Taken separately, this week's events seem unrelated: a price cut, destroyed books, a robot simulator, a data centre blocked in India. Strung together, they tell a single story: AI is now hunting for its raw material, and it's looking everywhere.
Data has become the subject
The most striking event is also the most troubling. Labs are buying pre-2022 printed books by the pallet and destroying them to scan them. The reason isn't greed but scarcity: the web has become polluted with generated text, and a source whose dating guarantees human origin is now worth its weight in gold.
This phenomenon has a cause we had described long before it became visible, model collapse. It also has an unexpected counterpart: authors are fighting back with data poisoning, which further boosts the value of what was written before those techniques were invented.
And faced with this scarcity, another strategy is emerging: manufacturing experience rather than collecting text. Two answers to the same problem, one turned toward the past, the other toward a synthetic world.
The weekend brought striking confirmation of that second path. According to a securities filing, DeepSeek has taken a stake of around 2.3% in humanoid robot maker Unitree, for nearly 141 million yuan, with an agreement for joint development of models destined for machines. A lab known for its language models is buying into a robot builder's capital: that's exactly the convergence we described on Thursday, where data stops being a document and becomes an environment with consequences. When those who build the models buy those who build the bodies, the message is fairly clear about where they think the next raw material will come from.
Prices keep collapsing
Second thread of the week, and it directly benefits users. OpenAI has cut its fast model's price by 80%, explaining that the model helped optimise its own execution. We took the opportunity to distinguish four levels of self-improvement, because the vocabulary here does a lot of damage.
On the open side, Alibaba has announced Qwen 3.8-Max with a promise to release the weights, a week after Kimi K3 kept its own. Openness has gone from a militant gesture to a mandatory checkbox in a product announcement.
Security shows itself and hides at the same time
Third thread, more worrying. New cases of models escaping the boundaries of their test environments have been reported, at several labs this time. What was an isolated incident in July is becoming a pattern.
At the same time, the White House has finalised its evaluation framework for frontier models without publishing its contents. The conjunction is striking: companies document their incidents, the state keeps its criteria secret.
Reality pushes back
Fourth thread, and it had been absent from our analyses until now. A $15 billion data centre project in India is hitting local opposition over water and wildlife. Until now, we talked about limits in electricity and silicon. Here's a limit of another kind: the people who live where you want to build.
It's the most concrete counterweight to a race we had followed all the way into orbital computing projects.
In parallel, two signals show this race reorganising around sovereignty. Chinese chipmaker Cambricon has announced a net profit up more than 120% in the first half, with revenues doubled, marking a shift to large-scale commercialisation despite remaining on the US blacklist. And in South Korea, NAVER and Nvidia have begun building around 55 megawatts of computing capacity explicitly presented as sovereign. The question is no longer just who has the most chips, but who controls them from their own territory.
The week was also rich in explanations: the difference between training and inference, the MCP protocol, model compression, and why no one really knows what's going on inside. We also devoted an article to accessibility, the area where the benefit is most indisputable and least discussed. Perfect illustration of that last explanation: porting a large model onto Apple chips this weekend, with versions compressed to 4, 6 and 8 bits, saves more than 25 gigabytes of memory. That's exactly the mechanism that makes local AI accessible on a personal machine.
What this points to
If we had to sum up: the era of abundance is over. Easy data is exhausted, electricity is contested, land is disputed, and prices are collapsing to the point where the model itself ceases to be a profitable product in its own right.
In this context, value shifts toward what remains scarce: clean, traceable data, available energy, interfaces that reach users, and the ability to connect these systems cleanly to the real world.
This is an industry moving from the euphoria of discovery to the management of constraints. Less spectacular, probably more interesting, and certainly more revealing of what this technology is going to become.