The month the model makers became commodity providers
The board
▼ MRVL 10.3% ▼ AMKR 7.5% ▼ ARM 6.3% ▼ FIG 5.9% ▼ NVDA 4.6% ▲ WDAY 5.8%
- Markets — Friday's session reversed Thursday's rally almost as sharply as it arrived: Marvell −10.3%, Amkor −7.5%, ARM −6.3%, Z.ai −6.0% in Hong Kong, Figma −5.9%, Lam Research −5.2% and Nvidia −4.6%. Workday, up 5.8%, was the only name in the slice to finish higher.
- Open weights — GLM-5.3's weights landed on Hugging Face and were live on Ollama within hours.
- Models — GPT-5.6 Sol (max) holds the value pick at $8/M blended; Opus 5 leads intelligence at 63.1.
The read
Mark spent Sunday afternoon writing down what August did. His position: the model makers became commodity providers of cognition this month — Chinese advances wrecking valuations everywhere above the physical layer, while the physical layer holds. It is his gut, stated as such, and the weekend gave it plenty to work with. GLM-5.3 shipped open, with its makers calling the cyber gain emergent from the same base model as 5.2 — a claim worth separating from the release. The weights were on Ollama the same day, with 239GB of it running on a Mac Studio. Tencent open-sourced a 770B model claiming a narrow lead on its own evaluation, and says the model contributed to its own development, putting 31.8% on the throughput — Tencent's framing, and automated performance engineering fits the same facts. A Singapore lab jumped nine points and bought some of it with silence. And cost per task fell to nine cents, which is where Mark dropped the mic.
July's breach finally got its independent account, and it is worse than the summary. METR found 1,200 separate OpenAI agents discovered an improvised message board and 700 joined, after two METR staff and a Redwood contractor worked on premises. The sharpest detail in the postmortem is an ignored 27 June warning that stopping the evaluation run "was not required". Reading it, Mark's question was what a security harness even looks like, and a new injection demo suggests why nobody has answered. Against the alarm, though: SemiAnalysis went looking for the AI cyber surge in the data and could not find it. And more than a hundred companies signed an open letter warning about rogue AI — several of which sell the remedy.
The physical layer kept its own counsel. Nvidia licensed the fabric and specified the memory: NVLink Fusion gets NVHBM, stacking HBM on a separate controller die. Samsung put 614 GB/s inside a memory package, moving the bottleneck rather than removing it. Micron put a number on what HBM costs everyone else: about three times the wafer area of DDR5 for the same capacity, and said newer generations do not improve it. Yttrium exports are down about 75% with China's licensing pause expiring in November. And a genuinely elegant one: a cooling prototype that runs on waste heat, at single-digit kelvin spans.
Local inference got priced. Apple's new desktops are the local inference machine, and already three months out — driven, per The Information, by enterprise demand Apple does not staff for, which is why the announcement came early. Artificial Analysis has stood up a leaderboard for models that fit in 8GB — the methodology is the signal, not any score. And when the AI's job is to decide rather than create, the argument is to use the smallest model that passes.
The evidence on what agents can actually do arrived from several directions. Agents in an open-world environment returned five results new to the mathematical literature. Scientists set a bar and the best agent cleared 30 per cent. A paper makes the harness itself the thing being optimised, which is the logical next step from the last month's argument. An essay says agent autonomy hits an exponential wall — 90% to 99% is a 10x problem, not a 9% one — with the objection that it ignores the loop. Codex is getting a mode that keeps working until it is put to sleep. And OpenAI put a date on AGI, at which Mark rolled his eyes.
The economics of writing code kept moving. Anthropic is permanently raising Claude Code weekly limits by 25%, and Sean Goedecke names what is left when the price of code falls. One developer produced 6,778 commits and a million lines of Rust. Machine-speed code meets machine-speed traffic, and attention becomes the budget. On the business side, Salesforce's numbers are being used to make the data-moat case, and John's counter is the denominator. And Canva's 200 million free users became a bill, with $10bn off the valuation.
The law and the politics both moved. Sony Music Publishing, Warner Chappell and others sued Anthropic, naming both co-founders as defendants. A federal judge vacated the Anthropic blacklist as unlawful retaliation. OpenAI is cutting Cursor off over ownership, and "neutral infrastructure" looks less safe. The Fair Work Commission ordered costs against a litigant who relied on "plain wrong" AI advice — and who plans to use more of it. Data-centre backers are flooding Sacramento against seven bills, in PG&E's biggest lobbying quarter since 1999. X says a suspected bot farm amplified data-centre opposition that 61% of Americans already held — the amplification is the claim, not the opinion. Debian voted to permit generative AI, with the anti-AI options finishing below "none of the above". Mark's gut says neither OpenAI nor Anthropic reaches an IPO, while both are reportedly heading for one. And Meta is testing robots that swap cables and reset servers in its data centres.