Money was the story this week. Two of the largest numbers in AI arrived within two days of each other, and neither of them was about a new capability. Both were about paying for compute.
Nvidia reported a quarter that beat almost every estimate, with data center revenue of 89 billion dollars, up 117 percent from a year ago. The forecast was the bigger number: the finance chief said capital spending at the top five hyperscalers should reach 1.3 trillion dollars next year, up from 800 billion this year. Alibaba raised 10.2 billion dollars in Hong Kong’s largest follow-on share sale and said all of it goes into AI infrastructure. The stock fell 8.5 percent on the news, which says something about how much patience is left for spending of this size.
The model news went the other way. Alibaba also shipped Qwen3.8-Flash-Next, and Simon Willison ran it on a DGX Spark at home. It is a mixture-of-experts model with 125 billion total parameters but only 6 billion active at any moment, and it doubles as an early preview of the Qwen4 architecture. So the same company is buying compute by the billion and giving away a model that runs on one desk. Robots are further behind. TechCrunch reported from the Actuate conference that physical AI is still in its GPT-2 era, slowed less by model quality than by a shortage of good training data.
Safety work continued quietly. Anthropic committed 5 million dollars to independent research on how AI affects user wellbeing, with results to be published as open-source evaluations that any developer can run. A new paper, Aligned Alone, Misaligned Together, found that agents which behave correctly on their own can be pulled off course by a small adversarial minority once they operate in a population. The practical part is that the authors can predict which populations are vulnerable by watching them before any attack starts.
Next week the forecasts start meeting reality. The question worth watching is whether anyone else follows Alibaba to the market with a raise this large.
T.
