AI slowdown isn't about ethics; it's about when the numbers stop adding up

Why AI is slowing down: scaling, inference cost, and lawsuits

Videos circulate of Chinese robots performing kung fu choreographies with the fluidity of martial arts actors, while in the West, there is a solemn announcement that it is time to slow down. The scene summarizes the problem: those who could stop, don't stop. And those who preach prudence have a bill they can't afford.

The reading that gains weight is uncomfortable. No one halts a profitable business out of scruples. They stop when the bottom line dictates silence.

The «voluntary slowdown» of AI, viewed through that lens, is a noble wrapping for a problem of margins.

The scaling law no longer pays for itself

For years, the trick was adding parameters: more data, more computing power, better models. That mechanism has been exhausted. Performance curves have flattened, and each additional point of improvement costs, according to circulating calculations, triple what it used to.
It is not prudence; it is the exhaustion of the mine. The same investment that bought a leap yesterday buys only a scratch today.

Serving AI costs more than training it

Training a model is a punctual and noisy expense. Serving it is a silent, daily drain: millions of trivial queries invoiced at subsidized prices. Add the electricity and water costs of data centers whose consumption effectively turns them into covert power plants. When the bill arrives and the neighbor protests about the noise, the spectacle is over. The breakdown of that cost per query, item by item, is where the narrative turns into numbers.

Who benefits from the slowdown: regulation, lawsuits, and bankruptcy

There is a more cynical, yet widely held, view regarding the AI slowdown: the slowdown benefits those at the top. Big tech companies stir up the narrative of existential risk and, in passing, push regulations that slow down Chinese competition and choke startups without the capital to comply.
It is even argued that regulation only serves to protect the wealthy. To this is added the judicial front regarding copyright, with lawsuits looming on the horizon, and the ghost of sectoral bankruptcy: gigantic investments yielding ridiculous returns that still demand more investment to keep up.

Less epic, but more common: installing AI in a company and having the invention fail to deliver doesn't help sell the product either.

The sobering fact is not sarracena; it is accounting: every additional improvement costs triple, and every query served subtracts from the margin. No one abandons a vein of gold on principle. They abandon it when the vein stops paying.