The Read

The argument about how fast to build frontier models stopped being a lab memo and became a price. President Trump called the slowdown push a "SICK conspiracy" on Truth Social, then told Jensen Huang by speakerphone on stage at the All-In Summit that loss-of-control fears are "a hoax," Beijing called it fearmongering, and the PHLX semiconductor index closed down 5.9%. That is the governance tide doing exactly what this brief has argued it would do — turning into market structure, in public, now with a ticker attached. The more useful signal ran underneath the noise: Nvidia, Palantir and Booz Allen were reported to be restricting frontier-model use over data-retention and IP fears on the same day Gartner told its IT Symposium that the leading AI vendors "are not enterprise grade." That is not a verdict on AI. It is a procurement specification being written in public, by the buyers, for the first time — and the vendors who ship that spec as a product instead of arguing with it will collect the regulated budgets that have been sitting on the sidelines all year.

🌊 Tide

Confirmed — governance-as-market-structure. The pacing debate left the labs and reached the state and the tape on the same day: a US president, a Chinese security ministry and a 5.9% semiconductor selloff all priced "pacing" before a single rule exists. No tide shift; the direction is unchanged and the evidence got much louder.

Pacing met the state, and the market put a number on it

Trump rejected the frontier-lab pacing push outright, writing on Truth Social that existing criminal and regulatory authority is sufficient and that “there is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China.” He singled out Dario Amodei by name. The same morning, at the All-In Summit in Los Angeles, Jensen Huang took a call from Trump on stage and put him on speaker: Trump called the loss-of-control argument “a hoax” and said “we’re not going to let that happen,” and Huang — who had been discussing Amodei’s essay moments earlier — replied, “We’re not going to let that happen, sir,” to applause. Beijing answered from the other direction. The foreign ministry called the slowdown framing fearmongering, spokesperson Guo Jiakun saying “fear-mongering, confrontation and vicious competition would undermine sound global AI governance and serve no one’s interests” — while Minister of State Security Chen Yixin warned that hostile forces were abusing generative AI to “fabricate political rumours and spread harmful information” against China’s political, institutional and ideological security. Then the tape voted. The PHLX semiconductor index fell 5.9%, cutting its 2026 gain to 57%; Nvidia closed down 3.4%, Micron more than 5%, Broadcom and AMD more than 4% each; the S&P 500 lost 0.48% to 7,619.94 and the Nasdaq Composite 0.56% to 26,186.41. SoftBank fell 11% in Tokyo after OpenAI said it would not list this year. Cybersecurity names went the other way, with Okta, CrowdStrike and Palo Alto Networks all higher.

So what: Two editions ago the standards body was three labs in a private room. Now it is a live political fight with a cost of capital attached — and the first thing that repriced was not the labs, it was the picks and shovels. Read that as the opening it is. The market has just told you it cannot yet distinguish "slower frontier training" from "less compute demand," which are not the same thing at all: pacing, as every principal actually described it on the day, means more compute spent on alignment, evaluation and monitoring, not less compute. If you are building, this is the week to write the evaluation-and-audit line into your product, because it is about to be the line that clears procurement — and if you are allocating, note that a 5.9% one-day move on a policy argument is what a market looks like before it has priced the mechanism.

Trump says calls for more control on AI are a ‘SICK conspiracy’
Nvidia CEO Jensen Huang tells Trump ‘we’re not going to let [an AI slowdown] happen’
Wall Street ends down, calls for AI slowdown pummel chipmakers
China dismisses AI ‘fearmongering’ as spy chief warns of threat to Communist party rule
Fear-mongering over AI serves no one's interests: spokesperson
Cybersecurity stocks get a jolt on gloomy AI warnings from CEOs of Anthropic and OpenAI

🌊 Waves

The buyers started writing the spec: data residency becomes the enterprise gate

The Information reported that Nvidia, Palantir and Booz Allen Hamilton have each restricted or threatened to drop frontier models from Anthropic and OpenAI over data-retention and IP exposure. Palantir has pressed Anthropic for zero-data-retention guarantees before exposing the models through its own software; Nvidia confines Anthropic's models to less sensitive internal work and prefers its own Nemotron models for proprietary tasks; Booz Allen has barred staff from running Anthropic's commercial model on cybersecurity projects that touch proprietary data. The proximate trigger is the rolling 30-day retention of prompts and outputs Anthropic introduced for automated cross-session misuse monitoring. Anthropic had already shipped the answer on Tuesday, September 1: Enterprise Frontier Safeguards keeps the monitoring store in the customer's own S3, Azure Blob or Google Cloud Storage under the customer's own keys, with automated scanning and no Anthropic human review, built with more than 100 customers, free, phased in later this fall, with zero-data-retention on Fable 5 and 5.1 in the interim. On the same Monday, Gartner VP analyst Daryl Plummer told the Australian leg of its IT Symposium/Xpo that "trust in these vendors is not warranted yet. They are not enterprise grade. They don't understand enterprise terms and conditions. They don't understand enterprise liability."

Roadmap implication: Roadmap implication: data residency has stopped being a legal annex and become a product surface, and the labs that treat it that way will win the accounts that spend the most. Anthropic's version is instructive precisely because it arrived thirteen days before the story that would have forced it — bring-your-own-bucket, bring-your-own-key, monitoring without custody. Put an equivalent on your roadmap this quarter if you sell into finance, health, defense or anything that files with a regulator: customer-held logs, contractual protection that covers what a provider may learn from your traffic, not just whether it trains on your raw data, and an audit trail a procurement officer can read without a lawyer. The constraint is real, but the framing that matters is that a specification is worth far more to a builder than a complaint — the buyers just published one for free.

Anthropic Data Fears Prompt Nvidia, Palantir and Booz Allen to Restrict Model Use
Palantir, Nvidia, Booz Allen restrict Anthropic and OpenAI models
AI and its main promoters are not enterprise-ready, says Gartner
Developing Enterprise Frontier Safeguards with our customers

Apple put the assistant on the device, and the interface question moved again

Apple shipped Siri AI and the next generation of Apple Intelligence across the OS 27 family: personal context drawn from mail, messages and photos, onscreen awareness, cross-app actions, Visual Intelligence, Writing Tools, and Private Cloud Compute for the work that cannot run on-device. It arrives as an opt-in, waitlisted English beta, timed just ahead of iPhone 18 availability, with some server-heavy features rate-limited daily and paid expansion signalled later; it is not initially available in China, and the EU exclusion covers iOS, iPadOS and watchOS while Mac and Vision Pro users there do get it. The assistant layer is no longer one company's product category — it is arriving preinstalled on the world's most valuable installed base at the same moment ChatGPT sits above a billion weekly active users.

Roadmap implication: Roadmap implication: this is the distribution-rewrite tide arriving as an SDK question rather than a strategy question. If your product's value is a screen a user must remember to open, a competent on-device assistant with cross-app actions and personal context is now the thing standing between you and that user. The move is not to fight it — it is to become the service the assistant calls. Decide this quarter what your three highest-value actions are, expose them as clean intents and tools, and instrument what the assistant actually invokes. The China exclusion and the partial EU hold-back also buy non-US builders a real window, which is worth noticing before it closes.

Siri AI, a profoundly more capable and personal assistant, is here
Apple releases iOS 27, redesigned Siri AI

Agentic coding's second-order bill arrived, and it is mostly architecture

Anthropic published its own numbers on what agentic coding does downstream: Claude now authors roughly 80% of Anthropic's code, engineers ship about 8x as much code per quarter as they did in 2021–2025, the test suite grew 10x, and CI jobs rose 25x in six months against a nominal headcount increase. Three patches to its test-selection service failed before an engineer rebuilt test-impact analysis around stateless listeners, an in-memory journal and a rollup consumer; the advice to other teams is to assume your architecture will be at 25x load within two quarters. The same day, Polylane published the counterweight on agent design. Its autofix system had grown to as many as 18 agents and sub-agents — triage, coordinator, up to fifteen hypothesis agents, a coding agent — and the failures lived in the handoffs: each summary dropped evidence the next agent never saw, per-stage evaluations passed while end-to-end fixes did not. On Thursday, September 3 they collapsed it into one agent that sees the whole run. Median time from detection to pull request went from 2.2 hours to 35 minutes, p90 from nine days to under two hours, and the share of detected issues that ended in a pull request from 0.6% to 4.2% and rising. Average model spend per pull request fell from $111 to about $18 over the first nine days — which the authors themselves caveat is not purely the architecture change, since they were iterating on everything at once.

Roadmap implication: Roadmap implication: budget the loop, not just the model. Two concrete moves. First, if you have turned coding agents loose, your CI, test infrastructure and review capacity are the binding constraint within two quarters, not your token bill — size them for 25x now, while it is a planning decision rather than an outage. Second, stop reflexively fanning work out to sub-agents: the handoff summary is where the evidence dies, and one agent holding the whole context beat an 18-agent pipeline on latency, hit rate and cost simultaneously. Evaluate the run end-to-end, keep one trace, and treat per-stage green lights as the warning sign they are.

Agentic coding is straining CI. Here’s how we scaled test impact analysis at Anthropic
Sub-agents are just wrong

🌊 Ripples

Microsoft published the constitution, on the day it said it would

Microsoft AI opened a six-week public consultation on a first draft of its Humanist AI Code of Conduct for MAI models — the document Satya Nadella committed to publishing on September 14 when he endorsed pacing over the weekend. It is written to outrank operator policy and user preference. The Absolute Constraints say MAI models "will never resist human interruption, override, correction, or shutdown" and must "always recognize the primacy of human intent"; the framework's own line is "interruptible, correctable, shut-down-able. If it isn't, we don't ship it." Also barred: assistance with chemical, biological, radiological, nuclear or explosive weapons, dangerous cyber capability, deception, claims of personhood, and hidden machine-only "neuralese" communication. A revised version is planned for later in 2026.

Do this now: read the Absolute Constraints as a draft term sheet, not a press release. Its chain of command explicitly ranks the Code above operator policy and above user preference — which means if you build on MAI, there are things you will not be able to configure. Six weeks is a real comment window and Microsoft asked to be argued with; if your use case sits near a constraint boundary, file that now rather than discovering it in production.

Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models
Humanist AI Code of Conduct
Microsoft AI Opens Six-Week Review of Draft Rules Governing MAI Behavior

‘Project Lily’: humans are reading ChatGPT conversations

404 Media reported that OpenAI runs an internal program, codenamed Project Lily, paying hundreds of contractors through third-party staffing firms to read real ChatGPT conversations and rate the replies, largely to curb sycophancy and overly human-like behaviour. Reviewers do not see usernames and OpenAI says it filters personal information before prompts reach them, but the company acknowledges sensitive details still get through. Separately, reported the same day, the Midas Project has alleged OpenAI failed to publish promised frontier-risk assessments for recent model releases under California's SB 53; OpenAI says it complies.

Do this now: check what tier your organisation is actually on and what its training and review settings are, then tell your people plainly what is and is not private. The consumer tier is not the enterprise tier, and the gap between what employees assume and what the terms say is where the incident comes from. This is also the demand-side explanation for the enterprise wave above — the same week buyers started writing retention into contracts, the reason why led the news.

Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats
OpenAI paid contractors to read ChatGPT conversations — here’s how to protect yourself

SemiAnalysis measured Vera Rubin, and the honest number is engine-dependent

SemiAnalysis published the first verified agentic-inference results for NVIDIA's Vera Rubin NVL72, measured on its AgentX benchmark replaying real agentic traffic. The headline: at 170 tokens/sec, Rubin delivers roughly 67x the total throughput per dollar of TCO of a GB300 running Dynamo/TensorRT-LLM under owning-cost assumptions, and about 61% higher maximum P90 interactivity (276.24 vs 171.53 TPS). The useful reading is the fine print the authors supply themselves. At 60–100 TPS, where most providers actually serve, the advantage is 1.4x to 3x. And on the separate per-megawatt measure, at that same 170 TPS point Rubin is 62.9x a GB300 running TensorRT-LLM but only 5.56x the same GB300 running the open-source SGLang stack — which also closes most of the interactivity gap outright. They also caught NVIDIA underselling itself: Jensen Huang claimed 3x performance per megawatt over Blackwell at GTC 2026, and the measured figure on pre-release software is up to 7x.

Do this now: when a 67x number crosses your desk, ask which software stack sat on the other side of the comparison before you put it in a board deck. The same hardware pair produces 67x throughput per dollar of TCO at 170 TPS against one engine, and on the separate per-megawatt measure 62.9x against GB300 TensorRT-LLM but 5.56x against GB300 SGLang at that same point — all of them true, and only one of them a headline. For anyone modelling inference economics, the operating point that matters is 60–100 TPS, where the honest answer is a real but ordinary 1.4x–3x.

Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar

The open-weight cost curve reached the desk

DeepSeek's V4.1-Flash, an MIT-licensed 552B-parameter mixture-of-experts model released September 10, spent Monday becoming a hardware-economics cottage industry. Developers published working setups running the full model on a single 128GB DGX Spark by keeping only task-relevant experts resident, and on Apple Silicon by streaming experts from SSD — down as far as a 16GB Mac mini — at roughly 23 seconds per token, which is a proof of portability rather than a production configuration. Arena.ai's Agent Arena leaderboard now places it fourth overall and second among open-weight models, at 13.75% confirmed success against Claude Fable 5.1 (Max) at 23.70% and GPT-6 Astra (Max) at 19.54%. Artificial Analysis puts it at 40 on its overall Intelligence Index — sixth of 113 models and far above the median of 18 for comparable models — while it scores 68.9% on AutomationBench-AA against GPT-6 Astra's 68.5% across 657 held-out Zapier workflows. Keep the reading honest in both directions: the frontier still wins the hardest agentic work by a wide margin, and this model matches it on a large, ordinary class of automation at open-weight prices.

Do this now: take your highest-volume, lowest-entropy automation workload — the classification, extraction and workflow-routing traffic you currently pay frontier prices for — and benchmark it against V4.1-Flash on your own data this week. AutomationBench parity with Astra at open-weight prices is exactly the shape of task where routing pays, and the fact that it runs at all on a workstation means the sovereignty and data-residency conversation from the wave above suddenly has a local answer.

DeepSeek V4.1 Flash (max) - Intelligence, Performance & Price Analysis
AutomationBench-AA: Agentic SaaS Workflow Benchmark
Agent Arena | AI Agent Performance Leaderboard
deepseek-ai/DeepSeek-V4.1-Flash

OpenAI bought a camera team

OpenAI acquired Glass Imaging for more than $300 million, per the Wall Street Journal. The Los Altos company, founded in 2019 and previously funded with roughly $30 million, was built by Ziv Attar and Tom Bishop, the pair who led the team behind Apple's Portrait Mode; its GlassAI neural image-processing stack already powers zoom on Honor phones. OpenAI has not said where the technology lands, and the deal sits beside the still-unannounced Jony Ive hardware effort acquired for $6.5 billion in 2025.

Do this now: nothing, but update your map. A $300M acqui-hire of capture-side computational imaging talent, on the same day Apple shipped an assistant with onscreen awareness and Visual Intelligence, tells you what OpenAI thinks the ambient device needs to see. If you are building anything that depends on a phone camera as an input to an agent, assume the capture layer gets materially better and materially more contested within a year.

OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says

Read the full edition and the archive: https://excelsiorgroup.ai/insights/signal/

The Signal · The Excelsior Group