TL;DR: Reckless company spending used to look like private jets and $400 expensed sushi dinners at Nobu. Today, it’s a $113,000-a-month Claude invoice. AI bills at some companies are now running so high that human labor may actually be cheaper than the tech that was supposed to automate jobs.

What happened: It’s suddenly a bad time to be a CTO who went all in on AI. Last week, an Nvidia exec told Axios that the cost of compute for his team was “far beyond the costs of the employees.” Uber’s CTO told The Information last month that the company already burned through its full 2026 AI budget and is “back to the drawing board.” A recent Goldman Sachs research note put it bluntly: Companies are “overrunning their initial budgets for inference by orders of magnitude.” Some software companies are already spending roughly 10% of their total engineering labor costs on AI, according to Goldman. And that could soon match what these companies pay humans.

The big picture: Worldwide IT spending is expected to hit $6.31 trillion in 2026, up 13.5% year over year, per business and tech insights firm Gartner. Compute costs are ballooning as companies turn to agentic AI tasks that are more complex and expensive to run; AI labs have also raised prices or shrunk how much AI use a dollar can buy. It’s enough to make human workers look like a bargain.

Budgetmaxxing: Just a blink of an eye ago, sky-high AI bills were a flex. Some tech firms, like Meta, even had internal leaderboards where employees raced to see who could burn through the most tokens. Over 30 days, Meta workers reportedly ran through 60 trillion Claude tokens—which, based on a calculation by Fortune, means the highest-ranked user on the AI leaderboard could have spent over $1 million in a month.

Now, the AI bill is coming due. Yet plenty of tech companies are reallocating headcount costs toward compute anyway (with some, including Block and several Wall Street banks, explicitly saying it’s to replace workers). Meta and Microsoft alone announced over 20,000 job cuts in late April, both to focus more on AI investments—and additional layoffs are likely coming at Microsoft even after its quarterly revenue grew. (Meanwhile, here’s how the great AI replacement is going on the other side of the world.)

Bottom line: For months, every CEO with a microphone has been warning that AI is coming for your job. The math on that may not work out (and, unlike AI, human workers can actually be held accountable for their mistakes). It’s unclear when—or if—costs will fall enough for human replacement to become a smart corporate strategy. —WK

Your AI is a yes-machine. Here's how to fix that

Last week I kicked off the column talking about Projects—dedicated workspaces that give your AI the context it needs. Since then, several of you have written in with some version of the question: "Why is AI so cloyingly sycophantic?" (Read: It basically agrees with everything you say.)

The short answer: AI tools are trained to produce responses that feel satisfying, which means they default toward telling you what you want to hear rather than what you need to know. Here's what you can do to avoid this as much as possible:

Start with "critical but supportive." At the start of any chat where you want honest feedback—a draft, a plan, a decision you're weighing—add this to your prompt:

"Please be critical but supportive. I want balanced feedback, not validation."

That one phrase can shift the tone of everything that follows. I use it in almost every chat.

Give it the full picture upfront. AI will work with whatever you give it, even if that's not enough. The more specific you are from the start, the more useful the response. If I'm asking about a health question, I include relevant history and what I've already tried. For career coaching, I give it my actual situation—not just the sanitized version. Then add:

"Do you have any questions before we get started? Feel free to ask for more information as we go."

Use a second chat as a reality check. For higher-stakes situations (like health, mental health, or a big career decision), ask the AI to save everything in the chat into a markdown file (your prompts and its responses both). Then drop that into a fresh chat and ask:

"Assess this conversation. Point out any inconsistencies, gaps in reasoning, or perspectives that weren't considered."

A second chat reviewing the first's work will often catch angles the original one didn’t. This is how I get something closer to a genuinely different perspective.

One other tip: If you're not sure what a tool can actually do, ask it upfront—then ask it to confirm before you get started. It won't always volunteer its limitations, but it'll usually tell you if you ask. —SM

There’s yet another podcast to add to your morning commute, right next to The Daily and Huberman Lab: Amazon’s new AI-generated audio clips, where two fake hosts discuss a product sold on the site like it’s a radio show. You can even “call in” with your own burning Q’s about the item—perfect if you’ve ever wanted to hear yourself on QVC. (Listen to an example with diaper rash cream here.) Amazon told Business Insider that this “Hear the highlights” feature now appears on products that “typically require consideration before purchase”—which apparently includes No. 2 pencils and a $6 stapler.

While it’s useful to be able to ask how long it might take to cobble together a 59-inch mid-century modern buffet sideboard in yellow walnut, it’s very much unclear why Amazon went with this particular out-of-left-field format. It could be a nostalgic callback to the halcyon days when home shopping reigned supreme on our TVs—or just a nod to the mega-popularity of podcasts today. Whatever the case, we eagerly await the video update to the product podcast, soon to clog up a vertical social media feed near you. —WK

Chaos Brewing Meter: /5