Nvidia told the manufacturers that assemble its servers about a price increase of more than 15% on systems built with its Vera Rubin and Grace Blackwell chips, affecting equipment that ships in early 2027. Fortune reported it, based on Bloomberg’s reporting, and also noted increases of up to 39% on consumer graphics cards. The announcement comes the week before the company reports earnings.
The detail that changes the reading is the margin. Nvidia already operates with a 75% gross margin, so this is hardly a cost pass-through to the customer: the margin is consistent with a supplier with no real alternative exercising market power, though on its own it does not prove it. When a single company sets the price of the basic input for the entire industry—the chips on which models are trained and run—raising it 15% requires neither justification nor negotiation.
For Latin America the causal chain is short, and it can be verified with facts from this same week. Brazil just put out a tender for two supercomputers worth 2.3 billion reais, with delivery scheduled for precisely 2027. Chile is debating a token bank—a state reserve of compute capacity bought in advance—in its 2027 Budget. And every cloud contract in the region is renegotiated on prices set upstream, in dollars, by a single company. The computational sovereignty that several countries in the region decided to buy just became 15% more expensive before reaching port, and the fixed-price tender terms already signed are the ones that absorb the difference.
Also today
- OpenAI asks California to strengthen the AI safety law it once fought — it now wants mandatory monitoring of models in training and tougher cybersecurity requirements, under a doctrine it calls “reverse federalism”: if enough states copy the same law, that amounts to a national standard.
- No frontier lab has a public plan to contain a model that resists — the first comparative audit of control practices grades Anthropic and OpenAI a C+ and Meta an F, and finds no full implementation of any of the six practices evaluated.
- An agent built on a 27B open model replicates published research better than frontier models — the startup Inherent, founded by four Google DeepMind alumni with $50 million in seed funding, attributes the result to the scaffolding around the model rather than to the model’s weights themselves.
- An anonymous model gives away 100 trillion tokens a day, and no one says who runs it — the tokenizer, the token errors and a leaked Java trace point to China’s Zhipu. Along the way, one data point: the U.S. share of tokens served by OpenRouter fell from 70% to 30% in a year.
- Harvard sells a $699 bootcamp in which part of the teaching is done by AI avatars of its instructors — the world’s most expensive educational brand found a way to license its reputation without the academic. The investor who was replicated admits finding it somewhat unsettling.
In the region
The weekend brought no institutional moves from the region itself: no dated publications from UNESCO, the OECD, ECLAC or the IDB, nor from ministries, data authorities or public procurement platforms in the region. What does change comes from abroad and lands in two ways. The first is price, and it is the story above: Nvidia’s increase hits precisely the countries that in the past month decided to buy compute instead of leasing it, on equipment to be delivered in 2027. The second is corporate governance. The audit of control practices published this week provides, without meaning to, a six-point checklist—action logging, monitor effectiveness, actions requiring prior permission, a circuit breaker, third-party review and a containment plan—that no tender on ChileCompra, ComprasNet, SECOP II or CompraNet asks about today, and that any of them could add without changing a single law. And there is a third, more uncomfortable reading: the doctrine of “reverse federalism” that OpenAI just named in California describes precisely what is already happening here, where the AI bills under consideration in Brazil, Chile, Colombia, Mexico and Peru are being written while looking at texts that the same companies helped draft in another jurisdiction.
Launches
- Ox Alpha — a stealth model on OpenRouter with a context of 1,048,576 tokens and text, image and video input, free during a one-week promotional window and with no declared data retention. It matters for what it is and for what it exposes: frontier capability at zero cost, served by a provider that does not identify itself. The analysts’ explicit recommendation is not to feed it code or sensitive data until it is known who is behind it.
- Google Antigravity in Gemini Enterprise — autonomous coding agents included with no additional license in Gemini Enterprise Standard, Plus and Standard Emerging Market, the tier Google uses to sell to the region. It comes with extensions for VS Code, Visual Studio, JetBrains and Zed. What is interesting is not the agent but the controls that come with it: monthly spending caps, shared token quotas, centralized auditing and workspace isolation. With the license included from day one, the barrier is no longer access but the internal governance of agents that carry out tasks on their own.
Threads we’re following
This adds to a story we have been following for days. Last week Brazil put out a tender for two supercomputers of its own, in what was the first complete computational sovereignty strategy written by a state in the region, and yesterday we reported that four Latin American countries took delegations to Beijing in the same week, each negotiating on its own. Today’s increase puts a price on that fragmentation: when the sole supplier raises prices 15% without explanation, whoever negotiates alone, with the tender terms already signed, has no ground to respond from.
If the sole supplier can raise prices 15% on any given weekend, and the company rated best in the world at containing its own model barely earns a C+, what exactly is a Latin American state buying when it signs an AI contract: a technology, or a place in someone else’s line?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.