Nvidia took an equity stake in Safe Superintelligence, the lab Ilya Sutskever founded in 2024 that to date has published nothing, has no product and has no revenue. The agreement also gives the lab access to the next-generation Vera Rubin systems: the compute available to it grows by an order of magnitude. The announcement does not disclose the amount (Bloomberg put it at about $5 billion, at a valuation of about $32 billion), but the part that is not money weighs more: Nvidia gains access to SSI’s research, and those findings will feed into the design of its next chips.
It is the second move of the same week. Two days earlier it emerged that Nvidia is negotiating to guarantee about $250 billion of the lease on OpenAI’s campus in Ohio. In forty-eight hours the company went from financing the debt with which its largest customer buys its chips to buying a position in the most secretive pure-research lab in the sector. Today it is, at the same time, supplier, creditor and shareholder of its two frontier customers. The practical consequence is that the price of compute stops being set in an open market and starts being set inside a web of cross-holdings.
For Latin America, the effect is indirect but real. The region buys that compute resold by cloud providers, two or three links further down from where these terms are negotiated, and no regional competition authority has jurisdiction or tools over that structure. And the same boom has already surfaced in a place where it is felt directly: Google confirmed that it is raising the price of the entire Pixel line because a gigabyte of RAM went from $2.80 to $12 in a year, after Samsung, SK hynix and Micron redirected their capacity toward memory for AI data centers. It is the first time the build-out of AI infrastructure has shown up on the price tag of a mass consumer product: about $100 more per phone.
Also today
- Hundreds of shared Claude conversations and Artifacts were indexed on Google — a
noindextag was missing: medical records, performance reviews and API keys ended up searchable, and someone backed up 11,241 messages before they were unpublished. - Artificial Analysis measures a 51% hallucination rate in Kimi K3 on the same day its weights are released — the largest open model in history knows more and also makes up more, and the figure does not appear in any chart published by Moonshot AI.
- Keiko Fujimori takes office as president of Peru with a plan that bets on predictive surveillance — C5i command centers with predictive analytics and a single health record linked to the national ID document, in the country with one of the most advanced AI frameworks in the region on paper.
- Spotify does not label AI-generated music, so volunteers built their own catalogs of synthetic artists — homemade spectral analysis to separate the human from the synthetic in a royalty pool that is zero-sum.
In the region
The regional move of the period comes from an incoming government rather than a legislative process. Peru inaugurates the presidency of Keiko Fujimori, the first woman elected by popular vote in more than two centuries of republican history, with a 2026-2031 plan whose artificial intelligence components focus on security: predictive policing, interconnected C5i command centers for anticipatory crime analysis, and a Single Electronic Health Record linked to the national ID document that would share data among MINSA, EsSalud and other systems, probably the largest database of sensitive personal data that state would ever have built. What is interesting is not the promise but the context: Peru already has Law No. 31814, its implementing regulation in force since September 2025, and a completed UNESCO readiness assessment. It is the first case in the region where one can observe whether an AI framework written before a government took office serves to set limits on that government, and the concrete test will be whether the National Authority for Personal Data Protection has the means to oversee the Ministry of the Interior. In parallel, and in a less predictable direction, the Peruvian private sector is offering itself to Mexico as a provider of AI, digital services and cybersecurity: a South-South operation that goes through neither Washington nor Shanghai. In Santiago, meanwhile, the fourth edition of the Chile Digital Summit is taking place at ECLAC headquarters, under the slogan “Data Hub of the Americas.” On the immediate calendar: on August 2 the transparency obligations of Article 50 of the European AI Act take effect.
Launches
- Celeris-1 — the first model from the lab founded by the creators of Marqo. It abandons word-by-word generation and uses diffusion to build the entire response at once: median latency of 157 milliseconds, about fifteen times less than GPT-5-mini, at the cost of giving up capability (76% on MMLU-Pro versus 81% for GPT-5). Available via an OpenAI-compatible API, without open weights. It matters because the region’s conversation with AI increasingly happens by voice (call centers, citizen services, messaging), and voice does not tolerate waiting.
- Build, in public alpha — Roblox — it turns a natural-language instruction into a playable game, with iteration via chat and continuity in Roblox Studio. It enters a free public alpha, only in New Zealand, for verified users aged nine and up: creating is allowed from age 9, publishing to a global audience only from 16. It is worth following even though it cannot yet be tried from the region: it is the first mass deployment of generative AI explicitly aimed at children, and its age-limit design is child-protection policy written by a company before any regulator.
Threads we’re following
We had been following the question of who can pay for the infrastructure to run frontier models, which last week emerged as the real bottleneck behind open weights. Today that story added a different chapter: the problem is no longer just how much compute costs, but who sits on each side of the table when the price is negotiated. When the maker of the silicon is also a creditor of one customer and a shareholder of the other, the cost of entry stops being a market figure and becomes the result of an agreement between related parties.
If the chip supplier is at once a creditor of one customer and a shareholder of the other, and on top of that takes both of their research to design the next generation of silicon, is there any price of compute left that a Latin American state can negotiate, or is the only thing negotiable from here on whom it rents from?
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.