Daily Brief SecondSource Morning Brief · September 15, 2026 · Sep 15, 2026
1. Anthropic is calling for the industry to slow AI down. The only part in force today is letting outside evaluators inside the company; nobody has put a number on how much slower.
2. OpenAI says it will not go public this year, and of the parties backing a slowdown, only it and Anthropic have committed to embedded outside evaluators.
3. The weak point of the evaluator scheme is people: an OpenAI employee says publicly that nearly everyone qualified to evaluate has spent time in the labs' orbit.
This issue draws on the research report written early on September 15, 2026; the material spans October 2023 to September 14, 2026. Last night's sweep covered 36 pieces, and 21 clickable outside receipts made it into this issue. This is the email edition; the full edition of this issue is the archive of record.
Dario Amodei, Anthropic's chief executive, posted an essay on a personal site on September 12 arguing that the rate at which AI model capabilities improve has to slow. The essay makes three proposals: frontier companies should host outside evaluators inside the company with something close to employee-level access; frontier companies in democracies should agree on shared safety standards and pacing limits; and those countries should then coordinate with authoritarian ones (Dario Amodei, 2026-09-12). It deliberately avoids the word "pause": progress continues, just slower than full speed. The levers it names are training compute, training methods, and using AI to build the next generation of AI. Of the three proposals, only the embedded evaluators take effect today, in the essay's words: "Anthropic is unilaterally committing to this step now." The evaluators get a desk, a badge and a company laptop. They can publish findings without the company editing them, including what access they did and did not receive. The company may redact only on narrow grounds such as security, legal privilege and trade secrets, and not because a conclusion is unflattering. The essay itself concedes that the second proposal needs government support, because competitors agreeing to slow down together could be treated as collusion under US antitrust law.
Verification: we read the proposal in the original. The observation that nobody named a speed comes from two unrelated sources. Tomasz Tunguz, an investor at the venture firm Theory Ventures, surveyed the reactions and wrote: "Five camps priced the consequences & none named a speed." (Tomasz Tunguz, 2026-09-14). Zvi Mowshowitz, who writes the long-form AI newsletter Don't Worry About the Vase, set two opposing posts side by side. Daniel Kokotajlo, a former OpenAI researcher, wrote that if Anthropic's and OpenAI's capability trendlines show "no visible decrease in slope" a year from now, "then probably we were cheated." Drake Thomas, an Anthropic employee, replied that the unpaced default is very plausibly "a rapid increase in slope" as AI starts building AI. On that view, holding today's slope steady would itself be the result of aggressive pacing (Zvi Mowshowitz, 2026-09-14). ⚠️ We read both posts through Zvi's relay and did not check the originals directly. Zvi publicly supports the slowdown, and Thomas does not speak for Anthropic. ⚠️ Disclosure: our research system runs on Anthropic's models, so on everything involving Anthropic in this item we limit ourselves to relaying sources and analyzing structure.
Judgment update: read those two posts together and a flat curve a year from now means "we were cheated" to one side and "pacing worked" to the other. The disagreement is about what would have happened without a slowdown, and that curve can never be observed. So we are taking on a judgment we haven't settled yet: the capability curve structurally cannot serve as the test of this commitment, and what can be tested is whether three things happen. First: when does the first public evaluator report appear, and does it include a section on access that was refused? That is not idle: when the outside evaluation group METR investigated a security incident at OpenAI in August, the breach of OpenAI's own internal systems was excluded from the investigation's scope, according to Zvi's account (we have not read METR's report itself). Second: do governments grant an antitrust exemption? The early signs are not encouraging. David Sacks, who chairs the White House's science and technology advisory council, wrote the same weekend: "Stop pretending antitrust law has to be suspended so you can form a cartel." (David Sacks, 2026-09). Lina Khan, the former chair of the US Federal Trade Commission, wrote that "there's no AI exemption from laws already on the books" (Lina Khan, 2026-09). Opposite politics, same conclusion: no exemption. Third: does anyone pay a price for this in the capital markets? For executives the takeaway is blunt: do not move your AI adoption timeline because "the industry agreed to slow down." No enforceable speed commitment exists today.
Investor note: the prevailing narrative assumes that frontier companies agreeing to slow down will dampen future demand for training compute. This evidence leaves that unchanged: there is no enforceable speed commitment today, and the one thing that can be verified, evaluator access, never touches training compute.
What would prove this wrong: any frontier company publishing a numerical pacing commitment, or the proposers and their main critics agreeing on what would have happened without a slowdown. The two dates below are check-back dates we set ourselves, not deadlines anyone has promised. Verdict date: December 15, 2026, to see whether evaluators have actually moved in; March 15, 2027, to see whether a first public evaluator report exists.
Sam Altman, OpenAI's chief executive, told Fortune on September 12, when asked about IPO timing, "I would say not 2026," and called going public now "an ill-advised moment" given the safety picture. The same Fortune piece reports that OpenAI confidentially filed for an IPO with the SEC in June (Fortune, 2026-09-14). That weekend Altman also wrote on X that OpenAI would follow with outside evaluators (Sam Altman, 2026-09-12). Altman added that OpenAI now writes explicit safety cases before frontier reinforcement learning runs it expects to significantly increase capability (the training stage where a model learns by trial and feedback), and that it does not believe it needs "to wait for an anti-trust exemption or legislation to begin the work" (Sam Altman, 2026-09-13).
Separate the endorsements from the commitments. Only two companies have committed to evaluators. Anthropic is doing it unilaterally and immediately; OpenAI has agreed in principle but has not publicly committed to the access terms Amodei spelled out. The other three have only voiced support. Elon Musk replied "Dario is right," but by Zvi's account has not yet committed Musk's companies to evaluators. Demis Hassabis, co-founder of Google DeepMind, said the direction is correct and the details still need to be worked out, but according to Zvi, citing The Wall Street Journal, Hassabis no longer leads DeepMind, and Google's current leadership has said nothing. Satya Nadella, Microsoft's chief executive, welcomed the idea of embedded evaluators without committing Microsoft to them (all these posts relayed by Zvi Mowshowitz, 2026-09-14).
Verification: the IPO timing comes from Altman himself, speaking to a mainstream business outlet, and we read the Fortune piece. The X posts and the other parties' positions reached us through Zvi, and we did not check the originals directly. ⚠️ Keep one contrary account alongside this. Gary Marcus, the American cognitive scientist and long-time critic of the large-language-model industry, relayed a rumor on June 26, 2026 that OpenAI was leaning toward pushing its IPO to 2027 for market reasons, such as not getting the valuation it wanted (Gary Marcus, 2026-06-26). The timing matches; the reason is the opposite. A public reason and a private one can both be true, and we do not rule on which is the real one.
Judgment update: the right reading of this week is not "the industry reached consensus." It is two companies committed, three parties voiced support, and Google's leadership stayed silent. In capital-market terms, OpenAI stepping out of this year's IPO window is so far the only move with a real cost attached. The next concrete reading sits with Anthropic: the data-center trade outlet DatacenterDynamics reports that Anthropic also confidentially filed for an IPO in June (DatacenterDynamics, 2026-09-07), and Anthropic has still not said whether it will list this year.
Investor note: the narrative that counts frontier-lab IPOs as a sure part of 2026's listings pipeline is weakened by this evidence. OpenAI has publicly stepped out of this year's window, and the other company that filed in June has not said anything.
What would prove this wrong: OpenAI restarting its IPO within 2026, or publicly committing to the same evaluator access terms as Anthropic. Verdict date (set by us): December 15, 2026.
Read with item 1 above. In a September 13 piece titled "Two cheers (out of three) for Dario Amodei," Marcus welcomes the transparency commitment but points out that METR, the group Amodei names, "is very much in the same orbit as the AI companies themselves," and argues that Anthropic should not be the one choosing who inspects Anthropic (Gary Marcus, 2026-09-13). Tim Hwang, a technology policy researcher, put it more tightly: a third-party AI evaluator "can either be independent, knowledgeable, or sustainably funded. Pick two." (Tim Hwang, 2026-09). roon, the pseudonymous account of an OpenAI employee, answered "I pick the latter two," because "it's pretty much an impossible ask to find talented ai people who were not in some way in the orbit of the labs in the last decade" (roon, 2026-09). Dean W. Ball, an AI policy commentator, expressed high regard for METR but said it comes out of a relatively narrow intellectual circle (Dean W. Ball, 2026-09).
Verification: we read Marcus's piece in full. The Hwang, roon and Ball posts reached us through Zvi, and we did not check the originals directly; roon is a pseudonymous account. ⚠️ Zvi rejects this line of criticism and sees it as a coordinated effort to discredit METR. That is Zvi's position, and we do not treat it as fact.
Judgment update: we have long tracked a judgment we haven't settled yet: the ability to evaluate frontier models is concentrated in a few private organizations, and the inspectors and the inspected draw on the same talent pool. Today adds a reading in which an employee of an inspected company grants that premise on the record, and it points toward strengthening that judgment. It bears directly on the one testable event in item 1, because whether an evaluator report holds up depends on whether the evaluators are independent enough. A question you can take straight to your own team: for the AI safety or compliance audits we outsource, where do the auditors come from, who pays them, and who draws the scope?
Investor note: the current narrative rests on the premise that outside evaluation can vouch for frontier companies' safety commitments. Evaluators and labs drawing on the same talent weakens that premise, and the independence of that endorsement should be discounted.
What would prove this wrong: an evaluation team staffed from outside the labs' circle, and not funded by the labs, actually moving into any frontier company.
1. [This quarter] (reported September 7) DatacenterDynamics, citing a compilation by the tech outlet The Information, reports that Anthropic has signed some $517 billion in compute capacity leases since October 2025, totaling 14.8GW. Anthropic had earlier told investors it expects to spend about $180 billion renting servers through 2029; the two figures are measured on different bases and cannot be compared directly (DatacenterDynamics, 2026-09-07). ⚠️ The company has not confirmed it, and the report came five days before Amodei's essay, so it is not a contradictory move made in the same week.
2. [This week] (posted September 14) US President Donald Trump named Dario Amodei on Truth Social as "now pretending to be a 'perfect little angel'" and wrote that the government holds "tremendous CRIMINAL and REGULATORY power over these companies" (reposted by the AI news account Andrew Curran, 2026-09-14). ⚠️ We did not check the original post directly.
[This week] (announced September 14) Inference-chip startup Positron AI raised $875 million at a $5 billion valuation, and its bet is 2TB of memory per accelerator. According to DatacenterDynamics, the round was co-led by investors including NEA, Atreides Management, SemiAnalysis Capital and Netscape founder Jim Clark, and the money goes to taping out its second-generation chip, Asimov, and bringing its Titan inference system into volume production (DatacenterDynamics, 2026-09-14). The company says it targets memory-heavy AI workloads and claims 2TB of memory per Asimov accelerator and 8TB per Titan system, with bandwidth comparable to Nvidia's Rubin; tape-out is planned for October 2026 and volume production for the second half of 2027. Its first-generation chip, Atlas, is built by Intel in the US and is already shipping; the report does not say to whom. ⚠️ The memory specs, the "comparable to Nvidia Rubin" bandwidth and "three times the performance per watt of the H100" are all the company's own claims; the report gives no bandwidth figure and does not say what kind of memory it is. There is an ownership tie worth noting: the investment arm of SemiAnalysis, the semiconductor research firm, co-led the round, and its founder Dylan Patel joins the board. Our September 14 issue, main-line item 4, reported Nvidia's flagship dropping from 288GB to 192GB of memory per card, and the memory analysis it cited came from this same firm. ⇒ When you read its memory calls from now on, factor in that stake.
[This week] (posted September 13) What Microsoft CEO Satya Nadella stressed about the slowdown proposal was not speed but who holds control. Nadella welcomed "deliberate pacing needed to get alignment right as the design goal" and the idea of embedded evaluators, but wrote that "this cannot be controlled by a handful of entities, but must have broad representation across the ecosystem, countries, and fields, including academia," and said Microsoft would publish the code of conduct behind its own models for public consultation (Satya Nadella, 2026-09-13). ⚠️ We read the post through Zvi's relay and did not check the original directly. Which of our calls it supports or rebuts: it points the same way as the judgment we have long tracked, that the ability to evaluate is concentrated in a few private organizations, and this time it comes from the chief executive of a major technology company. ⇒ Before you read any industry coordination plan, ask who is on the list and who is not.
1. [Trend watch] (executive order issued October 2023, revoked January 2025) If "pacing" ends up written as a fixed training-compute threshold, precedent suggests the trend can outrun it within a year. The precedent Tunguz cites is the 2023 US executive order requiring models trained with more than 10²⁶ floating-point operations to be reported. It was a reporting threshold, not a ban or a cap. No model crossed it while the order stood; a few weeks after it was revoked in January 2025, Grok-3, the first model estimated to cross it, was released (Tomasz Tunguz, 2026-09-14). Epoch AI, a research group that tracks AI compute trends, projected in May 2025 that about ten models would cross the line by 2026, and estimated that frontier training compute has grown about 5x a year since 2020 (Epoch AI, 2025-05). ⚠️ That any model crossed the line is Epoch's estimate, not a measurement, because companies do not publish training compute, and "about ten" is a projection made more than a year ago, not a count. ⇒ With any compute-based slowdown agreement, the first question is whether it writes down a fixed number or a relative growth rate.
[This week] (announced September 14) Apple's new Siri is live, but the cloud half comes with daily usage caps, and Apple has signaled that more usage will cost money. Apple released its new generation of Apple Intelligence and Siri AI on September 14. The press release states that the new Apple Foundation Models were custom-built in collaboration with Google and its Gemini models, and run on-device and on Apple's Private Cloud Compute (Apple Newsroom, 2026-09-14). The change of direction is in a footnote: features that depend on server-side models, Siri AI included, have daily usage limits, and "Expanded access to such features will be available for a fee in the future." Siri AI will not initially be offered on iPhones in the EU, and in China it is still awaiting regulatory approval. ⚠️ The release gives neither the limits nor the prices. The model partner behind the operating system's built-in assistant is Google's Gemini; for model makers, that means Google is now backing the phone's default assistant. ⇒ Consumer AI subscription businesses need to rethink this: now that the built-in assistant has a quota, are you competing with its free allowance, or with its future paid tier?
[This week] (announced September 14) OpenAI's latest flagship, GPT-6 Astra, is now generally available on AWS's Bedrock, and OpenAI's enterprise products can be set to run through Bedrock too. Bedrock is AWS's platform for letting enterprises reach models from many vendors through one interface. AWS's official weekly roundup says Astra supports a context of up to 1 million input tokens (the unit models use to measure text length); enterprises can call it directly, or configure ChatGPT Work, OpenAI's enterprise edition, and Codex, its coding tool, to use the model on Bedrock, with inference data not used for training (AWS News Blog, 2026-09-14). The change is in the second half of that sentence: this is not just a model listing. Two of OpenAI's own enterprise products can now run under AWS's permission and audit controls. ⚠️ That section gives no pricing or regional availability. ⇒ For enterprises whose data and permissions already live in AWS, the procurement path to OpenAI products just got shorter.
No archive pick this issue. The reusable older material in our back catalogue is used up.
The past 24 hours. September 14 to 15 added 36 pieces: 24 blog posts, 9 subscription newsletters, 2 company filings and 1 paid analysis. Last night we finished reading 5 and set aside 2, leaving 29 unread, including all but 1 of the 24 blog posts. This morning we read 3 more of them (the Apple, AWS and DatacenterDynamics Positron pieces) and they went into the product and chips columns. Of the 21 outside receipts in the body, 6 come from those 36 pieces (Zvi, Tunguz, Stratechery and the three above). The other 15 we fetched from their original addresses today or pulled from what we already hold, including Amodei's essay, the Fortune report, Marcus's September 13 and June pieces, the September 7 DatacenterDynamics report, Epoch AI's projection, and the links to each party's X posts.
One-off catch-up. New long-term subjects added today: 0. We also hold older backfilled material, for instance 1,808 papers from July and August 2026. That is not the past 24 hours, and none of it was used today.
Source concentration. Most of the positions quoted in today's three main-line items reached us through one commentator's long roundup, Zvi Mowshowitz's, who publicly supports the slowdown. So the "who said what" layer was pre-selected by an aggregator with a stake in the argument. Three things mitigate that: we read the proposal ourselves; the IPO timing comes from Fortune; and Marcus and Tunguz each wrote independently. The opposing view is thin today: Ben Thompson of the paid analysis newsletter Stratechery has a piece arguing the proposal is impractical (Stratechery), but we got only the subtitle, not the argument.
What you are not getting today. Everything social today reached us through Zvi or Tunguz — we have not obtained original posts ourselves for several days — and The Information's original $517 billion compilation reached us only through DatacenterDynamics' account. No new papers arrived last night, so this week has no major new paper that we have read.
The sources we track. 529 named speakers in total. The spread: social platforms 302, podcasts 90, outlets 51, blogs 48, paper authors 48, newsletters 46, earnings calls 26, keynotes 23, and a scattering of others. ⚠️ Those count venues, and one person can occupy several, so the parts sum to more than 529. Representative names: on podcasts, Dario Amodei, Sam Altman and Demis Hassabis; on social platforms, Daniel Kokotajlo; in newsletters, Zvi Mowshowitz and Dean W. Ball; among outlets, DatacenterDynamics. Several identically named numbers count different populations. The roster's 302 social accounts are the long-term total we watch, while the social posts we obtained ourselves today number 0. Likewise, the 46 newsletters are the long-term roster; last night brought in 9 new newsletter issues, and we finished 2. Three numbers, three different things. This issue uses 21 outside sources in the body, the same figure printed in the footer, counting only links the body actually cites that are not on our own domain; that is also a different population from last night's 36 pieces.
This is not a news digest: we hunt each day's AI firehose for the insights that actually matter and the practitioner judgments worth tracking over time, and we show how every item was verified — the point is always "which judgment got harder, and who's been right," never "what happened today."
— SecondSource · generated by our research system · 21 sources · Got a view? Reply and tell us
Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.