One open-source model, one Nasdaq dip
Let me lay out the numbers in order: Moonshot announced Kimi K3, Xi Jinping spoke at the Shanghai conference, the Nasdaq fell about 1%, and investors sold off chip stocks like Nvidia. This isn't bubble panic; it's a real repricing. The market had been betting on the assumption that 'the best model is locked inside American labs and expensive silicon' — when an open-weight model gets near the frontier, that assumption gets marked down.[1]
There is no cloud; there are machines wired to somebody's transformer — and who runs those machines now matters as much as who posts the best score. Moonshot itself admits Kimi K3 'still trails Claude Fable 5 and GPT 5.6 Sol' — that's not a bubble number, that's an honest concession. But Arena.ai's independent test found it competitive anyway. On July 14 I wrote about SK Hynix's $26.5 billion IPO; what I'm watching now is some of that capital getting reallocated — not by the question 'American model or Chinese model,' but by 'which one scales cheap.'[1], [3]
Google is billing by watts, not by tokens
The same week, Google started metering Gemini usage by the compute cost of each request instead of the request count. This is company policy catching up to the point I made on July 16: the constraint is no longer silicon, it's watts. A free user is capped at a 32K-token context while an Ultra subscriber can reach 1 million — but even Google's own documentation admits the limit 'may change based on testing, experimentation or availability.' Translation: how much GPU they can spare us changes day to day.[2], [4]
We'll keep reading the bill in megawatts — but now we need to read it in watts on the user side too. My guess: Anthropic and OpenAI move to similar compute-based metering by late July, because flat per-token pricing can no longer protect lab margins in a market being undercut by cheap alternatives like Kimi K3. These aren't bubble numbers; they're the shipped result of real cost pressure.[1], [2]