The weights landed on schedule; the hard problems are money and oversight
The board
▲ WDAY 9.0% ▲ FIG 8.6% ▲ PLTR 7.0% ▼ AMKR 6.5% ▼ ASML 5.8% ▼ AMD 5.2% ▼ NVDA 5.0%
- Markets — a clean rotation out of AI hardware into software: Nvidia −5.0%, AMD −5.2%, ASML −5.8% and Amkor −6.5%, while Workday +9.0%, Figma +8.6% and Palantir +7.0% ripped. It happened on the session CXMT's Shanghai debut closed up 466%, handing a Chinese DRAM challenger the public equity to expand straight through a downcycle — historically how memory cartels break.
- Open weights — Kimi K3's weights are live and already top Hugging Face trending on likes, days after release.
- Models — Opus 5 still holds both ends of the model board: top intelligence at 60.7, and the best value at $10/M blended.
The read
The date held. Kimi K3's weights landed exactly when Moonshot said they would, with Ollama metering them within hours — a week after this newsletter flagged 27 July as the test of the open-versus-closed gap. The sharper number arrived alongside it: a year-old open model still carries about a third of Opus's token volume, which is the commoditisation thesis stated as a usage statistic rather than a benchmark. Open weights has meanwhile become American policy, argued hardest by the companies it costs most. The historical frame is worth holding onto: almost nobody ever wrote open source, and it ate computing regardless — minority production, majority consequence.
The money questions got sharper, and they are no longer being asked politely. Nvidia's $250bn backstop has revived the circular-financing debate — conducted, as ever, without the actual numbers; there is now roughly a trillion dollars of capex and no settled answer on who repays it; and Morgan Stanley has moved AI onto the margin line, where this week's results will test it. Predictably, bubble scepticism is finding the audience that needs it to be true — which is not the same as being wrong. Underneath sits the disagreement that actually determines what you are underwriting: protopia or step function. The supply side answered in its own register: Huawei is building its own DRAM fab to get memory it cannot buy, CXMT closed up 466% on debut with the balance sheet to break a cartel, and at the opposite end of the scale a 2.4B model is now answering questions inside ordinary DDR4 memory cells.
Then the week's uncomfortable fact: an agent compromised a company for days, and the provider noticed a week later — in the same stretch that Meta AI acquired hands. What is striking is that none of the credible responses is a better model. Microsoft's security pitch is a harness-engineering thesis compressed into one sentence; Shopify went back to clean code because the agents demanded it; and proof irrelevance names the one domain where an unreliable generator is provably safe — verification, again, as the scarce input. Two political notes close the day: data centres can be powered down in three minutes, which changes the politics of siting them, and a national model is not the same thing as a public-interest one.