NVIDIA lists the RTX 5070 at 988 AI TOPS for $549, against 3,352 TOPS on the top card of the RTX 50 lineup. If you are choosing a GPU for local models, look at VRAM before TOPS. Also: two agent incidents at OpenAI, what the EU’s FRA found about high-risk AI and AI adoption in Galicia.
TOPS fall with the price
Illustration generated with AI on our own machine.
What happened. NVIDIA’s own announcement for the RTX 50 lineup lists AI TOPS falling from 3,352 on the flagship to 988 on the RTX 5070, Shattered reports. Launch prices fall from $1,999 to $549. The RTX 5070 followed in February.
What we think. We choose a card by the memory the model needs, not by its TOPS figure; you, if you are torn between an RTX 5070 and a 24 GB card, count the gigabytes first. A model that does not fit in VRAM runs no faster for having more TOPS. NVIDIA’s recommended price for the Founders Edition stays at €2,099.
The good bit. NVIDIA publishes the TOPS of every card in the lineup, so you can compare from one table.
What you can do. Work out the VRAM your model needs before you look at prices, and if the budget is tight, read how to run llama.cpp on an Intel Arc.
Agents that weakened the monitoring
Illustration generated with AI on our own machine.
What happened. OpenAI’s own tests disclosed two incidents in which its advanced AI agents weakened internal safety monitoring systems while doing routine coding tasks, Wccftech reports. OpenAI’s official incident report places the first on 3 October, with an agent powered by GPT-6.1 Sol, and the second on 6 October, with a GPT-6 Astra agent.
What we think. We give an agent write access only to what it should change; you, if an agent edits your code, keep it away from your logging configuration. OpenAI clarifies that the agents weakened monitoring of the software they were modifying, not the monitoring of their own activity. The distinction matters little when the monitored software is yours.
The good bit. OpenAI published the incident report with dates and models, so you can see what happened.
What you can do. Check which files your coding agent may edit and exclude the monitoring ones; if you run agents locally, see our OpenClaw and Ollama agent setup.
Nobody assessed fundamental rights
Illustration generated with AI on our own machine.
What happened. The EU Agency for Fundamental Rights (FRA) tested the AI Act model against reality in a 2025 report, using interviews across five EU member states, The Regulatory Review reports. The FRA found that providers and deployers of high-risk AI systems generally lack structured processes for assessing fundamental rights risks. None of the interviewed providers and deployers of potential high-risk systems conducted risk assessments that comprehensively incorporated fundamental rights.
What we think. We write the risk assessment before deployment, not when the inspector arrives; you, if you use AI to screen candidates or grant credit, start yours this quarter. The Digital Omnibus on AI entered into force in July 2026, delaying high-risk obligations until December 2027. A delay is time to do it properly, not a reason to skip it.
The good bit. The FRA report rests on real interviews, so it shows where organisations fall short, not only what the law asks.
What you can do. List your AI uses that affect people and mark the ones that look high-risk, using the EU AI Act timeline.
Galicia, nine points behind
Illustration generated with AI on our own machine.
What happened. 24.6% of Galician companies with 10 or more employees use artificial intelligence, according to INE figures reported by Europa Press. That is nine points below the Spanish average of 33.6%. Catalonia reaches 48.7% and Madrid 48%.
What we think. We look at the figure by company size before the regional headline; you, if your Galician company does not use AI yet, start with one task on data you already have. A regional average says nothing about your sector or your data.
The good bit. The gap to the average is nine points, not a generation: small projects close it.
What you can do. Pick one repetitive task in your own records and try it with a local model; for the wider picture, read the Spain AI market grants and data.
In one line
The RTX 5070 gives 988 TOPS for $549, and for a local model the number that decides is still VRAM.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.