Amazon Cuts 30,000 Jobs While Hiring Interns, Grads

Amazon Is Laying Off Staff and Hiring Interns and New Grads at the Same Time

AWS CEO Matt Garman has consistently pushed back against the idea of freezing entry-level hiring because of AI. In a June 24 interview with Platformer, he said Amazon plans to hire 11,000 interns and new graduates this year. That doesn’t mean all 11,000 will land full-time roles.

Since October 2025, Amazon has cut roughly 30,000 corporate jobs. CEO Andy Jassy has said that as AI drives efficiency gains, the company’s overall corporate headcount could keep shrinking. But that forecast isn’t proof that AI was responsible for all 30,000 of the cuts.

Layoffs and new hiring can happen at the same time, because they target different roles and different timelines. What caught my attention wasn’t the contrast between the two numbers—it was the AI-usage race reportedly playing out inside Amazon, where some employees allegedly padded their usage stats to climb internal rankings, a practice dubbed “tokenmaxxing”¹.

Usage Rankings Started Changing Employee Behavior

Back in May, the Financial Times reported something striking: inside Amazon, there was a leaderboard called KiroRank. It ranked employees by token²** consumption — essentially, how much they used Kiro, Amazon’s internal AI coding tool. Amazon had set a goal for more than 80% of its developers to use AI tools at least once a week.**

The FT reported that some employees were boosting their AI usage numbers with an internal agent³ tool called MeshClaw, which can automate things like code deployment and message processing. There’s nothing inherently wrong with that kind of automation. The problem arises when the reason for using a tool becomes “how much it helps my ranking” rather than “whether the task actually needs it.”

Employees interviewed by the FT described feeling pressure to use AI tools, along with a sense that managers were informally tracking usage. Amazon has stated that usage does not factor into performance reviews. Security concerns were also raised about the system access and execution permissions granted to MeshClaw. When companies push usage before working out actual performance metrics and permission boundaries, cost and governance problems tend to follow. We saw this same tension play out in the Uber case discussed earlier — how to manage the cost of AI usage was central there too.

Told to Spend Half Your Salary on Tokens… How Did That Go?Uber’s AI budget spending and cost management problemsinlevel9.com

At Meta, too, an employee-built token usage leaderboard called Claudeonomics reportedly ran for a while before being shut down. It handed out titles to top users. Just knowing the reported total token volume doesn’t let you calculate what the company actually paid — per-model pricing, input/output ratios, caching, and contract terms all differ.

At Microsoft, President Julia Liuson sent an internal memo saying “using AI is no longer optional — it’s core to every role at every level.” Nvidia’s Jensen Huang went even further, saying publicly: “If I have an engineer making $500,000, and they’re not spending $250,000 in tokens, I’d be deeply worried.”

Encouraging AI adoption and treating token consumption as a performance metric are two different things. The fact that multiple companies have made statements encouraging usage doesn’t mean they’re all running identical leaderboards — or that they’ve all admitted the same failure.

If Usage Went Up, Did Results Go Up Too

In the interview, Garman explained that when you build the wrong metric, people end up focused on hitting that metric rather than on the actual goal it was supposed to represent. A goal that maximizes AI usage isn’t the same as a goal that produces good outcomes.

So what should companies actually look at to verify whether AI adoption and investment are paying off?

Combined 2026 capital expenditures (CAPEX)⁴ for Amazon, Alphabet, Meta, and Microsoft are estimated at roughly $700 billion. AI infrastructure is the main driver of this spending increase, but that doesn’t mean a company’s entire capex can be classified as “AI spending.” For a large investment to pay off, you need services customers keep using and actual revenue to show for it.

Internal token consumption shows that employees have access to the tools. That number alone can’t explain whether work actually improved, or whether there’s real demand from paying external customers. Concluding that the tokenmaxxing incident inflated overall infrastructure demand across the board would be the same kind of error.

Garman said that when he asked a room of roughly 100 CIOs whether they’d already seen returns on their AI investment, or had a clear path to seeing returns within a few months, about 90% raised their hands. This isn’t a representative survey of companies, nor is it a statistic showing 90% have realized returns. He also cited examples like reduced processing time for insurance claims and improved success rates for AI agents on certain tasks.

Amazon’s own recruiting tool, Amazon Connect Talent, automates tasks like scheduling and voice interviews. Garman explained the intent was to free up recruiters to focus on finding candidates and building relationships with them. Automating part of the hiring process is a different decision from not hiring new employees at all.

Garman also said many companies are shutting down PoCs⁵ that haven’t produced results. There’s a gap between the fact that a tool was adopted and the fact that it generated profit in actual operations. You can only see that gap by looking at usage, cost, and outcomes together.

Goodhart’s Law⁶ is often invoked to describe this kind of situation. It holds that when a measured metric becomes a reward or a target, people act to raise that metric — which can pull them further away from the actual outcome the metric was meant to track. Tokenmaxxing shows this same problem can surface in the process of AI adoption.

New hires also need real chances to learn on the job

Garman pointed to the earlier example of computers entering the workplace: even when new technology changes how work gets done, people can still learn to work alongside new tools. I think that to properly evaluate that optimism, we also need to ask whether new hires are actually getting the chance to learn.

A Stanford research team looked at ADP payroll data and observed that employment among 22–25 year-olds in occupations with high AI exposure has been relatively weak. In a February 2026 explainer, the researchers described a relative gap of about 16%, based on data through October 2025. That doesn’t mean overall youth employment fell by 16%, nor does it establish that AI caused all of that decline. But it’s a good enough reason to keep watching how opportunities are changing for early-career workers.

In Korea, the job platform Catch tallied its own listings and found that postings for full-time entry-level positions at large companies fell 43% year-over-year in 2025, and 67% in IT and telecom specifically. Job postings aren’t the same as actual hires, and other factors — the broader economy, corporate investment levels — also play a role. These numbers alone can’t tell us that AI is the cause behind Korea’s shrinking entry-level hiring.

Still, there’s a training problem that companies need to solve. Junior developers learn how systems work by writing basic code; junior analysts often pick up business context by cleaning data. If these tasks get automated, companies need to redesign what work new hires do to build foundational skills in the first place. Keep cutting hiring and training, and a few years down the road, it may become hard to find experienced talent in-house.

IBM said it plans to triple entry-level hiring in the US in 2026, while reshaping what those jobs actually involve for the AI era. Cognizant, meanwhile, said it hired about 20,000 new graduates in 2025. Even as automation expands, how companies choose to develop talent still shapes their hiring decisions.

The willingness to learn that Garman emphasized matters. But that willingness only turns into real capability if companies also give people actual work to do and real feedback on it.

Oswarld’s Lens

Having built go-to-market and technology strategies for companies, I’ve seen again and again that what you measure changes how employees behave. If you make subscriber count the only goal, for instance, usage and return visits after sign-up can get neglected. The same goes for code commits or token consumption. Metrics that count activity can be useful, but treating them as performance in themselves can steer people toward the wrong behaviors.

When I listen to a vendor’s pitch, I also weigh their business interests. AWS is an infrastructure provider whose revenue grows as customers’ AI usage grows. That doesn’t mean Garman’s forecast is wrong. It means the success stories a vendor presents need to be separately verified against your own company’s costs and results.

When evaluating AI adoption, it helps to first define the business outcomes you’re measuring — processing time, error rates, customer satisfaction, revenue. If a case study claims it cut insurance claim processing time, for example, check exactly how many days were actually saved, and what happened to error rates and rework costs. Only by looking at those outcomes alongside the cost of use can you make a real investment decision.

Closing

It’s hard to judge AI’s employment impact from Amazon’s layoff numbers and intern/new-hire counts alone. We need to look at which jobs are shrinking and which new roles are emerging. The same goes for AI adoption inside companies — it’s not enough to note that usage has gone up. We need to check how much work has actually improved and what employees have learned along the way.

Have you ever measured AI usage or adoption at your company? Tell me in the comments how usage compared to actual results.

References & Further Reading

Primary sources

• Casey Newton, “The CEO of AWS on why Amazon is hiring 11,000 interns and junior employees”, Platformer, 2026.6.24. : This piece includes AWS CEO Garman’s own remarks about his “spreadsheet analogy” and the token leaderboard.
• Financial Times, “Amazon employees inflate AI usage on internal leaderboard”, 2026.5.11. : The article that first reported the tokenmaxxing phenomenon, including interviews with Amazon employees.
• The Information, “Meta’s Claudeonomics leaderboard”, 2026.4.7. ··· Reports on the shutdown of Meta’s internal token-usage ranking board.

Background

• Stanford researchers explain their analysis methods and timing, 2026.02.09.

• IBM on expanding entry-level hiring in the US and redesigning work, 2026.02.

• Andy Jassy on generative AI and the corporate workforce outlook, 2025.06.17.

• Erik Brynjolfsson, Bharat Chandar, Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, Stanford Digital Economy Lab, 2025. : An empirical paper analyzing declining employment among early-career workers aged 22 to 25. Read it carefully, distinguishing the analysis period from the comparison groups used.

• Stanford HAI, “AI Index Report 2026”, 2026.4. : An annual report synthesizing research on AI adoption and its economic and employment effects.

• Catch (a Korean recruitment platform), “Analysis of 2025 Large-Corporation Entry-Level Hiring Postings”, 2025.12. : The source of the data showing a 67% plunge in entry-level IT/telecom hiring at major Korean corporations.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Tokenmaxxing: The practice of artificially inflating AI tool usage to climb an internal leaderboard. A blend of “token” (the unit of data an AI processes) and “maxxing” (maximizing), this phenomenon surfaced simultaneously across several Big Tech companies in 2026. ↩

2. Token: The smallest unit of data an AI model uses to process text. How characters map to tokens varies by language and model. API pricing typically applies different rates for input and output tokens. ↩

3. Agent: AI software that uses tools to perform tasks within a defined goal and set of permissions. Unlike a simple chatbot, it can autonomously handle practical work such as deploying code, processing emails, or managing schedules. ↩

4. CAPEX (Capital Expenditure): Spending by a company on long-term assets such as data centers, servers, and real estate. Its composition varies by company; AI-related CAPEX includes items like GPU servers and power infrastructure. ↩

5. PoC (Proof of Concept): A small-scale experiment to verify whether a new technology or idea actually works. If it succeeds, it moves to full-scale adoption; if it fails, it’s discontinued. ↩

6. Goodhart’s Law: A principle proposed by British economist Charles Goodhart, meaning “when a measure becomes a target, it ceases to be a good measure.” It applies here to explain how efforts to boost token-usage rankings can drift away from actual productivity. ↩