Closing Time: The slowdown just got signatures

• Anthropic CEO Dario Amodei published We Must Pace the Frontier, a three-step plan to slow capability gains, and said Anthropic is unilaterally giving independent evaluators permanent employee-level access, CNBC reported.

• OpenAI CEO Sam Altman and Elon Musk both backed the call on Saturday, with Altman saying OpenAI will also commit to independent evaluators with employee-like access, CNBC said.

• Nvidia is in talks to anchor Anthropic’s planned IPO with a stake of as much as $10 billion as Anthropic eyes raising up to $100 billion at around a $2 trillion valuation, Reuters reported.

• The UAE is redesigning its 5 GW AI campus away from a single Abu Dhabi megasite toward a multi-site network with air defenses and possible underground builds after Iranian attacks, Reuters reported.

• Meta suspended its Model Capability Initiative after employees revolted over mandatory keystroke and screen tracking for AI training, and the program remains suspended as opt-in only, Business Insider reported.

---

The slowdown just got signatures

CNBC reports that Anthropic CEO Dario Amodei published a Saturday essay urging frontier labs to intentionally pace how fast they raise model capabilities. Anthropic is committing unilaterally to the first step: permanent, employee-level access for independent evaluators who can verify safety work and publish findings without Anthropic’s editorial control.

The second step asks leading labs in democratic countries to set common safety standards together. The third reaches for coordination between democratic and authoritarian governments. Amodei was careful to say pacing is not a halt on training; it is time for alignment, safeguards, and outside confirmation before the next capability jump.

That matters because the week’s resignation drama just got a CEO-level policy answer. After Jacob Coxon’s public quit and a string of agent security scares, the industry’s most safety-branded lab put a concrete voluntary slowdown on paper, and rivals did not laugh it off.

The one move is to watch whether Altman’s promised evaluator access and any shared safety bars land before the next major model drop, because a Saturday essay only counts if the next training run actually waits. Read CNBC

Nvidia may underwrite the Anthropic float

Reuters reports that Anthropic is talking with Nvidia about an anchor investment of as much as $10 billion in what could be the largest IPO on record. Anthropic is seeking to raise as much as $100 billion at a valuation around $2 trillion, with a listing still expected before the November midterms, according to people familiar with the talks.

The plans remain fluid, and both companies declined to confirm the discussions on the record. Nvidia already sits deep inside Anthropic’s compute stack through prior partnership capital and Azure capacity powered by Nvidia chips, while Anthropic also leans on Amazon, Google, and Broadcom for diversifying silicon.

That matters because the same week Amodei asks the industry to slow capability gains, the chip supplier that enables those gains is being courted to stamp the biggest public-market bet on Claude. Capital and caution are moving in the same headline cycle.

The one move is to track whether Nvidia’s check becomes a filed commitment in the IPO roadshow, because an anchor that large would set the tone for every other frontier listing that follows. Read Reuters

Meta’s own staff killed the data grab

Business Insider reports that Meta’s Model Capability Initiative, a mandatory program to capture employee keystrokes and mouse movements for AI training, was suspended after an internal revolt and a summer data leak. More than 1,800 employees signed a petition, organizers papered campuses with flyers, and workers nicknamed the recorder icon the Eye of Sauron.

The push started inside Meta’s elite TBD AI unit as a way to teach models the messy desktop behaviors that still trip agents. CTO Andrew Bosworth initially said there would be no opt-out. After engineers found MCI data had been viewable company-wide, Meta shut the program off and said any future version would be opt-in. Two months later, it remains suspended.

That matters because the AI data race just hit a hard internal wall. The people closest to the models refused to become the training set, and a hyperscaler with hundreds of billions in AI spend had to blink.

The one move is to watch whether Meta restarts MCI as a true opt-in or quietly replaces it with synthetic desktop data, because the next agent leap still needs human computer traces from somewhere. Read Business Insider

---

Deals Desk

This week’s compute and power tape from the tracker, newest first.

----------

G42 (UAE AI campus / Stargate UAE) × OpenAI; Oracle; Nvidia; SoftBank; Cisco — Sep 11, 2026

AI data center campus redesign (5 GW plan redistributed) · United Arab Emirates (multi-site network after original single Abu Dhabi campus plan)

5 GW overall plan under redesign; first phase Stargate UAE cited as $30B / 1 GW cluster with first 200 MW due in 2026

Status · under construction · Source: Reuters

--------------------

Nvidia × Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC, AirTrunk — Sep 10, 2026

AI factory / data center capacity partnerships (DSX platform) · Australia

Up to 2 GW of AI-related capacity by 2027 (partners’ sites; Australia’s current computing capacity cited at 1.6 GW)

Status · announced · Source: Reuters

--------------------

Fluidstack × Cameron County (worksite permit); AEP Texas; East Rio Hondo Water Supply — Sep 10, 2026

AI data center campus (phase-one site work / worksite permit) · Unincorporated Cameron County, Texas, US

$4B phase one on >1,100 acres; critical IT load and anchor tenant not disclosed; phase-one completion targeted mid-2028

Status · under construction · Source: Cameron County

----------

Full tracker → https://www.theaigentic.com/deals