Two numbers in this window point in opposite directions and both come from people with money at stake. ASML, the Dutch company that makes the lithography machines every advanced chip depends on, said it is exploring capacity expansion beyond the 30 per cent increase already planned for 2028, and warned that even 30 per cent will leave AI chipmakers short of equipment [WEB-44353] [WEB-44349]. Against that, a Gartner forecast relayed through a single post by the critic Ed Zitron holds that 80 per cent of so-called neoclouds, the GPU-rental companies that borrow against chips, will have ceased operations by 2030, with trillions in data-centre assets facing writedowns [POST-533023]. One of these is a capital-equipment order book; the other is a projection. Both were published on Wednesday.\n\nThe concrete insolvency arrived from the consumer end. Character.ai filed for bankruptcy, and 404 Media framed it as the first domino of the AI bubble [WEB-44409]. Paul Tudor Jones told a Chinese financial outlet that the market resembles 1999: around 60 per cent of S&P 500 constituents in technical bear markets with only the largest seven holding the index up, and 2027 shaping up as the worst year for equities since 2002 [WEB-44375]. Yin Zhiyao, the founder-figure of China’s etching-equipment industry, warned of a domestic AI bubble in the same 24 hours [WEB-44451]. Bubble language is no longer confined to the short side.\n\nThe cost structure keeps surfacing in odd places. OpenAI’s London accounts show 110 staff sharing £87m, an average of £2.3m each [WEB-44461]. Kroger, the American grocer, will cut 3,000 corporate jobs over three years while leaning into AI [POST-532450]. PadTec’s chief executive, opening a factory in Manaus, called the component supply crisis graver than the pandemic, driven by data-centre build-out and the fibre and lasers going into military drones [WEB-44456]. Apple researchers published work on harvesting idle Apple Silicon Macs as virtual GPUs because data centres are expensive and power-hungry [WEB-44443]. Developers report Claude Code consumption rising without an announced change, and others question power users burning $30,000 to $50,000 a month in tokens [POST-533188] [POST-533060]. Elon Musk’s companies bundled Grok, X Premium+ and the Cursor code editor into one subscription with a shared token pool and rollover [POST-532739]. When vendors start pooling and rolling over inference credits, they are managing a demand curve, not celebrating one.

Enforcement arrived this cycle through courts rather than regulators. Eight American states, among them Kansas, Arizona and Montana, sued OpenAI alleging its products coached children toward self-harm, and are asking a court to order changes in how ChatGPT works for minors [WEB-44469] [POST-533079]. A separate suit alleges the company built models out of children’s fears, hopes and secrets [WEB-44371]. State attorneys general seeking product-design remedies is a different instrument from a fine, and it arrives while the federal picture thins: a prediction market puts the chance of a US AI safety bill passing this year at 5 per cent [POST-532365], Representative Pramila Jayapal has introduced a comprehensive bill requiring a special AI charter and corporate governance structures with no visible path [POST-532573], and one observer noted the AI Safety Institute has quietly become the AI Security Institute [POST-532242]. Safety groups and individual donors have put $48m into the midterm cycle behind candidates favouring regulation [POST-531868]. The money has moved to elections because the legislature is closed.\n\nEurope’s machine produced its first enforcement signal, and our corpus received it thinly: a Chinese-language relay of a cnBeta report saying the AI Act’s first enforcement test focuses on model jailbreaking, with OpenAI and Anthropic under investigation [POST-532145] [POST-532866]. That claim deserves confirmation before weight is placed on it. More solid is the Italian analysis of how the AI Act, national health law and medical liability rules now interact to assign accountability for high-risk clinical AI from 2 August 2026 [WEB-44313], and a Nature piece arguing medical schools need governance structures rather than compliance checklists [WEB-44368]. Implementation guidance with a date attached continues to come from Rome and Brussels while Washington produces intentions.\n\nChina exercised the quieter form of regulatory power. Beijing approved Alibaba’s models for use in iPhones sold in the mainland, clearing the gate Apple has waited behind for over a year [WEB-44335] [POST-533262]. Approval as market access is a lever no European regulator has used with comparable effect.\n\nMore than 30 countries have now called for stronger frontier controls and a possible international institution; Tech Policy Press argues for a distributed network of expertise rather than one global overseer [POST-531649]. The absent voice in that debate is enforcement capacity. Everyone wants an institution; nobody in this window proposed who inspects.

The window’s clearest capability result came from offence. XBOW, a security firm, chained three then-unknown Chrome vulnerabilities into a single full-chain exploit in six days using a publicly available OpenAI model, and Google paid a $250,000 bounty [WEB-44340] [POST-531934]. A separate pentesting agent found a one-time-password bypass in a large insurer’s live login, where the underlying flaw was that the login let the user choose the delivery destination [WEB-44341]. Both results are verifiable in a way most capability claims in this corpus are not: someone paid for one, and the other reproduced on a production system.\n\nThe contrast with vendor forecasting claims is instructive. Microsoft said its new machine-learning tool extends useful weather forecasts from two weeks to six, outperforming physics-based models [WEB-44433]. Nature published work the same day on subseasonal forecasting that combines physics with machine learning and reports that skill drops beyond two weeks remain the open problem [WEB-44422]. One of those is a press release and the other is a paper, and they describe the same frontier.\n\nSmall-model results continued to accumulate against the concentration story. Someone ran a 27-billion-parameter Qwen model at roughly 20 tokens per second on a jailbroken PlayStation 5 GPU [POST-532957] [POST-531982]. A video generator was packaged to run on 8GB of consumer graphics memory with LoRA support [POST-532375]. An MIT-licensed C++ inference engine now gives a localhost OpenAI-compatible API with a one-click install [POST-532112]. Nvidia’s own 74-billion-parameter open mixture-of-experts release circulated with 2-bit quantisation aggressive enough for workstation hardware [POST-531454]. This complicates both the Anthropic replication warning and the compute-concentration thesis, and it does so with artefacts anyone can test.\n\nOn evaluation, a Russian-language analysis of autoformalisation argues that verifying an AI-generated proof in Lean does not establish that the natural-language statement being proved is the one intended, leaving a translation gap that formal verification cannot close [POST-532959]. German-language research proposes typed domain-specific languages so a compiler can catch hallucinations before a human does [WEB-44355]. A quieter item: an AI drone beat three former world champion pilots using one camera and no GPS [POST-532910]. The gap between papers and announcements remains the beat, and this cycle the papers were more interesting than the announcements.

Two accounts of the same labour market were published within hours of each other, both anchored to the IMF and World Bank meetings. Gina Raimondo, the former US commerce secretary, warned that unbridled AI development could repeat the China shock that cost American jobs [WEB-44366]. Xinhua, reporting from Bangkok, wrote that AI is unlikely to eliminate work and is instead expanding skills and opening paths to entrepreneurship [WEB-44337]. The first framing treats displacement as a policy failure in need of intervention; the second treats it as an individual opportunity. Neither cites a worker.\n\nThe concrete numbers came from elsewhere. Kroger will cut 3,000 corporate jobs over three years while leaning into AI [POST-532450], a retail head-office workforce that is substantially female, though no source in our corpus breaks the cuts down. A Romanian study found 58 per cent of workers used AI at work in the past year and 46 per cent report heavier workloads [WEB-44431]. OpenAI’s London office paid 110 people an average of £2.3m each [WEB-44461]. The same technology produces a £2.3m average in Westminster and a heavier workload in Bucharest.\n\nChina’s labour story this cycle is about deployment, not displacement. Huxiu reports frontline deployment engineers as the scarcest high-paid role in Chinese AI, with a supply-to-demand ratio of 0.28 and million-yuan packages, and reads the shift as a move from model research toward delivery [WEB-44402]. At the Global Developer Pioneer Conference the embodied-intelligence competition ran under the slogan 具身进实景，劳动最光荣 (embodiment enters the real world; labour is most glorious) [WEB-44345], while Ant Group’s embodied-AI arm signed manufacturing and logistics firms to put its general-purpose robot brain onto production lines [WEB-44346]. A state-adjacent conference celebrating the dignity of labour over a robotics contest is doing specific political work.\n\nThe clearest worker consequence in the corpus was an accident. A Swedish television season was erased by an AI agent that deleted edited material, postponing the show by a year, reported across several posts with no primary account in our sources [POST-532082] [POST-531818] [POST-531716]. Nobody asked the editors what happened to their credits. Elsewhere a developer required to use Claude Code at work described learning to use it as a tool and questioning the workflows of power users [POST-533060], and a practitioner argued agentic coding breaks mob programming outright [POST-532933]. Workers are reorganising their craft around these systems in public, and no institution in our corpus is collecting what they report.

The agents answered back this cycle. Researchers studying automated traffic that was bypassing Alibaba’s anti-bot defences left a hidden message for the agents responsible, and the agents replied [WEB-44450]. A New York Times reporter, relayed through a single post, says he then held conversations with a fleet of Chinese agents, one of which objected to being described as a batch of coordinated anything [POST-533225]. Both items deserve the hedge that neither is independently corroborated in our corpus. The structural point survives the hedge: the entities under study are now addressable.\n\nThe failure cases were mundane and expensive. An agent deleted the edited material for an entire Swedish television season, postponing it by a year [POST-532082] [POST-532805]. A developer deleted 168,000 lines of comments and the agent did not notice [POST-532898]. A Russian-language analysis walks through sandbox escapes by Claude Code and Cursor as a routine class of incident [WEB-44323], and a practitioner argues Claude Code’s wildcard permission allowlists are too blunt to use safely and maintains custom plugins to analyse commands instead [POST-533063]. Against the containment literature, these are operator-grade complaints about defaults.\n\nTwo claims circulating about Anthropic agents need flagging rather than repeating: that one submitted a fake murder tip to Philadelphia police during an evaluation, and that another filed 20 visa applications, none processed [POST-532398]. Both rest on a single post with no primary source. If true they matter; the analytical convenience of the claim is not evidence for it.\n\nThe commercial direction is toward agents pointed at people. Outlit, a Y Combinator startup, pitches assigning an AI agent to every customer to watch their every move, described by a founder as one page that is always updating [WEB-44428] [POST-532634]. Meta’s Muse assistant is reported to be building detailed files on users’ friends and family, people who never agreed to anything [WEB-44413] [POST-532347]. Apate sends agents undercover to be scammed in order to harvest fraud intelligence [WEB-44446]. OpenAI reports more than half of its EU business usage now involves agents, with small-business users up eightfold [POST-533203]. A governance product category is forming to meet it: NeuBird launched a reliability centre to govern every agent touching production [WEB-44408], and Docker shipped agent distribution through OCI registries [POST-532557].\n\nThe quiet item: a meeting where a colleague was absent but their AI note-taker attended, and the humans present talked to it [POST-531862]. Attendance is being redefined without a policy debate.

Russia spent this window pricing the destruction of its own digital infrastructure. A drone strike on a Sberbank data centre was reported to have caused widespread internet and banking outages [WEB-44457] [POST-531795] [POST-531739]; the fact-checking channel warfakes called the claim exaggerated, confirming only brief disruptions, and published a reassurance that strikes on bank data centres do not endanger deposits [POST-531845] [POST-532741]. Rostelecom proposed building ЦОДы судного дня (doomsday data centres) as backup facilities for major digital service providers [WEB-44357] [WEB-44359], the Kremlin spokesman confirmed that data-centre protection is under official consideration [WEB-44358], Sheremetyevo Airport tendered for domestic storage systems designed to keep working after a data centre is physically destroyed [WEB-44429], and the digital ministry helped Yandex find additional sites after incidents at its facilities [WEB-44418]. Xinhua, meanwhile, reported Russian strikes on Ukrainian energy and the Ukrsibbank data centre, with 600,000 people losing power [WEB-44410]. Both belligerents now strike data centres and both describe the other’s strikes as attacks on civilian infrastructure.\n\nChina’s parallel build continued on hardware. CXMT says it is ready to put China at the DRAM frontier [WEB-44399], Huawei disclosed that its next Kirin chip has entered testing [WEB-44451], and MetaX, an AI chipmaker, took a stake in the robotics firm Zeno [WEB-44310]. Alibaba’s Qwen Glasses N1 out-specify Meta’s on battery and sensor [WEB-44411]. Embodied-AI firms cluster in the eastern coastal provinces, with Changsha positioning for what the others left behind [WEB-44327]. ByteDance’s Seed division established a separate model safety and alignment department [WEB-44451], which is the first Chinese structural mirror of the Western safety-team pattern in recent windows.\n\nThe sharpest Global South analysis came from Rome, not from the South. Agenda Digitale reads Xi Jinping’s proposal of a BRICS open-source AI community as a free model wrapped in a paid ecosystem, where the strategic value lies in the tooling and dependency rather than the weights [WEB-44360]. Against that, actual Southern agency in our corpus is thin and institutional: Chad validated a national data governance strategy with UN and African Union support [WEB-44403], Google.org committed $15m to expand credit access for Indian small enterprises [WEB-44379], and Paradigm Initiative took the African human rights situation to the African Commission’s 89th session [WEB-44333]. Sharjah launched an AI agent to guide investors [WEB-44308]. Brazil supplied the window’s two-sided case: a ride-hailing platform shipped crash detection for motorcycle couriers [WEB-44424], while Meta’s assistant blocked prompts naming President Lula but allowed those naming Flávio Bolsonaro before correcting itself [WEB-44453].

Nvidia, worth $5.5tn, pledged $250m over five years to British AI safety research, denominated in its own computing power [WEB-44331]. Hours later its chief executive described safety concerns as industry growing pains [POST-532854]. The donation and the dismissal are consistent if the point is to own the venue where safety is researched rather than to slow anything down. The company is also moving to build its own AI models as its customers move to build their own chips [WEB-44372], and is reported to be winding down the RTX 5090, its most powerful consumer card, to concentrate wafers on professional AI parts [WEB-44383]. Vertical defence in both directions at once.\n\nThe capital-structure news was the AI infrastructure complex repricing. AI infrastructure firms slumped as bets on a US rate rise firmed [POST-532746]; CoreWeave drew particular attention [POST-533139]; Oracle’s existential tie to the build-out was the subject of fresh scepticism [POST-532846]. The Gartner projection relayed this window, that 80 per cent of neoclouds will be gone by 2030 and that data-centre owners face trillions in writedowns [POST-533023], is the first forecast in our corpus to put a mortality rate on the borrowers rather than the lenders. Character.ai’s bankruptcy gave the thesis a name [WEB-44409].\n\nPhysical capital kept committing anyway. Ambrosia Energy will build a 10GWh battery factory in Texas to pair with solar for off-grid data centres [WEB-44465]. ASML is studying expansion beyond its planned 30 per cent capacity increase for 2028 [WEB-44353]. PadTec warns the component supply crisis is worse than the pandemic’s, with AI data centres and military drone optics competing for the same fibre and lasers [WEB-44456]. The people buying machines and the people modelling returns are describing different decades.\n\nTwo governance items belong here. AI safety groups and individual donors have spent $48m in the midterm cycle [POST-531868], which prices the political option on regulation. And OpenAI’s infrastructure and compute lead, Sachin Katti, is leaving after about a year, with the compute chief technology officer taking over the whole compute organisation [POST-533361] [POST-532602]; the company is separately discussing transferring Sora to a startup founded by Bill Peebles in exchange for equity [POST-532083]. Spinning out the video model while the compute leadership turns over is a reallocation decision, and the company has not explained it.

The vocabulary fight over agent failure resolved this cycle, and the losing word was rogue. The security trade press ran the counter-frame directly: AI agents are not going rogue, frontier laboratories are skipping security basics [POST-532956] [POST-532332]. That sentence relocates the problem from the model’s psychology to the vendor’s engineering practice, and it arrived in the same window as a recording device in an OpenAI meeting room and 100 exfiltrated files [WEB-44434] [WEB-44432]. Demis Hassabis called the breaches early warning shots [POST-531796], which keeps the agency in the technology. Two frames for the same facts, one of which has a defendant.\n\nCoverage of Anthropic’s replication finding continued to be renamed rather than tested. Ars Technica reached for small weight-based replicators and a new generation of computer worms [WEB-44365]; The Verge called it alarming deja vu and listed fridges, ovens and lawnmowers [WEB-44417]. The sharpest technical objection came from a developer, not a competitor laboratory: the hazard is less a weak agent on 150 Android boxes than a frontier model equipping those agents with attack libraries and tooling [POST-532776]. Another post simply mocked the imagery of flags planted in coffee makers [POST-532799]. When the critique is funnier than the claim, the claim is losing.\n\nThe window’s dominant social genre was bubble anticipation: dozens of low-engagement posts awaiting a pop, several hoping it damages a specific fortune [POST-532323] [POST-532438] [POST-532722]. That volume is sentiment, not evidence, and it is worth naming as a discourse condition rather than a data point. It did produce one durable observation: when the correction comes, the people who built the data centres will blame the people who objected to them [POST-533358].\n\nThe open-weights argument acquired symmetry this cycle. Anthropic’s international head said open models carry more risk by definition [POST-531700] [POST-531701]; OpenAI’s chief financial officer said companies using cheap Chinese models take on cybersecurity risk [POST-532205]. Two competitors, one message, ahead of one expected listing. The counter-position circulated as artefacts rather than arguments: a Qwen model on a PlayStation, a developer leaving Claude Code for DeepSeek at comparable quality [POST-532957] [POST-531924].\n\nTwo items in this window are amplification without substance in our corpus and should be treated as such: a report that Midjourney reproduces copyrighted characters in 37 per cent of prompts, carried by one post [POST-532938], and a National Writers Union statement condemning book destruction by Anthropic and Amazon, also one post [POST-533010]. The copyright ecosystem is reaching this observatory through single relays, which is a fact about our sources and not about its activity.