The truce among AI owners

In March 2023, a letter circulated asking for a six-month halt to the training of systems more powerful than GPT-4. Tens of thousands of people signed it. Elon Musk was one of them.

Four months later he founded xAI. In November he launched Grok. When someone pointed out the contradiction on his own social network, he answered without makeup: he signed knowing it was useless, he just wanted to be on the record as someone who had recommended a pause.

There's the entire genre summed up.

Today the scene returns, with more signatures and better production. Amodei, Altman, Gates and Musk himself agree that speed needs to be cut. It's being discussed as if it were an ethical awakening in the industry. It's worth reading it differently: only someone who has their hand on the accelerator can ask for a pause. The rest of us are commenting on traffic from the sidewalk.

It's worth rereading that 2023 letter. It didn't ask to stop artificial intelligence. It asked to stop systems more powerful than GPT-4. The threshold left the model that was ahead at that moment undisturbed and froze everyone who was running behind. It could be a technical coincidence. It could also not be.

The pattern repeats in Brussels. While Sam Altman toured capitals asking for regulation, his company had already done the opposite work: in September 2022 it delivered a seven-page document to the European Commission and Council arguing that general-purpose systems should not be classified as high-risk. The final version of the regulation eliminated that classification. An advisor for European Digital Rights summed it up in four words: they got what they asked for.

Asking for rules on stage and negotiating exceptions in the hallway is not incoherence. It's a strategy, and it works.

But there's a hypothesis more uncomfortable than hypocrisy, and that's if the truce is not moral but physical.

In February 2025 Altman wrote that his company had run out of GPU and that's why he had to stagger the rollout of GPT-4.5. It wasn't a figure of speech. The International Energy Agency calculates that data centers consumed 415 terawatt-hours in 2024, around 1.5% of global electricity, and projects the figure will double to about 945 by 2030. The problem isn't consumption. It's the queue. Connecting to the grid takes years, a transmission line takes between four and eight, and the same agency estimates that about 20% of announced projects risk being delayed.

Epoch AI reached a similar conclusion by another route. Of the four constraints it analyzed for scaling through 2030—energy, chip manufacturing, data, and latency—energy is the tightest in the short term, followed by advanced packaging capacity and high-bandwidth memory. None of that is fixed with a product decision. It's fixed with substations, environmental permits, and ten-year power contracts.

So the question no one in the industry has wanted to answer: if the ceiling is copper and silicon, not conscience, what exactly does it mean to ask for a pause right now?

A formal truce doesn't stop anyone. It freezes positions. Whoever is ahead keeps the lead while securing energy and factory capacity. Whoever is behind loses the months they needed to catch up. And whoever has absolutely nothing to pause applauds from the stands.

I'm not saying there's a coordinated plan. I'm saying the incentive is visible, the timeline fits, and none of those who sign have offered a better explanation.

The criticism isn't new either. When the 2023 letter came out, a group of researchers from the DAIR Institute responded that the document fed the industry hype and diverted attention from the damage that was already happening. Three years later that diagnosis has aged pretty well.

Up to here the industry. Now the country.

Costa Rica cannot accelerate this or slow it down, and that's not an opinion: it's arithmetic. In 2025 the United States produced fifty notable artificial intelligence models and all of Europe produced two. Global private investment hovered around 581 billion dollars, of which more than 344 billion stayed in a single country. Nvidia controls more than 60% of global computing capacity. Latin America contributes 6.6% of global output and receives 1.12% of that investment, while accounting for 14% of global visits to AI tools with only 11% of internet users.

We consume like giants and produce like dwarfs.

The 2025 Latin American Artificial Intelligence Index says it with its own numbers. Costa Rica is fifth in the region, fourth in governance with 77.55 points and eighth in research, development and adoption with 41.3, below the 47.74 from the previous year. We write rules better and better and build things worse and worse.

That's why the moratorium debate feels so comfortable to us. It allows you to have a position without having capacity. It doesn't require a budget, doesn't open a legislative proceeding, doesn't force any official to answer anyone about anything. It's the only conversation about artificial intelligence in this country that doesn't cost anybody anything.

And we know how to do something else. In 2017 Costa Rica chaired the United Nations conference that adopted the Treaty on the Prohibition of Nuclear Weapons, a legally binding instrument, without having a single warhead. That influence didn't come from signing someone else's letter. It came from drafting the text and chairing the table.

Compare it with February 2026, when eighty-six countries signed in New Delhi the broadest diplomatic agreement ever reached on artificial intelligence, without a single binding commitment. The difference isn't in the country's size. It's in whether you arrive with an instrument or arrive with an adhesion.

Meanwhile, there are issues that do depend on us and are barely discussed. Between January 2018 and August 2025, 40,457 reports of cybercrime were registered in the country, and 84% correspond to computer fraud and identity theft: exactly the two behaviors that generative AI made cheap. In July 2026 the free trade zones regime lost jobs for the first time in seven years.

Each one of those decisions has a responsible party, a budget, and a timeline. None of them need Silicon Valley's permission.

One reasonable objection remains: small countries do shape global norms. It's true, and we've done it. But that influence is built by being specific, not solemn. A criminal provision for AI-assisted impersonation gets copied. A public procurement notice demanding auditability and portability gets copied. A rule about what isn't automated in the classroom, and from what age, gets copied. A signature at the bottom of a declaration doesn't get copied: it gets filed away.

Those who today call for a truce are going to launch a bigger model as soon as they have the power to train it. It happened once. It will happen again. The only thing really in question is what we do with those months: write something someone else will want to copy, or keep applauding someone else's fight from the sidelines.

Sources#

Future of Life Institute, Pause Giant AI Experiments: An Open Letter (March 2023). https://futureoflife.org/open-letter/pause-giant-ai-experiments/

Elon Musk's statement on signing the letter, November 5, 2023, reported by Benzinga/Nasdaq. https://www.nasdaq.com/articles/elon-musk-admits-signing-open-letter-on-ai-limits-was-futile-but-he-just-wanted-to-be-on

Billy Perrigo, "Exclusive: OpenAI Lobbied the E.U. to Water Down AI Regulation", TIME, June 20, 2023. https://time.com/6288245/openai-eu-lobbying-ai-act/

DAIR Institute, Statement on the "AI Pause" Letter, March 31, 2023. https://dair-institute.org/blog/letter-statement-March2023/

Kyle Wiggers, "OpenAI CEO Sam Altman says the company is 'out of GPUs'", TechCrunch, February 27, 2025. https://techcrunch.com/2025/02/27/openai-ceo-sam-altman-says-the-company-is-out-of-gpus/

International Energy Agency, Energy and AI (2025), executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary

Jaime Sevilla et al., Can AI Scaling Continue Through 2030?, Epoch AI, August 20, 2024. https://epoch.ai/blog/can-ai-scaling-continue-through-2030

Stanford Institute for Human-Centered Artificial Intelligence, The 2026 AI Index Report. https://hai.stanford.edu/ai-index/2026-ai-index-report

CEPAL, Latin America and the Caribbean Accelerate the Adoption of Artificial Intelligence, October 3, 2025. https://www.cepal.org/en/pressreleases/latin-america-and-caribbean-accelerate-adoption-artificial-intelligence-though

CENIA and CEPAL, Latin American Artificial Intelligence Index (ILIA) 2025. https://www.cepal.org/en/publications/84471-latin-american-artificial-intelligence-index-ilia-2025-executive-summary

United Nations, The UN Adopts the First Legally Binding Treaty Prohibiting Nuclear Weapons, July 7, 2017. https://news.un.org/es/story/2017/07/1382151

National University, LabCIBE, Cybersecurity Status in Costa Rica 2025. https://d1qqtien6gys07.cloudfront.net/wp-content/uploads/2026/04/UNA-Estado-de-la-Cyberseguridad-en-Costa-Rica-2025-v4-1.pdf

Sofía Chinchilla, "For the First Time in 7 Years Costa Rica Loses Jobs in Free Trade Zones", La Nación, September 5, 2026. https://www.nacion.com/economia/por-primera-vez-en-7-anos-costa-rica-pierde-empleo/FOEXQV6VSRGLTBZRKKRQGOJ5C4/story/

Government of Spain, Spain Promotes an International Coalition to Protect Minors in the Age of Artificial Intelligence, July 6, 2026. https://www.lamoncloa.gob.es/serviciosdeprensa/notasprensa/transformacion-digital-y-funcion-publica/Paginas/2026/060726-coalicion-proteccion-menores-ia.aspx