In 2025, tech companies laid off 150,000 people. The stated reason, more often than not, was AI efficiency. Companies said they could do more work with fewer people because their AI tools had gotten good enough. This is the public story: the machines are coming for the jobs, and the machines are already here.
The private story is different. Inside those same companies, AI spending went up 400% while revenue per employee dropped. ML engineers were laid off while their companies posted new openings for ML engineers. The contradiction is not subtle, and it is not a bug in the numbers: the layoffs were never a cost-cutting exercise that happened to succeed. They were a narrative signal sent to capital markets.
Think about who got cut first. ML engineers operated as something like a guild inside tech companies: special budgets, opaque roadmaps, metrics that nobody outside the team could audit. They commanded high salaries, flew under normal performance review processes, and made promises about future capability that finance could not independently verify. The AI efficiency narrative gave CFOs a weapon to reassert control over a department that had escaped normal business discipline. This was never about whether the models worked. It was about who got to decide what working meant.
And it worked beautifully. Companies that laid off workers citing AI efficiency saw their stock prices rise. The signal reached investors: we are serious about AI, serious enough to sacrifice our own people for it. The actual operational reality, whether the remaining team could still ship, mattered less than the story the layoff told about corporate commitment. Signal over substance, every time.
Then came the contractors. A quiet pattern emerged: companies fired full-time ML engineers, then rehired the same people as contractors at higher rates. Not because the models could replace them, but because only those engineers understood why the models worked. The undocumented 2am intuition that made a particular experiment converge, the hunch that a learning rate schedule was wrong even though the loss curve looked fine, the memory of a failed approach that nobody had written down, all of this left when the employees left. Six months after a layoff, teams found their models drifting silently. Logs do not capture the hmmm weird moment.
The contractor pipeline reveals the real economic transaction. The layoff was not about eliminating work. It was about category shifting: turning employees into expenses. Employees cost headcount on a balance sheet. Contractors cost an operational line item. The same person doing the same work for more money, but the optics for investors are completely different. The company can say they automated the role while quietly paying the same person to do it through a different channel.
This creates a strange dynamic for the survivors. The people who keep their jobs are the ones whose proximity to the model is direct enough to be framed as indispensable. The data scientists, the prompt engineers, the fine-tuning specialists, they watch their friends get cut while the justification for cutting those friends is the very technology they maintain. It creates a caste system within tech: the AI-adjacent insiders who are close enough to the magic to survive, and everyone else who is fungible. The survivor dynamic is not comfortable. You do not sleep well when you know the story that saved you is the same one that took your teammate.
There is a darker layer underneath all of this. The ML engineers being laid off are the same people who fine-tuned the models that now generate the reports used to justify their own removal. They built the knife. They sharpened it. They handed it to management and said here, this will save us money. They did not realize they were writing their own termination notice.
This is the recursive part. The industry sells a story about automation to clients, uses that story to justify layoffs, lays off the people who could build actual automation, uses the savings to buy AI tools from external vendors who are also laying off their builders, and the circle tightens. The people building AI are being automated out by the idea of AI, not by AI itself. The concept of the replacement precedes the replacement. The story does the work before the technology can.
An infrastructure team at a major AI company spent two years building a training pipeline for a flagship model. The pipeline was excellent: automated data preprocessing, failure recovery, allocation of compute during training runs. It worked so well that the team was laid off because the pipeline they built automated their own operational role. They were fired for succeeding. The success condition for their project was the evidence used to eliminate their jobs. This is not a story about innovation. It is a story about what happens when the tool becomes the justification for removing the toolmaker.
The numbers do not add up because they were never meant to. 150,000 people laid off citing AI efficiency. AI spending up 400%. Revenue per employee down. If the layoff was a genuine efficiency play, the spending and revenue data should show some relationship to headcount. They do not. The layoff was a narrative for capital markets, a power struggle between engineering guilds and finance departments, a category shift from full-time headcount to contract labor, and a signaling game where companies compete to show who is most serious about AI by sacrificing the people who actually understand it.
None of this means AI is not changing work. It is. But the layoff wave of 2025 was not driven by actual automation displacing actual workers. It was driven by a story that companies needed to tell, and the people who got cut were the ones who could not control that story. The technologists lost the narrative war before they lost their jobs. The models they built were used as evidence in a trial they did not know they were attending.