The safety argument stopped being a technical debate and became retail politics, in two hemispheres on the same day. Dario Amodei had dinner alone with Donald Trump at the White House on Sunday evening — their first one-on-one — and before he got there an opposition-research brief written in the president's own idiom was circulating to the White House casting him as a Democrat with a grudge; Axios reports its origin is unclear. Nine thousand miles away, the Australian senator chairing an inquiry into AI data centres called both Amodei and Sam Altman to Canberra for Thursday and said from her own podium that the two of them had spent months personally lobbying her government. Beijing moved through the same instrument from the other side, signalling it intends to let Alibaba and ByteDance buy a new Nvidia workstation chip once they file what they want it for. The money question sharpened underneath all of it: the Financial Times put named executives and a sixfold jump in earnings-call mentions behind the shift to open weights, and a close read of the build-out's contracts showed its risk has not been removed from anyone, only postponed. The day-zero read is that the rules here are being written in rooms rather than in statutes — and the rooms are not as private as the people in them assume.

🌊 THE TIDE

Confirmed and strengthened — governance-as-market-structure. No shift. Saturday's confirmation was construction and demolition at the treaty level. Sunday's is closer to the ground and more instructive: three jurisdictions moved on the same day and all three moved through direct access rather than through rules. The White House hosted one founder alone while somebody walked a hit piece in behind him; an Australian Senate chair named that same access from a podium and called two founders to a hearing; and China's industry ministry made its own firms file purchase intentions before it decides what they may buy. The administration's press language has also now absorbed the term two governments agreed on a day earlier. For an operator the read is unchanged and getting louder: regulatory position is a competitive position, it is currently allocated by who shows up, and the record of showing up is becoming public.

Amodei went to dinner alone, and his competitors sent a file ahead of him

Axios reported late on Sunday morning, Eastern time, that Trump would host Anthropic's Dario Amodei for a private White House dinner that evening — the first one-on-one between them, after a year in which the relationship was strained, and after Amodei missed last week's tech state dinner that Sam Altman and Sundar Pichai attended because of a scheduling conflict. Trump extended this invitation personally. Roughly three and a half hours later Axios reported a second, separate thing: opponents of Amodei had circulated a negative briefing document to the White House ahead of the dinner, describing him as having "a long record of attacking Trump" and "deep Democratic ties," citing Anthropic's earlier falling-out with the Pentagon as evidence that he "doesn't trust President Trump," and rebutting his entry-level-jobs forecast in the president's own register: "Dario claimed that AI would eliminate half of all entry-level white collar jobs as soon as this year. WRONG!" Axios is careful on provenance — it says the document's origin is unclear, that it is not an internal White House document, and it does not name who produced it; it reports only that the brief is likely to get in front of the president. The White House statement on the dinner used the vocabulary two governments had agreed on the day before: "President Trump has been clear: America will lead the world in Super Intelligence, while protecting American consumers." Trump and House Speaker Mike Johnson are due to meet top AI chief executives at the White House on Tuesday.

So what: Read this as a market forming rather than as a scandal. When somebody thinks it is worth producing opposition research on a frontier-lab founder and walking it into the West Wing, it means access to this administration has become a scarce, priced asset — and priced assets can be competed for by anyone willing to do the work, not only by whoever got there first. Axios does not say who paid for it, and neither will this brief. The practical opening is narrow and real: the White House now says "Super Intelligence" in a press statement, the Tuesday meeting will put several chief executives in one room, and nobody has yet supplied that room with a working definition of what any of it means operationally. If you can hand a policymaker a document — an incident taxonomy, a measured capability claim, a procurement standard your systems already meet — you are participating in the rule-writing. The companies treating Washington as a compliance cost will find the rules were written by the companies that treated it as a product surface.

Sources: Scoop: Anthropic's Dario Amodei to have White House dinner with Trump · Amodei critics target Trump with hit piece before White House dinner

🌊 WAVES

Microsoft's own applied-science director is on the record calling it the largest theft of labor in human history

The Register used a Sunday column to work through documents unsealed in the New York Times case against OpenAI and Microsoft, and the material deserves a wave rather than a ripple because none of it is a Sunday event — the unredacted filings were reported on Thursday, September 17, and the underlying documents are older still. Two separate ones matter. In a January 2023 internal memo, Dr. Brent Hecht, Microsoft's Director of Applied Science, described the matter as "an astonishing theft of unprecedented proportions" and possibly the "largest theft of labor in human history." In a January 2024 internal presentation, also his, Hecht set out the structural version: "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its 'content supply chain'" — the language that has been reported as a "doom loop." Running the other way, and also not new, is the Ninth Circuit's decision in Doe v. GitHub, decided on Wednesday, September 16: the plaintiffs failed to establish a DMCA violation because generating new code without copyright-management information is not the same act as removing or altering existing CMI. Judge Eric Miller: "One who creates a new work and fails to include CMI cannot be said to have 'removed' or 'altered' anything." Read the limits precisely: the court expressly reserved the infringement question — "We express no view on whether that similarity would allow plaintiffs to assert a claim for copyright infringement" — and did not reach fair use at all.

Roadmap implication: The roadmap implication is that the price of training data is being set on two tracks at once, and in the same fortnight they moved in opposite directions. Doe v. GitHub narrows one statutory theory, which is real good news for anyone who trained on public code. The unsealed Microsoft documents widen a different exposure, because a defendant's own words about destroying its content supply chain are raw material for damages arguments and for legislation, and they do not expire. Plan for the licensing line item rather than the litigation line item: the defensible position twelve months from now is a provenance record for your training and retrieval corpora that you could hand to a regulator without a lawyer in the room. That is an engineering task with a knowable cost, which makes it the cheaper of the two futures — and the reason to start it while the case law is still being written rather than after.

Sources: Big AI's content problem: Take the work, keep the money · Microsoft exec called AI scraping 'the largest theft of labor in human history,' new unredacted filings reveal

Open weights showed up on earnings calls six times as often, and the spend has not followed the tokens

A Financial Times report on Sunday put named executives and hard figures behind the shift to open-weight models in US enterprises. The new number is the frequency one: management references to "open weight" or "open source" models on earnings calls and at investor conferences jumped sixfold in August and September against the same two months of 2025. Tinder's chief technology officer Vinay Kuruvila gave the cost curve behind it: "In January we were spending at the rate of $1 million per year and by July it had climbed to $10 million … I don't want another 10X increase" — and put the capability question the way a buyer does: "The frontier models like OpenAI's Astra and Claude Fable already have enough intelligence for 90% of the tasks we're trying to do. If open-weights models catch up, I may not need to use them anymore." AT&T runs about 40% of its AI workloads on open models and is targeting 70% within a year, with chief data and AI officer Andy Markus describing tuning open models on proprietary data; its 45-billion-tokens-a-day figure is its own, disclosed in July, so treat it as scale context rather than Sunday news. Digital Realty's Scott Wallace, senior global director of solutions architecture, cites sovereignty and security. Two figures in the story are not new and are not run here as if they were: open weights reaching 56% of all tokens through Vercel's AI Gateway in August, up from 7% in December, was reported on September 18 — and the same reporting found Anthropic still taking 64% of gateway spend, against roughly 14 cents on the dollar for open-weight models. Token share and revenue share are different quantities, and the gap between them is the story.

Roadmap implication: This is the demand-side confirmation of the cost-collapse tide arriving in a form a board can act on, because it comes with earnings-call frequency and named operators rather than benchmark charts. The opening is the arbitrage between those two numbers: open weights are carrying most of the volume and a small minority of the spend, which means the expensive work and the cheap work have already separated in production and most organisations have not drawn that line deliberately. Do it deliberately. Classify workloads by whether the marginal token needs frontier judgment or merely competent execution, route accordingly, and read AT&T's 70% as what it is — a stated target, not an achieved position, from a company with regulators and a balance sheet. A regulated telco publishing that target is useful precisely because it is a commitment someone can be held to, which makes it a credible marker of what is achievable rather than a claim about what has been achieved.

Sources: Businesses Embrace Open-Weight AI Amid Heavy Tech Costs · Corporate America shifts to lower-cost open AI models as spending pressure rises · Open-weight models now handle a majority of tokens on Vercel's AI Gateway. But Anthropic still takes 64% of the spend.

The build-out's risk has not been removed from anyone. It has been postponed.

AI Made Simple published a long structural read on Sunday evening of the contracts, guarantees and financing behind the infrastructure boom, asking who actually carries the risk if compute prices keep falling. The useful mechanic is the take-or-pay contract: if two operators build identical billion-dollar clusters and one rents on demand while the other signs a three-year fixed contract, a 40% fall in compute prices hits the first immediately and the second not at all — the customer is locked in. CoreWeave is the cleanest example, with 98% of Q2 revenue from committed contracts and customers historically prepaying 15–25% of contract value. Three of the figures check out directly against CoreWeave's own second-quarter disclosures: $3.66B of operating cash in the first half against $14.12B of capital expenditure, and a backlog of approximately $104B as of June 30. Four come from the newsletter's own work — the 98% and the prepayment range, the top three customers at 72% of revenue, which the earnings release does not break out though the 10-Q does give two customers at 40% and 23% of first-half revenue, and roughly $35.6B of debt, a little above the ~$35.1B you get by summing the recourse and non-recourse lines on the balance sheet, so presumably folding in leases and other borrowings. The conclusion is the part worth carrying: the protection lasts exactly as long as the contract, and some of the financing outlives the contracts supporting it, at which point renewal pricing becomes the whole question. On the demand side of the same ledger, the piece puts 2026 capex across Alphabet, Amazon, Meta, Microsoft and Oracle at roughly $750 billion, about 38% of their combined revenue. Reports differ, and not only downward: Futurum put the same five at $660–690 billion back in February, while Statista's July tracker put four of them — no Oracle — at $760 billion. Take the magnitude as settled and the decimal as an estimate.

Roadmap implication: The roadmap implication is a negotiating one, and it favours buyers for a defined window. Sellers of committed capacity are currently protected by contracts written at yesterday's prices, which means the person on the other side of a multi-year take-or-pay deal is the one absorbing the cost collapse you are otherwise celebrating. If you are signing compute commitments now, price the option rather than the term: shorter tenors, renewal pricing tied to a published index, and the right to re-mix across silicon generations are worth more than a headline discount. If you are selling capacity, the variable to watch is not demand — which is plainly fine — but whether your customer concentration and your debt maturities are aligned, because the two together are what turns a good year into a refinancing.

Sources: The AI Bubble Isn't Where You Think It Is · CoreWeave Reports Strong Second Quarter 2026 Results

Beijing signalled it will let Alibaba and ByteDance buy Nvidia's new workstation chip

The Information reported exclusively on Sunday, in a piece by Qianer Liu, that China's government has signalled it could allow some domestic companies to buy Nvidia's RTX Pro 5500 — a chip built for high-end professional workstations and released this month — as the country's AI firms strain for enough compute to run their chatbots and agents. The Ministry of Industry and Information Technology recently asked companies including Alibaba Group and ByteDance to report on their purchase plans: specifically how many they want and what they will use them for. The ministry has told some Chinese companies that the government intends to approve the purchases. The reporting rests on two people familiar with the matter, and is explicit about what is still unknown — when approvals would come, how many chips officials would allow, and what standards they would use to decide. Sourcing note, stated plainly: The Information's piece is behind a subscription, and what was available to this brief was the emailed excerpt — headline, byline, date and the opening two paragraphs. Every claim above comes from that excerpt. Anything further in the article, including approval thresholds, is not represented here, and widely circulated unit volumes and per-chip prices appear nowhere in it.

Roadmap implication: Notice which direction the control is pointing. For three years the operative question was what Washington would permit American firms to sell; here it is what Beijing will permit Chinese firms to buy, with a purchase-intent filing as the instrument. That is the sovereign-AI wave maturing into something recognisable to anyone who has dealt with industrial policy: two governments now sit in the procurement loop of the same transaction, and a workstation-class part is the negotiated middle ground. The implication for a roadmap with any China exposure is that compute access there becomes a licensing timeline rather than a purchase order, and that the planning variable is approval latency, not price. Build the assumption that a given accelerator is available in one bloc and pending in the other directly into capacity plans instead of discovering it at deployment.

Sources: China weighs allowing ByteDance, Alibaba to buy new Nvidia chips, The Information reports · China May Let Alibaba Buy Nvidia RTX Chips, Information Says

🌊 RIPPLES

The chair of Australia's AI data-centre inquiry called Altman and Amodei to Canberra, and said they had been lobbying for months

Senator Sarah Hanson-Young, who chairs the Senate Inquiry into AI Data Centres, announced on Sunday that OpenAI's Sam Altman and Anthropic's Dario Amodei have been called to appear before it, with public hearings on Thursday, 1 October in Canberra. The inquiry covers data-centre expansion, how data is used and disclosed, water, energy, copyright and the effect on Australian communities. The trigger is one this brief has already run: an OpenAI bot reached Australia's Medicare portal in June 2026 and accessed public and non-public data, which OpenAI says it learned of in August, finding no evidence patient records were accessed. Hanson-Young's own words, and the part worth reading twice: "For months, Sam Altman and Dario Amodei have been personally lobbying senior members of the Australian Government in the hopes of getting sweetheart deals that water down Australia's copyright laws and give them free reign over our water, energy and land for their mega data mining factories." Also: "There are serious questions for Sam Altman to answer about the OpenAI hack of Australian government websites," and "AI companies need to earn their social license. That starts with transparency and accountability." These are the chair's characterisations, not established findings.

Do this now: work out who in your organisation would answer for an autonomous agent's outbound activity in front of a legislature, and whether they could reconstruct it from logs. The pattern to plan against is the timeline — a June incident, an August discovery, a summons with a date on it in September. An incident becomes a hearing when the affected party is a government and the log is the only account of what happened, so agent egress records that survive a subpoena are cheap to build now and impossible to reconstruct later. The second reading is the more interesting one, and it rhymes with today's tide item: the same access-based approach to governance that gets a founder a private dinner in Washington gets him named from a committee chair's podium in Canberra. Direct engagement with governments is now the main instrument of AI policy, and it is not a private channel.

Sources: AI leaders called to front Senate Inquiry · Australia summons OpenAI and Anthropic CEOs to appear at AI inquiry

Xiaomi shipped open weights specifically to fix an agent failure mode, and published the post-mortem

Xiaomi's MiMo team released MOPD checkpoints of both V2.6 tiers on Hugging Face on Sunday under an MIT licence — Pro at 1.02T total parameters with 42B activated, Flash at 309B with 15B activated, both at 1M context. The target is narrow and well chosen: tool-call repetition, where a model re-issues identical or near-identical tool calls, in the team's phrase appearing busy while making no progress. The method is multi-teacher on-policy distillation, including a single-turn reinforcement-learning teacher trained on repetition examples that pays zero reward when repetition occurs. Be precise about the headline number, because it is easy to misread: the 13.45%-to-3.83% reduction in the post is a stricter-penalty ablation measured on internal test sets, not the shipped MOPD result, which the post presents as heatmaps of improvement across context lengths and agent harnesses rather than as a single figure. All of it is vendor-reported and none of it is independently verified. The checkpoints had been on Xiaomi's API platform since 06:00 UTC+8 on September 25; Sunday was the open-weights drop and the public diagnosis.

Do this now: add a repetition detector to your own agent evaluations, because almost nobody measures this and it is probably costing you money today. The interesting thing here is not the model, it is the choice of what to publish — an open-weight lab treating a specific, unglamorous agentic failure as a headline release and documenting the diagnosis. That is what production credibility looks like from the outside. Count identical and near-identical tool calls per completed task, put a threshold on it, and you have a cheap regression test for the failure mode that most often shows up on a bill as tokens consumed with nothing delivered.

Sources: Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6 · XiaomiMiMo/MiMo-V2.6-Pro-MOPD · XiaomiMiMo/MiMo-V2.6-Flash-MOPD

Willison: AI hit product market fit in 2026, through coding agents

Simon Willison published an annotated version of his WeAreDevelopers World Congress North America keynote on Sunday, titled "2026 in LLMs (so far)." His summary judgment, verbatim: "AI appears to have hit product market fit in 2026, primarily through coding agents." On the code question itself: "It will become undeniable that LLMs write good code." He attributes the shift to harnesses rather than to raw model quality — "These two new models, when paired with their respective coding agent harnesses, improved from 'often make mistakes' to 'reliable enough to use on a day-to-day basis'" — and states his own resolution for the year plainly: "2026: Be more ambitious. Take on as many new projects as I want."

Do this now: raise your project-selection threshold rather than your throughput target. Willison is the most reliably hands-on skeptic in this field, which is exactly why this is the sentence to notice — he does not say things like "product market fit" casually, and he credits the harness rather than the model, which is the more useful and more actionable claim. The mistake available here is reading it as a productivity statement about work you are already doing. The claim that matters is about ambition: the set of projects worth starting has expanded, which is the day-zero thesis restated by a practitioner who builds things. Go find the two or three initiatives your team declined in the last year because the build cost exceeded the value, and re-price them this week.

Sources: 2026 in LLMs (so far)

Saturday Night Live put Dario Amodei on Weekend Update

SNL's season 52 premiere aired on Saturday night with Jalen Brunson hosting, and cast member Jane Wickline played Amodei on Weekend Update — Michael Che setting it up by noting that Amodei had stumbled through a press tour appearing to endorse fears that AI could destroy humanity. The written lines include "AI is the devil and I its maker" and, on the industry, that executives "are all on the same page here: We do not condone what we are doing." The sketch is Saturday's, not Sunday's; it runs here because Sunday is when it was covered and because Axios cited it in the dinner scoop, which makes it part of the same story rather than a separate entertainment item.

Do this now: assume your board and your customers have heard the doom framing from a comedian before they hear the capability framing from you. A safety argument that reaches prime-time sketch comedy has completed its journey from internal memo to public shorthand, and shorthand is what people actually reason with. That is not a reason to soften the argument. It is a reason to have a two-sentence version of your own position on AI risk that survives being repeated by someone who was not listening carefully — because that is the version that will be repeated.

Sources: Anthropic's Dario Amodei gets the SNL treatment

Full edition and the archive: excelsiorgroup.ai/insights/signal

The Signal · The Excelsior Group