AI doubles layoffs: the split nobody disputes

54,836 AI-related layoffs in 2025: the dystopia now has a date

Sinan Can Demir was twenty-four years old, studying computer science at the University of Texas at Dallas, and had racked up more than twenty rejected internship applications. He spent the last week of July polishing his résumé, as half the world would: digging through open GitHub projects for something to fix and show off with. He found a proposed change inside a network scanning tool, with a malware downloader hidden behind an impeccably clean bug fix. He alerted the maintainer. The episode sums up the pattern: it is no longer clear who writes, who signs off, and who checks.

The layoffs companies are already attributing to AI

According to Challenger's records, US companies attributed 54,836 job cuts to artificial intelligence throughout 2025. In the first half of 2026, the reported figure stands at 101,743: nearly double in half the time. For four consecutive months, AI was the most frequently cited reason when announcing layoffs.

The nuance matters and should not be brushed aside: Challenger collects corporate announcements and the reasons companies themselves give, and does not audit whether an algorithm actually replaced each position. It measures what companies are willing to declare, not actual replacement. Even so, when the leading cause of layoffs becomes a technology, the narrative has already shifted the ground on which negotiations take place.

The split: AI accelerates a shift that started long ago

The second front is not how much wealth artificial intelligence creates, but who keeps it. The figure that explains almost everything is labor's share of income: what portion of the value produced is counted as wages and what portion remains as profits and non-labor income. The shift toward capital was not started by AI; it has been tilting that way for decades. AI is not the origin, it is the accelerant.

When verifying costs more than faking

The cost of manufacturing a fake piece of evidence —a voice, an image, a video— has collapsed far faster than the cost of checking it. You don't need to claim that truth has become unknowable. It is enough to accept something almost as uncomfortable: that producing credible evidence can be faster and cheaper than verifying an authentic one.

88 hours, 10,000 agents, and a door-slamming resignation

On September 8, 2026, OpenAI published a 166-page paper and a Lean formalization showing that the three-dimensional Navier-Stokes equations can develop a finite-time singularity. According to the company, it was solved by an internal model more capable than GPT-6 Astra, supported by around 10,000 concurrent agents, in 88 hours. The publication arrived wrapped in a dispute over credit and private data.

That same day, Jacob Coxon, a 27-year-old researcher with three years of pretraining work behind him, announced he was leaving Anthropic and the sector. He argues that none of the big companies is acting responsibly and that the people building these systems genuinely believe they could kill us all before the decade is out. His account is not that of an outside critic, but of someone who trained the very models he now talks about.

---

Meanwhile, Spanish tech employment remains concentrated in a handful of parks: the Tecnológico de Andalucía in Málaga accounts for around 3.1% of the workers in its province. The question is no longer whether there will be work. It is in whose hands.