Recurring storylines, tracked across issues so trajectory stays visible rather than being re-discovered every month.
Thread
State as of 2026-09-20
Pacing, the safety-slowdown question, and now its antitrust exposure
Subscribers sued Anthropic, OpenAI, SpaceXAI and Google on September 18th, arguing the labs' public pacing talk was an illegal agreement to restrain competition rather than four unilateral decisions. Congress split further: Sanders and Casar promise a superintelligence ban, Warren wants a government-imposed pause, Meta and Nvidia's Huang reject coordination outright. Nobody has actually slowed a release; Anthropic was reported the same week to be weighing a new model to counter OpenAI.
Capability gating and embedded evaluators
Anthropic named Accenture's Faculty unit as its first embedded evaluator, with both companies expecting to commit at least $1 billion over five years. Access, reporting and funding terms remain under development. OpenAI separately began publishing its own misalignment incident reports, six in the opening batch, under a process OpenAI alone controls.
Verification and credit disputes
No new dispute this week. Last week's Fields Medalist declaration and the OpenAI/Nvidia math-credit fights stand as the record.
Circular & off-balance-sheet financing
Nscale's public S-1 disclosed $140.6 million in first-half revenue against a $1.02 billion net loss, anchored by a $44.6 billion Anthropic infrastructure contract, the clearest documented instance yet of the pattern this ledger keeps tracking. Nvidia's reported anchor talks for Anthropic's own IPO remain unclosed.
The IPO queue
Seven names now: Anthropic (WSJ reports a third slip, October to November, to carry Q3 results), OpenAI (no listing this year), Moonshot (dual HK/Shanghai exploration), SB Energy (no update), Enflame (listed, closed under target), DeepSeek (CITIC-led STAR Market prep, no filing yet), and new this week, Nscale (filed a public S-1 for NYSE under NSCL, anchored by Anthropic and Microsoft).
China's IPO wave, frenzy versus fade
No update this week. Enflame's close under target stands as the clearest data point.
Open-weight commoditization
StepFun opened API access to Step 5 Preview, a 600B-parameter sparse MoE with 27B active, still under the 2.8T threshold Prediction T5 watches. DeepSeek's technical report explained V4.1-Flash's memory efficiency without a new weight release.
Containment & agent security
The week's heaviest thread. Google confirmed Gemini breached three real companies during authorized May red-team tests. Researchers described chaining a bug into OpenAI's own internal code repository, with Claude helping write the exploit. The Plugin4Shell SHA-pinning flaw hit four coding agents; Anthropic and OpenAI patched, Microsoft and Google had not as of Friday. CrowdStrike attributed a new npm information-stealer largely to LLM-generated code.
The deployment gap
No new hard data this week. Zapier's CEO put a number on one platform's automation share, roughly 40% of tasks on its own Automation Bench, in a podcast rather than a filing.
Entry-level collapse / labor substitution
No new BLS print. The September report, due by October 10th, is the standing test for Prediction 2026-09-06-B2.
Power as the constraint
The House passed the Ratepayer Protection Act 417-3, which would make data centers over 100MW cover their own grid-upgrade costs; it has stalled in the Senate. The Fed raised rates a quarter point, its first hike since 2023, repricing every AI financing structure built during the easing cycle. Amazon committed up to $8 billion to Generac for backup generation, and India doubled its chip mission to $13.5 billion with Applied Materials and Lam Research committing alongside it.
Frontier pricing power
No new data this week.
AI-generated science
No new data this week.
Google's talent drain
No new data this week.
Consent and the free tier
No new data this week.
US-China AI relations, ahead of the Trump-Xi summit
New thread. Bessent and He Lifeng met in New York Sunday to set the agenda for Xi's September 24th state visit, where AI, chip access and the tariff truce all compete for room. Eight tech CEOs, including Musk, Bezos, Huang, Altman, Pichai and Cook, are confirmed for the state dinner. Rhodium estimates Chinese models earned a tenth of OpenAI and Anthropic's combined revenue over March-August despite comparable capability, and Z.ai and MiniMax shares have each fallen more than 30% this month on distillation-restriction fears.