SpaceX's $40B chip deal, Tencent's Hy4, and banks blocking AI agents

The most revealing number in today's news isn't a benchmark score, it's $40 billion. That's the size of the financing package Goldman Sachs and Apollo Global Management are now trying to syndicate to investors, all of it tied to the AI chips SpaceX wants to buy. I find this more telling than any model launch this week. We have moved from AI being funded by venture rounds and corporate cash flow to AI hardware being packaged, sliced and sold like infrastructure debt. That is how railways and power plants got built, and it carries the same risk: when chips become collateral, their depreciation curve matters enormously. Chips that are obsolete in three years sit awkwardly beside financing meant to last longer. By the way, the fact that SpaceX is a customer at this scale tells you how thin the line has become between "space company" and "AI compute company."

Meanwhile, the model race keeps moving on a different axis than raw intelligence. Tencent has released Hy4, an open-source model that combines stronger visual understanding with, for the first time in the Hunyuan line, the ability to operate a computer. Tencent calls it frontier-class, and I'd treat that claim with the usual caution until independent evaluations land. What is genuinely notable is the direction. Computer use is where Anthropic, OpenAI and now Tencent are converging, and an open-weight model with that capability changes the economics for anyone building agents. You no longer need to rent a closed API to automate a desktop workflow. The geopolitical angle is hard to ignore, too: Chinese labs keep shipping open models that narrow the gap, and Western export controls on chips look less decisive each time they do.

That agentic push is also reaching the enterprise software layer. Adobe has made its CX Enterprise Coworker, a marketing agent for orchestrating customer experience, available inside both ChatGPT and Claude. I think this is a quiet but important shift. Software vendors are no longer insisting that users come to their interface; they are bringing their agents to wherever the user already works. Chat assistants are becoming the new operating shell, and everyone else is turning into a plugin. That is good for convenience and less obviously good for Adobe's leverage over its own customers.

Two research items today are worth reading side by side. One audit questions how reliable LLM judges are for evaluating clinical text, noting that most existing evidence is thin. Another study injects "thoughts" directly into models to test whether they can recognise their own internal states. Both circle the same uncomfortable question: how much do we actually know about what these systems are doing? We are using models to grade other models in medicine, while still mapping whether they have any accurate view of themselves. Is that a foundation we are comfortable building on?

I suspect the answer will come from the money before it comes from the science. If lenders start demanding proof that these systems perform reliably, the evaluation gap stops being an academic footnote and becomes a pricing input.