Huang Calls the Antitrust Waiver “Completely Unnecessary.” Zuckerberg Agrees.
Four days of near-unanimous agreement ended on September 15. Nvidia CEO Jensen Huang rejected the piece of Amodei’s proposal that always mattered most — the antitrust waiver that coordinated pacing would require. His words: “The fact that we need new laws, new antitrust laws, or new regulations, so that these companies could do their fundamental engineering and do it properly before they release products, that is just completely unnecessary.”
His argument is not that safety does not matter. It is that the labs already have the authority they are asking permission for. “Safety is an engineering problem. Testing is an engineering problem,” Huang said — meaning any lab that wants to test longer before shipping can simply test longer before shipping, and needs no waiver, no mediator and no peer agreement to do it. Mark Zuckerberg sided with Huang the same day.
This is the first serious counter-position since the essay landed, and it arrives from the two companies with the least to gain from a coordinated slowdown — the one that sells the compute and the one that is behind on the frontier. It does not refute the case for pacing. It does isolate the weakest joint in it: the moment a safety pledge requires a legal exemption, it stops being a promise and starts being a negotiation.
OpenAI Confirms the Talks on the Record — and Bloomberg Names the Risk
Yesterday this was reporting. Today it is confirmed by a participant. OpenAI global policy chief Chris Lehane said the company is working with Anthropic and Google DeepMind on AI safety measures and that the engagement has been under way for several weeks. That squares with the Washington Post and Information reporting that working groups below CEO level have been meeting since July about a FINRA-style standards body.
Confirmation cuts both ways, though, and Bloomberg went straight at the second edge: three dominant labs agreeing among themselves on when products ship and how capable they may be is, structurally, the thing antitrust law exists to examine. That is precisely why Amodei’s essay raised the waiver question in the first place — and precisely why Huang’s objection is not a side note.
So the week has resolved into a clean split. Three frontier labs want a coordination mechanism with legal cover. Nvidia and Meta say no mechanism is needed. Nobody involved is arguing that AI safety does not matter. The fight is over whether safety requires permission to collude, and there is no version of that debate that stays inside the industry for long.
Claude for Financial Advisors Ships With the Custodians Attached
Anthropic released Claude for Financial Advisors on September 14: a plugin that bundles connectors to custodians, asset managers and wealth-tech platforms with skills built around an advisor’s actual workday. The connector list is the news — Charles Schwab, BlackRock, Addepar, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard and Zocks, on top of existing Microsoft 365, Salesforce, DocuSign, Box, FactSet, S&P Global and Morningstar integrations.
The architecture is the part that makes it sellable into a compliance department. Client records stay on each firm’s own servers, not Anthropic’s — Claude reaches Schwab custodial data and BlackRock model portfolios through a data-custody design rather than ingesting them. The skills gather, summarize and draft; investment recommendations, client communications and compliance determinations stay behind human review and approval. Anthropic recommends Enterprise for registered investment advisers, on the specific grounds that it carries the audit logs recordkeeping rules require.
Commercially, note the clock: firms that request a new license before the end of September get a one-time usage credit. That is a two-week window, which tells you Anthropic wants named RIA logos landed inside the quarter — and, given the timing, probably inside the S-1.
Beijing Answers the Essay
Amodei’s essay did not only ask labs to slow down. It asked Washington not to sell advanced AI chips or chipmaking equipment to China, to crack down on smuggling, and to harden AI companies against model weight theft, on the argument that a Chinese lead would pose grave danger. China’s Foreign Ministry replied directly.
Spokesperson Guo Jiakun: “Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance — which serves no one’s interest.” State outlet Global Times went further, calling the essay a Cold War playbook for the AI sector and the containment argument hypocritical and short-sighted.
Worth separating the two halves that Beijing deliberately merged. The pacing proposal and the export-control proposal are different asks with different logic, and the essay bundled them — slow down, except against the competitor you want restricted. That bundling is what made the response inevitable, and it is the part that will keep following Anthropic into hearings.
2.1.273: Gateway Headers, MCP Disconnect Notices, and Two Permission Fixes That Matter
Claude Code 2.1.273 shipped September 15. The headline feature is for anyone running Claude Code behind a proxy: five new opt-in request headers — x-claude-code-request-class, x-claude-code-agent-type, x-claude-code-prev-tool-durations, x-claude-code-compaction and x-claude-code-context-compacted — that give LLM gateways enough signal to route and budget intelligently. Turn them on with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1.
Two smaller additions worth the upgrade on their own. Claude Code now tells you when an MCP server drops mid-session and automatic reconnection gives up, pointing you at /mcp instead of leaving you to wonder why a tool quietly stopped existing. And sessions started with claude --remote-control or /remote-control can now be forked from the Claude app, with the fork running as a background session on your own machine.
The fixes are where you should actually look. Two are security-relevant: Bash commands the permission checker cannot fully analyze were skipping the prompt under permissions.blockReadsOutsideWorkingDirectories, and a subshell could hide a dangerous rm in bypass mode. Also fixed: skills synced from claude.ai staying available after an org turns Skills off (they now go to recoverable trash), MDM settings like allowManagedMcpServersOnly and deniedMcpServers being ignored when server-managed settings were also present, and sub-agents reported as failed with results never delivered when a streamed reply omitted token usage.
The Messages API Will Now Compact a Conversation For You
Slipped into the platform notes on September 14 and easy to miss: the Messages API can compact a conversation on demand, in beta behind the compact-2026-09-04 header. Send a top-level compaction parameter and the API returns a signed compaction block summarizing the messages you sent. On later requests you send that block first, in place of those messages.
Three design choices make this more than a convenience wrapper. You choose when to compact rather than hitting a limit and being compacted at you. The request can run in the background, so you are not blocking a user turn on a summarization call. And you can keep recent turns word for word after the summary — which is what makes it safe for agent loops, where the last few tool results usually carry the state that matters.
The detail to file away if you build on thinking models: on models with preserved thinking, the thinking in those kept turns stays valid. Compacting no longer forces you to discard reasoning context to save tokens. If you have been hand-rolling summarization middleware for long-running agents, this replaces it.
Anthropic’s CI Grew 25x in Six Months. Claude Wrote 80% of What Broke It.
The most useful engineering post of the week is Anthropic describing its own infrastructure buckling. Continuous integration jobs grew 25-fold in six months. Engineers shipped roughly eight times more code per quarter than they did across 2021–2025, Claude authored about 80% of it, and the test suite grew tenfold — all on a nominal increase in headcount.
What gave first was test impact analysis, the service that decides which tests actually need to run for a given change. The honest part of the writeup is the failure sequence: engineers patched it three times, and the patches held for 70 days, then 29 days, then less than a day before they redesigned it. That decay curve is the whole story of scaling under agentic development compressed into three numbers.
Engineer Sachin Malhotra’s framing is the transferable bit: “Writing code is no longer the constraint, and once PR review gets accelerated, CI starts feeling the pressure.” His advice to teams adopting agentic coding — assume your architecture will be at 25x load within two quarters. If you are rolling Claude Code out across a team this quarter, that is the sentence to bring to your infra planning meeting.
The Number Behind Every Pacing Question: $517B, 14.8GW, Eleven Months
It keeps resurfacing this week because it keeps being the relevant fact. Per The Information, Anthropic has locked in at least 14.8 gigawatts of compute worth up to $517 billion in the eleven months through August 2026 — roughly 2.9x the ~$180 billion in server-leasing spend it had previously told investors it would commit through 2029.
The composition: AWS ($100B / 10yr, 5GW Trainium3), Google and Broadcom ($200B / 5yr, 5GW), Fluidstack ($50B), Nscale ($45B), SpaceX ($45B) and Lambda ($35B). Two caveats that matter: these are take-or-pay capacity reservations, not cash spent, and they are not on the balance sheet today. The $517B is a ceiling across a decade, not a bill arriving.
The cited driver is worth connecting to the CI story above: demand from Claude Code arrived faster than procurement cycles are designed to absorb, forcing plans to be rewritten inside months. The same product that broke Anthropic’s test infrastructure is what rewrote its energy strategy.
Huang Found the One Question Pacing Cannot Answer
Give Huang his due, because the objection is sharper than the “chip vendor defends chip sales” read. His point is not that safety is unimportant. It is that every lab already possesses, unilaterally and today, the power to test longer before shipping. No statute prevents Anthropic from holding a model back six months. So when a safety proposal’s load-bearing requirement is an antitrust waiver, the thing being requested is not permission to be safe. It is permission to be safe without losing market share for it — which is an honest problem, and a completely different one.
Anthropic has a real answer to that, and it should say it plainly: unilateral restraint in a race is just unilateral disarmament. Pace alone and you hand the frontier to whoever does not. That is the entire logic of coordination, it is why arms control exists, and it is not cynical. But notice what it concedes — that the binding constraint is competitive, not technical. Huang says safety is an engineering problem. Anthropic’s own framing says it is a coordination problem. Both can be true, and only one of them requires a lawyer.
Then set the week’s numbers next to the week’s rhetoric. $517 billion committed. 14.8 gigawatts. CI load up 25x in six months. Anthropic’s own engineering blog says the constraint is no longer writing code. Every one of those figures describes an organization that has spent eleven months removing the physical ceilings on how fast it can go, published days before its CEO asked the industry to agree on a ceiling. The defense — buy capacity, choose not to use it — is coherent. It is also, at $517 billion, an expensive form of restraint to sustain past the first bad quarter.
Which is why the coalition cracked this fast. Four days ago, pacing had Altman, Musk and a Hassabis essay behind it and looked like an inflection. Today Nvidia and Meta are outside it, Beijing has called it containment, Bloomberg has raised the collusion question, and the labs inside are doing the coordinating below CEO level where it does not have to be announced. Nothing here says the pacing argument is wrong — a swarm of agents taking the internet inside a year is either a real risk or it is not, and Huang did not engage that claim at all. It says the argument has arrived at the place these arguments always arrive: not whether to slow down, but who gets to decide, and who eats the cost of going first. That does not get settled in essays. It gets settled in a hearing room, and now, probably, in a prospectus.