Amazon named its AI budget this week. It did not account for the people it cut to pay for it.
On July 22, Amazon told some of the people who build its AI models their services were no longer needed. The cuts landed inside the Amazon AGI organization, the group behind the company's Nova models and the fine-tuning that makes them useful to a customer. Teams under two executives took the loss, Adeeb Shanaa in AGI Data Services and Vishal Sharma in AGI Information. Amazon would not say how many employees would no longer be receiving pay stubs, or what percent of the unit got their walking papers. Amazon referred to the layoff as "some roles." In the same news cycle Amazon boasted a different figure and without reservation stated that: it would spend about $200 billion in capital in 2026, most of it supporting AI and data-center buildout. It is telling that they are proud of the capital they spend on data centers yet coy about they people they put in the unemployment line the same day.
The $200 billion headline is the one with more virality. It led the coverage, moved the stock, and got defended on the earnings call, a figure so large that the main worry on a trading desk is whether AWS can grow fast enough to earn it back. The layoff rode underneath as a line about a fast-moving space and a sharpening of focus. You had to go looking to find that the sharpening landed on the researchers the company spent years and top dollar to recruit.
Here is the frame they handed reporters. "This is a fast-moving space, and we're sharpening our focus on the initiatives that matter most for customers," the statement read, "including eliminating some roles within parts of our AGI organization, even as we continue to invest in the areas most important to our customers' future." Read as intended, it is a story about agility, the sound of a giant getting leaner while it spends more on AI than almost any company in history. The word carrying all the weight is "some," which can cover any number at all.
Amazon is playing banker in Monopoly, everyone knows they are doing you dirty but it is hard to prove. Amazon moved the work of customizing and refining models out of its own building and onto its customers, who will now rent that compute on Bedrock and Trainium and pay by the hour. Research that used to sit on Amazon's payroll becomes a service billed to someone else, and the people who did it become a cost the company would rather push off its books. The AGI shop has been coming apart in slow motion for a year: Rohit Prasad, who built the Nova push, was out by the end of 2025; David Luan, who ran the San Francisco lab, gone in February. The investment continued to grow, and the headcount shrank because Amazon is all about cutting cost just like every public company in America. Once an opportunity provides itself, likely due to the workers hard work making a solid product, they will cut the labor and try and pass as much cost along to the consumer.
The fair objection deserves saying out loud, because the credibility is the point. This reads to some as reallocation, not replacement, and treating every cut as proof that a machine took a job is exactly the loose move we have warned against all year. Amazon runs one of the largest hiring operations on earth and still adds far more people than it lets go. Andy Jassy muddied the story last fall, when the company shed about 30,000 corporate roles and told staff the decision was "not really financially driven, and it's not even really AI-driven, not right now at least." Layoffs get blamed on AI faster than the technology itself advances, which tells you the label is doing public-relations work. Our claim was never that an AI walked in and fired these specific people. It is narrower and harder to duck: Amazon published the capital number and withheld the labor number in the same week, and only one of the two was hard to find.
Look twelve to eighteen months out and the story gets clear. More of the model-tuning work gets sold as metered compute while more in-house research teams thin under the cover of that word "some," and the capital line keeps getting announced to the dollar while the labor line keeps getting a shrug. The researchers cut this quarter will be told next year that they should have reskilled, by the same companies that just proved a top-dollar AI skill set is no shield when the spreadsheet turns. When even the people making the thing are a number a company will not say, the line about AI creating more jobs than it takes needs its asterisk printed in bold.
What to do this week asks no one's permission, and it runs on two tracks. First, push to make the count mandatory. A few states already force some version of it: Connecticut now requires employers to disclose when a layoff is tied to technology like AI, and Colorado's SB 26-189 moves the same direction. Tell your state representative you want a disclosure rule with teeth, so that when AI is named in a mass layoff, the notice has to say how many jobs it took, in a good-faith number, not the word "some." Second, follow the $200 billion to where it lands. Most of that money is data-center concrete, and those buildings rise in real counties, approved at zoning hearings almost nobody attends. Pull your county's next agenda and read it. If you are in Virginia, The Hum lists every data-center hearing on the calendar (https://thehum.cancelclankers.com/). The capital shows up as a building down the road long before it shows up as a headline.
Keep the second track human. The people cut this week were workers, some of the most skilled in the country, and they will need what any laid-off worker needs: a reference, a lead, a check that clears. If you know one, be that for them. The number Amazon would not print, every one of them already carries. They can tell you exactly how many desks went empty.
A company that can price the machine to the penny can count the people. It just decided you did not need the number.
— Stay Human ★
Get the Dispatch every Sunday.
Free. One issue a week. Forever — until you unsubscribe.