to the report, which links the leaderboard to 60 billion tokens used in 30 days.
Microsoft removed access for some employees after high token costs per engineer and moved its GitHub Copilot coding assistant to credit-based pricing, ET reports from the bank’s compilation.
Uber used up its 2026 AI budget in four months, according to Goldman Sachs’ ‘Connecting the Dots’ report; Amazon and Walmart moved to team token allowances and Microsoft moved GitHub Copilot to credit-based pricing (Goldman Sachs report, as reported by The Economic Times, 19 October 2026).
What Was Posted on X
Cognition is taking aim at “tokenmaxxing” with a system that grades engineers on their AI efficiency and adjusts their spending limits based on performance.
The Information reported on 17 October that Cognition is taking aim at “tokenmaxxing” with a system that grades engineers on their AI efficiency and adjusts their spending limits based on performance. As we read it, that is a vendor building the same kind of usage limit that Goldman’s report lists at its customers.
Why the Bills Are Hard to Predict
ET notes that much of this spending comes down to tokens, the units of text that AI models read and produce, which providers commonly use to calculate charges. Even small charges can add up when employees use the tools throughout the day, ET notes.
The report cites Cognizant’s warning that 85% of enterprises were missing their AI spending forecasts. In its summary on Capgemini, it says: “Cost of tokens can balloon.” Capgemini is focusing on smaller models built for specific types of work, according to the report.
Smaller Models, Lower Cost per Task
AT&T, which the report puts at eight billion tokens of average daily usage, changed the software that sends work to AI systems so it could use “smaller agents depending on the task”. That change cut costs by 90%, according to the report.
Other companies are judging spend by results, ET reports from the bank’s summary. Zopa flagged unpredictable token costs and is directing its AI budget towards improving resolution rates. Starbucks is building its own AI tools to reduce its software budget, and Spotify is weighing the cost of employees against spending more on AI tokens.
AT&T changed the software that directs work to AI systems to use smaller agents depending on the task, a change that cut costs by 90%; the report also cites Cognizant’s warning that 85% of enterprises were missing AI spending forecasts (Goldman Sachs report, as reported by The Economic Times, 19 October 2026).
The Structural Reading
As we read it, the report shows AI usage becoming a budget line that companies ration: team allowances at Amazon and Walmart, credits at GitHub Copilot, and priority rules at Uber.
Our earlier reporting covered the same pattern at two other companies: Accenture capping Claude funding at $29 a month per employee, and Santander cutting serving costs per interaction in experiments by moving to smaller specialised models. As we read it, Goldman’s list suggests those were not isolated cases.
The Structural Read
As we read it, the examples ET reports fall into two groups: limits on how much staff can use (Uber, Amazon, Walmart, Microsoft, Meta’s leaderboard) and changes to what each task costs (AT&T’s smaller agents, Capgemini’s smaller models).
As we read it, the report pairs usage limits with Cognizant’s warning that 85% of enterprises were missing their AI spending forecasts.
As we read it, AT&T’s 90% cut, as the report gives it, comes from routing each task to a smaller agent, the same lever Santander described when it moved to smaller specialised models.
Goldman Sachs, ‘Connecting the Dots’ report, as reported by The Economic Times, 19 October 2026
“Many enterprises are currently capping AI spend, given elevated AI costs,”
Three Implications
MODEL VENDORS The report lists companies moving to smaller AI models designed for a narrower range of work, including Capgemini and AT&T, according to ET.
CODING-ASSISTANT PRICINGMicrosoft removed access for some employees after facing high token costs per engineer and shifted GitHub Copilot to credit-based pricing, the bank’s compilation shows.
FINANCE TEAMS The report cites Cognizant’s warning that 85% of enterprises were missing their AI spending forecasts, and Zopa directing its AI budget towards resolution rates.
The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, the companies in Goldman’s report are the buyers on the other side of that engine, now budgeting the marginal cost of every query. The framework was published in March 2026 and does not discuss these companies.
What Is Not Established
We did not read the Goldman Sachs report itself; every figure and example here is as The Economic Times reports it. ET does not give the report’s date, its authors, or how the company examples were gathered. ET’s article carries no comment from Uber, Amazon, Walmart, Meta or Microsoft.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Goldman Sachs says in a report, as The Economic Times reported on 19 October 2026, that many enterprises are capping AI spend because of elevated costs. Its examples include Uber using its 2026 AI budget in four months, Amazon and Walmart moving to team token allowances, and AT&T cutting costs by 90% with smaller task-specific agents.
A note on sourcing. This piece is based on The Economic Times’ report of 19 October 2026 on a Goldman Sachs report titled ‘Connecting the Dots’, read in full. We did not read the Goldman Sachs report itself, so every company example and figure is attributed to Goldman as ET reports it. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
