The Army runs out of AI budget, Pakistan's courts show what makes AI pay, and power bills climb

The US Army opened an AI workspace to its staff in 2025, and it went from almost no users to about 19,000 in under 45 days. This month the Army told those users to cut back, because the tokens it had paid for were running out far ahead of schedule.

The service, called Ask Sage, runs on a five-year contract worth $49 million, and staff draw on a yearly allotment of tokens across models from OpenAI, Google, and others. The budget broke because of how these tools are billed. You pay by the token, so the bill grows with how much work the software does, and heavy users can run up far more than a per-seat plan assumes.

A simple question runs one pass through a model. An agent handed a whole task breaks it into steps and can run twenty passes for a single request, and every pass adds to the total. A promise of open access, then, is a promise to pay whatever the usage turns out to be.

The Army is not the first. Uber, Meta, Microsoft, and Amazon have each run through their AI budgets faster than they planned, some within months of giving staff wide access.

For every dollar it spent, Pakistan's court system got back about $38 in value, by the measure of a large field experiment run with the country's judiciary.

Researchers led by Elliott Ash built a custom assistant, JudgeGPT, for the trial courts and divided 1,559 judges across 118 courts into three groups. One got the tool along with training built around how to use it. One got the tool with only general training on technology and law. One got the general training and no tool. The judges given the tool and the training built around it adopted it more, kept using it over time, and their districts went on to resolve more cases. The judges who got the tool with only general training used it less and gained less.

At a typical district, pairing the tool with that focused training corresponded to 1,848 additional cases resolved in a year, in a system carrying more than two million pending cases.

The models themselves keep getting cheaper to run. The electricity behind them does not.

US data centers will draw about a fifth of the country's electricity by 2035, up from under 6% today, BloombergNEF said this week, roughly four times their current use. Nearly half of that capacity goes to training and running AI.

These centers connect to the same grid everyone else uses. When a utility upgrades that grid to feed a new one, it spreads the cost across every customer on those lines. In states where the centers cluster, such as Virginia and Texas, the share of local power they take is higher still.

For a manufacturer, a hospital, or a cold-storage warehouse in one of those regions, the industrial power rate climbs whether or not the business ever runs a single AI query.

In an Albertsons warehouse, a program now inspects each pallet of fresh strawberries and tells a worker whether it meets the grocer's quality standard. It is not the kind of AI that makes headlines. The same company also launched a shopping chatbot that turns a request for something spicy into a full basket of products. Industry analysts say the strawberry checker is worth far more today.

The returns in grocery are turning up in the back rooms: forecasting fresh orders, grading produce, setting how much to reorder. Hy-Vee is using AI to sharpen how it forecasts and replenishes fresh product, and Grocery Outlet is bringing in a system to tighten ordering across departments.

Take produce ordering. A model can read years of sales by store and season, the weather, and current stock, then place an order that leaves less on the shelf to spoil. The buyer no longer works from a best guess. What holds most grocers back is the data underneath, often messy and spread across old systems.

Most companies buying a robot for a warehouse or hospital floor never see the security flaws that ship inside it. The machines arrive able to move, sense, and decide on their own, and many carry known weaknesses in their control software that the buyer is never told about.

The World Economic Forum flagged this in its 2026 cybersecurity outlook. As order-picking robots and port machines change from simple equipment into systems that learn and adapt, their behavior gets harder to predict, and someone who compromises the control software can change what a machine does to people and goods within seconds.

Responsibility for that harm usually falls on the operator, not the manufacturer. The EU's AI Act begins applying formal duties to high-risk systems, robots among them, in August 2026.

In June, the US government reached into a private company's product and switched it off. Citing export-control powers, the Commerce Department ordered Anthropic to block all foreign nationals from its two newest models, Fable 5 and Mythos 5. Unable to verify each user's nationality in real time, Anthropic took both offline for every customer worldwide.

Most vendor contracts do not cover that kind of interruption. A business that kept a second model or an earlier version ready could switch over; one that relied on a single top-tier model waited out the full shutdown, which ran 19 days.